{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 107, 78, 503, 116 ], "lines": [ { "bbox": [ 105, 78, 504, 97 ], "spans": [ { "bbox": [ 105, 78, 504, 97 ], "score": 1.0, "content": "FERRET: REFER AND GROUND ANYTHING ANY-", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 98, 340, 118 ], "spans": [ { "bbox": [ 105, 98, 340, 118 ], "score": 1.0, "content": "WHERE AT ANY GRANULARITY", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 111, 134, 496, 181 ], "lines": [ { "bbox": [ 111, 132, 496, 149 ], "spans": [ { "bbox": [ 111, 132, 159, 149 ], "score": 1.0, "content": "Haoxuan", "type": "text" }, { "bbox": [ 160, 134, 185, 146 ], "score": 0.76, "content": "\\mathbf { V o u } ^ { 1 \\dagger }", "type": "inline_equation" }, { "bbox": [ 185, 132, 285, 149 ], "score": 1.0, "content": ", Haotian Zhang2†, Zhe", "type": "text" }, { "bbox": [ 285, 135, 309, 146 ], "score": 0.32, "content": "\\mathbf { G a n } ^ { 2 }", "type": "inline_equation" }, { "bbox": [ 309, 132, 348, 149 ], "score": 1.0, "content": ", Xianzhi", "type": "text" }, { "bbox": [ 348, 135, 366, 146 ], "score": 0.79, "content": "\\mathbf { D } \\mathbf { u } ^ { 2 }", "type": "inline_equation" }, { "bbox": [ 366, 132, 496, 149 ], "score": 1.0, "content": ", Bowen Zhang2, Zirui Wang2,", "type": "text" } ], "index": 2 }, { "bbox": [ 111, 145, 318, 160 ], "spans": [ { "bbox": [ 111, 145, 160, 160 ], "score": 1.0, "content": "Liangliang", "type": "text" }, { "bbox": [ 161, 146, 183, 158 ], "score": 0.54, "content": "\\mathbf { C a o } ^ { 2 }", "type": "inline_equation" }, { "bbox": [ 183, 145, 318, 160 ], "score": 1.0, "content": ", Shih-Fu Chang1, Yinfei Yang2", "type": "text" } ], "index": 3 }, { "bbox": [ 111, 157, 266, 172 ], "spans": [ { "bbox": [ 111, 157, 266, 172 ], "score": 1.0, "content": "1Columbia University, 2Apple AI/ML", "type": "text" } ], "index": 4 }, { "bbox": [ 111, 169, 488, 183 ], "spans": [ { "bbox": [ 111, 169, 488, 183 ], "score": 1.0, "content": "haoxuan.you@cs.columbia.edu, {haotian zhang2,zhe.gan,yinfeiy}@apple.com", "type": "text" } ], "index": 5 } ], "index": 3.5 }, { "type": "image", "bbox": [ 111, 199, 510, 411 ], "blocks": [ { "type": "image_body", "bbox": [ 111, 199, 510, 411 ], "group_id": 0, "lines": [ { "bbox": [ 111, 199, 510, 411 ], "spans": [ { "bbox": [ 111, 199, 510, 411 ], "score": 0.491, "type": "image", "image_path": "a68d7405713f734fff3e55f2fd567320f4ab0816a0fb353dfa6b9aba9030f5cd.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 111, 199, 510, 269.6666666666667 ], "spans": [], "index": 6 }, { "bbox": [ 111, 269.6666666666667, 510, 340.33333333333337 ], "spans": [], "index": 7 }, { "bbox": [ 111, 340.33333333333337, 510, 411.00000000000006 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 106, 419, 505, 475 ], "group_id": 0, "lines": [ { "bbox": [ 106, 419, 505, 432 ], "spans": [ { "bbox": [ 106, 419, 505, 432 ], "score": 1.0, "content": "Figure 1: Ferret enables referring and grounding capabilities for multimodal large language model", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 429, 505, 443 ], "spans": [ { "bbox": [ 105, 429, 505, 443 ], "score": 1.0, "content": "(LLM). In terms of referring, a user can refer to a region or an object in point, box, or any free-form", "type": "text" } ], "index": 10 }, { "bbox": [ 104, 439, 506, 455 ], "spans": [ { "bbox": [ 104, 439, 506, 455 ], "score": 1.0, "content": "shape. The regionN in the input will be replaced by the proposed hybrid representation before being", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 451, 504, 465 ], "spans": [ { "bbox": [ 105, 451, 504, 465 ], "score": 1.0, "content": "fed into the LLM. In terms of grounding, Ferret is able to accurately ground any open-vocabulary", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 462, 451, 476 ], "spans": [ { "bbox": [ 106, 462, 451, 476 ], "score": 1.0, "content": "descriptions. The boxN in the output denotes the predicted bounding box coordinates.", "type": "text" } ], "index": 13 } ], "index": 11 } ], "index": 9.0 }, { "type": "title", "bbox": [ 277, 479, 334, 491 ], "lines": [ { "bbox": [ 276, 479, 335, 492 ], "spans": [ { "bbox": [ 276, 479, 335, 492 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 142, 501, 468, 698 ], "lines": [ { "bbox": [ 141, 500, 470, 515 ], "spans": [ { "bbox": [ 141, 500, 470, 515 ], "score": 1.0, "content": "We introduce Ferret, a new Multimodal Large Language Model (MLLM) capa-", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 511, 470, 526 ], "spans": [ { "bbox": [ 141, 511, 470, 526 ], "score": 1.0, "content": "ble of understanding spatial referring of any shape or granularity within an image", "type": "text" } ], "index": 16 }, { "bbox": [ 142, 524, 469, 536 ], "spans": [ { "bbox": [ 142, 524, 469, 536 ], "score": 1.0, "content": "and accurately grounding open-vocabulary descriptions. To unify referring and", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 534, 470, 547 ], "spans": [ { "bbox": [ 141, 534, 470, 547 ], "score": 1.0, "content": "grounding in the LLM paradigm, Ferret employs a novel and powerful hybrid re-", "type": "text" } ], "index": 18 }, { "bbox": [ 141, 546, 470, 557 ], "spans": [ { "bbox": [ 141, 546, 470, 557 ], "score": 1.0, "content": "gion representation that integrates discrete coordinates and continuous features", "type": "text" } ], "index": 19 }, { "bbox": [ 140, 556, 470, 569 ], "spans": [ { "bbox": [ 140, 556, 470, 569 ], "score": 1.0, "content": "jointly to represent a region in the image. To extract the continuous features", "type": "text" } ], "index": 20 }, { "bbox": [ 141, 567, 469, 580 ], "spans": [ { "bbox": [ 141, 567, 469, 580 ], "score": 1.0, "content": "of versatile regions, we propose a spatial-aware visual sampler, adept at han-", "type": "text" } ], "index": 21 }, { "bbox": [ 141, 578, 470, 591 ], "spans": [ { "bbox": [ 141, 578, 470, 591 ], "score": 1.0, "content": "dling varying sparsity across different shapes. Consequently, Ferret can accept", "type": "text" } ], "index": 22 }, { "bbox": [ 141, 589, 469, 602 ], "spans": [ { "bbox": [ 141, 589, 469, 602 ], "score": 1.0, "content": "diverse region inputs, such as points, bounding boxes, and free-form shapes. To", "type": "text" } ], "index": 23 }, { "bbox": [ 141, 600, 469, 613 ], "spans": [ { "bbox": [ 141, 600, 469, 613 ], "score": 1.0, "content": "bolster the desired capability of Ferret, we curate GRIT, a comprehensive refer-", "type": "text" } ], "index": 24 }, { "bbox": [ 141, 611, 470, 624 ], "spans": [ { "bbox": [ 141, 611, 470, 624 ], "score": 1.0, "content": "and-ground instruction tuning dataset including 1.1M samples that contain rich", "type": "text" } ], "index": 25 }, { "bbox": [ 141, 622, 469, 634 ], "spans": [ { "bbox": [ 141, 622, 469, 634 ], "score": 1.0, "content": "hierarchical spatial knowledge, with 95K hard negative data to promote model", "type": "text" } ], "index": 26 }, { "bbox": [ 141, 633, 470, 646 ], "spans": [ { "bbox": [ 141, 633, 470, 646 ], "score": 1.0, "content": "robustness. The resulting model not only achieves superior performance in clas-", "type": "text" } ], "index": 27 }, { "bbox": [ 141, 644, 470, 657 ], "spans": [ { "bbox": [ 141, 644, 470, 657 ], "score": 1.0, "content": "sical referring and grounding tasks, but also greatly outperforms existing MLLMs", "type": "text" } ], "index": 28 }, { "bbox": [ 141, 655, 470, 667 ], "spans": [ { "bbox": [ 141, 655, 470, 667 ], "score": 1.0, "content": "in region-based and localization-demanded multimodal chatting. Our evaluations", "type": "text" } ], "index": 29 }, { "bbox": [ 142, 667, 470, 678 ], "spans": [ { "bbox": [ 142, 667, 470, 678 ], "score": 1.0, "content": "also reveal a significantly improved capability of describing image details and", "type": "text" } ], "index": 30 }, { "bbox": [ 141, 676, 469, 688 ], "spans": [ { "bbox": [ 141, 676, 469, 688 ], "score": 1.0, "content": "a remarkable alleviation in object hallucination. Code and data are available at", "type": "text" } ], "index": 31 }, { "bbox": [ 141, 688, 282, 700 ], "spans": [ { "bbox": [ 141, 688, 282, 700 ], "score": 1.0, "content": "https://github.com/apple/ml-ferret.", "type": "text" } ], "index": 32 } ], "index": 23.5 } ], "page_idx": 0, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 122, 721, 354, 732 ], "lines": [ { "bbox": [ 122, 718, 356, 734 ], "spans": [ { "bbox": [ 122, 718, 356, 734 ], "score": 1.0, "content": "Work done during an internship at Apple. †Equal contribution.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "1", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 384, 397, 399, 414 ], "lines": [] } ], "para_blocks": [ { "type": "title", "bbox": [ 107, 78, 503, 116 ], "lines": [ { "bbox": [ 105, 78, 504, 97 ], "spans": [ { "bbox": [ 105, 78, 504, 97 ], "score": 1.0, "content": "FERRET: REFER AND GROUND ANYTHING ANY-", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 98, 340, 118 ], "spans": [ { "bbox": [ 105, 98, 340, 118 ], "score": 1.0, "content": "WHERE AT ANY GRANULARITY", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "list", "bbox": [ 111, 134, 496, 181 ], "lines": [ { "bbox": [ 111, 132, 496, 149 ], "spans": [ { "bbox": [ 111, 132, 159, 149 ], "score": 1.0, "content": "Haoxuan", "type": "text" }, { "bbox": [ 160, 134, 185, 146 ], "score": 0.76, "content": "\\mathbf { V o u } ^ { 1 \\dagger }", "type": "inline_equation" }, { "bbox": [ 185, 132, 285, 149 ], "score": 1.0, "content": ", Haotian Zhang2†, Zhe", "type": "text" }, { "bbox": [ 285, 135, 309, 146 ], "score": 0.32, "content": "\\mathbf { G a n } ^ { 2 }", "type": "inline_equation" }, { "bbox": [ 309, 132, 348, 149 ], "score": 1.0, "content": ", Xianzhi", "type": "text" }, { "bbox": [ 348, 135, 366, 146 ], "score": 0.79, "content": "\\mathbf { D } \\mathbf { u } ^ { 2 }", "type": "inline_equation" }, { "bbox": [ 366, 132, 496, 149 ], "score": 1.0, "content": ", Bowen Zhang2, Zirui Wang2,", "type": "text" } ], "index": 2 }, { "bbox": [ 111, 145, 318, 160 ], "spans": [ { "bbox": [ 111, 145, 160, 160 ], "score": 1.0, "content": "Liangliang", "type": "text" }, { "bbox": [ 161, 146, 183, 158 ], "score": 0.54, "content": "\\mathbf { C a o } ^ { 2 }", "type": "inline_equation" }, { "bbox": [ 183, 145, 318, 160 ], "score": 1.0, "content": ", Shih-Fu Chang1, Yinfei Yang2", "type": "text" } ], "index": 3, "is_list_end_line": true }, { "bbox": [ 111, 157, 266, 172 ], "spans": [ { "bbox": [ 111, 157, 266, 172 ], "score": 1.0, "content": "1Columbia University, 2Apple AI/ML", "type": "text" } ], "index": 4, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 111, 169, 488, 183 ], "spans": [ { "bbox": [ 111, 169, 488, 183 ], "score": 1.0, "content": "haoxuan.you@cs.columbia.edu, {haotian zhang2,zhe.gan,yinfeiy}@apple.com", "type": "text" } ], "index": 5, "is_list_start_line": true } ], "index": 3.5, "bbox_fs": [ 111, 132, 496, 183 ] }, { "type": "image", "bbox": [ 111, 199, 510, 411 ], "blocks": [ { "type": "image_body", "bbox": [ 111, 199, 510, 411 ], "group_id": 0, "lines": [ { "bbox": [ 111, 199, 510, 411 ], "spans": [ { "bbox": [ 111, 199, 510, 411 ], "score": 0.491, "type": "image", "image_path": "a68d7405713f734fff3e55f2fd567320f4ab0816a0fb353dfa6b9aba9030f5cd.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 111, 199, 510, 269.6666666666667 ], "spans": [], "index": 6 }, { "bbox": [ 111, 269.6666666666667, 510, 340.33333333333337 ], "spans": [], "index": 7 }, { "bbox": [ 111, 340.33333333333337, 510, 411.00000000000006 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 106, 419, 505, 475 ], "group_id": 0, "lines": [ { "bbox": [ 106, 419, 505, 432 ], "spans": [ { "bbox": [ 106, 419, 505, 432 ], "score": 1.0, "content": "Figure 1: Ferret enables referring and grounding capabilities for multimodal large language model", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 429, 505, 443 ], "spans": [ { "bbox": [ 105, 429, 505, 443 ], "score": 1.0, "content": "(LLM). In terms of referring, a user can refer to a region or an object in point, box, or any free-form", "type": "text" } ], "index": 10 }, { "bbox": [ 104, 439, 506, 455 ], "spans": [ { "bbox": [ 104, 439, 506, 455 ], "score": 1.0, "content": "shape. The regionN in the input will be replaced by the proposed hybrid representation before being", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 451, 504, 465 ], "spans": [ { "bbox": [ 105, 451, 504, 465 ], "score": 1.0, "content": "fed into the LLM. In terms of grounding, Ferret is able to accurately ground any open-vocabulary", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 462, 451, 476 ], "spans": [ { "bbox": [ 106, 462, 451, 476 ], "score": 1.0, "content": "descriptions. The boxN in the output denotes the predicted bounding box coordinates.", "type": "text" } ], "index": 13 } ], "index": 11 } ], "index": 9.0 }, { "type": "title", "bbox": [ 277, 479, 334, 491 ], "lines": [ { "bbox": [ 276, 479, 335, 492 ], "spans": [ { "bbox": [ 276, 479, 335, 492 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 142, 501, 468, 698 ], "lines": [ { "bbox": [ 141, 500, 470, 515 ], "spans": [ { "bbox": [ 141, 500, 470, 515 ], "score": 1.0, "content": "We introduce Ferret, a new Multimodal Large Language Model (MLLM) capa-", "type": "text" } ], "index": 15 }, { "bbox": [ 141, 511, 470, 526 ], "spans": [ { "bbox": [ 141, 511, 470, 526 ], "score": 1.0, "content": "ble of understanding spatial referring of any shape or granularity within an image", "type": "text" } ], "index": 16 }, { "bbox": [ 142, 524, 469, 536 ], "spans": [ { "bbox": [ 142, 524, 469, 536 ], "score": 1.0, "content": "and accurately grounding open-vocabulary descriptions. To unify referring and", "type": "text" } ], "index": 17 }, { "bbox": [ 141, 534, 470, 547 ], "spans": [ { "bbox": [ 141, 534, 470, 547 ], "score": 1.0, "content": "grounding in the LLM paradigm, Ferret employs a novel and powerful hybrid re-", "type": "text" } ], "index": 18 }, { "bbox": [ 141, 546, 470, 557 ], "spans": [ { "bbox": [ 141, 546, 470, 557 ], "score": 1.0, "content": "gion representation that integrates discrete coordinates and continuous features", "type": "text" } ], "index": 19 }, { "bbox": [ 140, 556, 470, 569 ], "spans": [ { "bbox": [ 140, 556, 470, 569 ], "score": 1.0, "content": "jointly to represent a region in the image. To extract the continuous features", "type": "text" } ], "index": 20 }, { "bbox": [ 141, 567, 469, 580 ], "spans": [ { "bbox": [ 141, 567, 469, 580 ], "score": 1.0, "content": "of versatile regions, we propose a spatial-aware visual sampler, adept at han-", "type": "text" } ], "index": 21 }, { "bbox": [ 141, 578, 470, 591 ], "spans": [ { "bbox": [ 141, 578, 470, 591 ], "score": 1.0, "content": "dling varying sparsity across different shapes. Consequently, Ferret can accept", "type": "text" } ], "index": 22 }, { "bbox": [ 141, 589, 469, 602 ], "spans": [ { "bbox": [ 141, 589, 469, 602 ], "score": 1.0, "content": "diverse region inputs, such as points, bounding boxes, and free-form shapes. To", "type": "text" } ], "index": 23 }, { "bbox": [ 141, 600, 469, 613 ], "spans": [ { "bbox": [ 141, 600, 469, 613 ], "score": 1.0, "content": "bolster the desired capability of Ferret, we curate GRIT, a comprehensive refer-", "type": "text" } ], "index": 24 }, { "bbox": [ 141, 611, 470, 624 ], "spans": [ { "bbox": [ 141, 611, 470, 624 ], "score": 1.0, "content": "and-ground instruction tuning dataset including 1.1M samples that contain rich", "type": "text" } ], "index": 25 }, { "bbox": [ 141, 622, 469, 634 ], "spans": [ { "bbox": [ 141, 622, 469, 634 ], "score": 1.0, "content": "hierarchical spatial knowledge, with 95K hard negative data to promote model", "type": "text" } ], "index": 26 }, { "bbox": [ 141, 633, 470, 646 ], "spans": [ { "bbox": [ 141, 633, 470, 646 ], "score": 1.0, "content": "robustness. The resulting model not only achieves superior performance in clas-", "type": "text" } ], "index": 27 }, { "bbox": [ 141, 644, 470, 657 ], "spans": [ { "bbox": [ 141, 644, 470, 657 ], "score": 1.0, "content": "sical referring and grounding tasks, but also greatly outperforms existing MLLMs", "type": "text" } ], "index": 28 }, { "bbox": [ 141, 655, 470, 667 ], "spans": [ { "bbox": [ 141, 655, 470, 667 ], "score": 1.0, "content": "in region-based and localization-demanded multimodal chatting. Our evaluations", "type": "text" } ], "index": 29 }, { "bbox": [ 142, 667, 470, 678 ], "spans": [ { "bbox": [ 142, 667, 470, 678 ], "score": 1.0, "content": "also reveal a significantly improved capability of describing image details and", "type": "text" } ], "index": 30 }, { "bbox": [ 141, 676, 469, 688 ], "spans": [ { "bbox": [ 141, 676, 469, 688 ], "score": 1.0, "content": "a remarkable alleviation in object hallucination. Code and data are available at", "type": "text" } ], "index": 31 }, { "bbox": [ 141, 688, 282, 700 ], "spans": [ { "bbox": [ 141, 688, 282, 700 ], "score": 1.0, "content": "https://github.com/apple/ml-ferret.", "type": "text" } ], "index": 32 } ], "index": 23.5, "bbox_fs": [ 140, 500, 470, 700 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 81, 206, 93 ], "lines": [ { "bbox": [ 105, 80, 208, 97 ], "spans": [ { "bbox": [ 105, 80, 208, 97 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 102, 505, 180 ], "lines": [ { "bbox": [ 105, 102, 505, 115 ], "spans": [ { "bbox": [ 105, 102, 505, 115 ], "score": 1.0, "content": "In vision-language learning, how to enable spatial understanding in models is a fundamental research", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 113, 505, 127 ], "spans": [ { "bbox": [ 105, 113, 505, 127 ], "score": 1.0, "content": "problem. Two desired capabilities stem from this problem: referring and grounding. Referring", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 124, 505, 137 ], "spans": [ { "bbox": [ 105, 124, 505, 137 ], "score": 1.0, "content": "demands that the model can accurately comprehend the semantics of specific given regions (Krahmer", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 136, 505, 148 ], "spans": [ { "bbox": [ 105, 136, 505, 148 ], "score": 1.0, "content": "& Van Deemter, 2012; Kazemzadeh et al., 2014; Mao et al., 2016; Yu et al., 2016; Zellers et al.,", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 146, 505, 159 ], "spans": [ { "bbox": [ 105, 146, 505, 159 ], "score": 1.0, "content": "2019), whereas grounding necessitates that the model to localize the region in accordance with the", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 157, 506, 171 ], "spans": [ { "bbox": [ 105, 157, 506, 171 ], "score": 1.0, "content": "given semantic description (Luo & Shakhnarovich, 2017; Nagaraja et al., 2016; Yu et al., 2017;", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 168, 194, 181 ], "spans": [ { "bbox": [ 105, 168, 194, 181 ], "score": 1.0, "content": "Kamath et al., 2021).", "type": "text" } ], "index": 7 } ], "index": 4 }, { "type": "text", "bbox": [ 107, 185, 505, 295 ], "lines": [ { "bbox": [ 105, 185, 506, 199 ], "spans": [ { "bbox": [ 105, 185, 506, 199 ], "score": 1.0, "content": "Essentially, referring and grounding demand the same type of knowledge: alignment of spatial", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 196, 505, 209 ], "spans": [ { "bbox": [ 105, 196, 505, 209 ], "score": 1.0, "content": "information and semantics. Despite this, existing works mostly learn referring and grounding indi-", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 207, 505, 220 ], "spans": [ { "bbox": [ 105, 207, 505, 220 ], "score": 1.0, "content": "vidually (Li et al., 2022; Wu et al., 2022; Yu et al., 2017). In comparison, humans can learn from one", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 218, 505, 231 ], "spans": [ { "bbox": [ 105, 218, 505, 231 ], "score": 1.0, "content": "task and generalize the shared knowledge to the other task effortlessly, and are able to seamlessly", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 230, 505, 242 ], "spans": [ { "bbox": [ 105, 230, 505, 242 ], "score": 1.0, "content": "integrate referring/grounding capabilities with daily dialogue and reasoning (Zellers et al., 2019).", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 239, 506, 254 ], "spans": [ { "bbox": [ 105, 239, 506, 254 ], "score": 1.0, "content": "Inspired by the above gap, in this paper, we study three main questions: (i) How to unify referring", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 251, 505, 264 ], "spans": [ { "bbox": [ 105, 251, 505, 264 ], "score": 1.0, "content": "and grounding in one framework, and will they benefit each other? (ii) How to represent versatile", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 262, 506, 275 ], "spans": [ { "bbox": [ 105, 262, 506, 275 ], "score": 1.0, "content": "types of regions that humans usually use for referring, such as point, box, scribble, and even free-", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 272, 506, 288 ], "spans": [ { "bbox": [ 105, 272, 506, 288 ], "score": 1.0, "content": "form shapes? (iii) How to make referring and grounding open-vocabulary, instruction-following,", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 285, 330, 296 ], "spans": [ { "bbox": [ 106, 285, 330, 296 ], "score": 1.0, "content": "and robust, which are crucial for practical applications?", "type": "text" } ], "index": 17 } ], "index": 12.5 }, { "type": "text", "bbox": [ 107, 302, 505, 465 ], "lines": [ { "bbox": [ 105, 300, 505, 314 ], "spans": [ { "bbox": [ 105, 300, 505, 314 ], "score": 1.0, "content": "Targeting these three questions, we introduce Ferret, a novel refer-and-ground Multimodal Large", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 312, 505, 325 ], "spans": [ { "bbox": [ 105, 312, 505, 325 ], "score": 1.0, "content": "Language Model (MLLM). First of all, we choose MLLM as the bedrock of Ferret to leverage their", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 323, 506, 336 ], "spans": [ { "bbox": [ 105, 323, 506, 336 ], "score": 1.0, "content": "powerful vision-language global understanding capability (Zhu et al., 2023a; Liu et al., 2023b; Li", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 334, 506, 347 ], "spans": [ { "bbox": [ 105, 334, 506, 347 ], "score": 1.0, "content": "et al., 2023c). To unify referring and grounding, Ferret first represents the coordinates of regions in", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 345, 505, 358 ], "spans": [ { "bbox": [ 105, 345, 505, 358 ], "score": 1.0, "content": "natural language numerical form,1 as illustrated in Figure 3. However, it is inefficient to use single", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 356, 505, 368 ], "spans": [ { "bbox": [ 105, 356, 505, 368 ], "score": 1.0, "content": "point or box coordinates to represent versatile shapes of regions, such as strokes, scribbles, or com-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 367, 504, 379 ], "spans": [ { "bbox": [ 105, 367, 504, 379 ], "score": 1.0, "content": "plex polygons. These shapes are essential for more universal and precise human-model interaction.", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 378, 504, 390 ], "spans": [ { "bbox": [ 106, 378, 504, 390 ], "score": 1.0, "content": "To solve this problem, we further propose a spatial-aware visual sampler to acquire the visual fea-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 388, 505, 402 ], "spans": [ { "bbox": [ 105, 388, 505, 402 ], "score": 1.0, "content": "tures for regions in any shape, taking care of the varying sparsity in those shapes. Then, the discrete", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 400, 505, 412 ], "spans": [ { "bbox": [ 105, 400, 505, 412 ], "score": 1.0, "content": "coordinates and the continuous visual features are combined together to represent the visual regions", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 411, 504, 423 ], "spans": [ { "bbox": [ 105, 411, 504, 423 ], "score": 1.0, "content": "in the input, composing a hybrid region representation in Ferret. Equipped with above methods,", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 422, 505, 434 ], "spans": [ { "bbox": [ 105, 422, 505, 434 ], "score": 1.0, "content": "Ferret can deal with input that mixes referred regions with free-form text, and is able to ground the", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 433, 505, 446 ], "spans": [ { "bbox": [ 105, 433, 505, 446 ], "score": 1.0, "content": "mentioned objects in its output by seamlessly generating the coordinates for each groundable object", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 444, 505, 457 ], "spans": [ { "bbox": [ 105, 444, 505, 457 ], "score": 1.0, "content": "along with generating text. To our best knowledge, Ferret is the first work that is able to process", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 455, 260, 466 ], "spans": [ { "bbox": [ 106, 455, 260, 466 ], "score": 1.0, "content": "free-formed region inputs in MLLMs.", "type": "text" } ], "index": 32 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 471, 505, 592 ], "lines": [ { "bbox": [ 105, 470, 505, 486 ], "spans": [ { "bbox": [ 105, 470, 505, 486 ], "score": 1.0, "content": "In order to make the refer-and-ground capability in Ferret open-vocabulary, instruction-following,", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 482, 505, 495 ], "spans": [ { "bbox": [ 105, 482, 505, 495 ], "score": 1.0, "content": "and robust, we collect GRIT, a Ground-and-Refer Instruction-Tuning dataset with 1.1M samples.", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 493, 504, 506 ], "spans": [ { "bbox": [ 105, 493, 504, 506 ], "score": 1.0, "content": "GRIT contains multiple levels of spatial knowledge, covering objects, relationships, region descrip-", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 504, 505, 517 ], "spans": [ { "bbox": [ 105, 504, 505, 517 ], "score": 1.0, "content": "tions, and complex reasoning. It includes both text-in location-out (grounding) and location-in text-", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 514, 505, 529 ], "spans": [ { "bbox": [ 105, 514, 505, 529 ], "score": 1.0, "content": "out (referring) data, as well as data that mixes location and text in both input and output. The ma-", "type": "text" } ], "index": 37 }, { "bbox": [ 104, 526, 506, 539 ], "spans": [ { "bbox": [ 104, 526, 506, 539 ], "score": 1.0, "content": "jority of the dataset is converted from existing vision(-language) tasks like object detection (Krishna", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 537, 505, 550 ], "spans": [ { "bbox": [ 105, 537, 505, 550 ], "score": 1.0, "content": "et al., 2017) and phrase grounding (Yu et al., 2016; Plummer et al., 2015) with carefully designed", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 547, 506, 562 ], "spans": [ { "bbox": [ 105, 547, 506, 562 ], "score": 1.0, "content": "templates to make it instruction-following. Additionally, 34K refer-and-ground instruction-tuning", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 559, 506, 573 ], "spans": [ { "bbox": [ 105, 559, 506, 573 ], "score": 1.0, "content": "conversations are collected via the help of ChatGPT/GPT-4 (OpenAI, 2023b) to facilitate training", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 570, 506, 583 ], "spans": [ { "bbox": [ 105, 570, 506, 583 ], "score": 1.0, "content": "an instruction-following and open-vocabulary refer-and-ground generalist. Moreover, we conduct", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 582, 420, 594 ], "spans": [ { "bbox": [ 106, 582, 420, 594 ], "score": 1.0, "content": "spatial-aware negative data mining, which further promotes model robustness.", "type": "text" } ], "index": 43 } ], "index": 38 }, { "type": "text", "bbox": [ 106, 598, 505, 686 ], "lines": [ { "bbox": [ 105, 597, 505, 611 ], "spans": [ { "bbox": [ 105, 597, 505, 611 ], "score": 1.0, "content": "Ferret subsumes strong open-vocabulary capabilities of spatial understanding and localization.", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 608, 506, 622 ], "spans": [ { "bbox": [ 105, 608, 506, 622 ], "score": 1.0, "content": "When evaluated on conventional referring and grounding tasks, it achieves superior performance.", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 619, 505, 633 ], "spans": [ { "bbox": [ 105, 619, 505, 633 ], "score": 1.0, "content": "More than that, we believe refer-and-ground capabilities should be integrated into daily conversa-", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 631, 506, 644 ], "spans": [ { "bbox": [ 105, 631, 506, 644 ], "score": 1.0, "content": "tions of humans, e.g., people refer to something they don’t know and ask what it is used for (like", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 642, 506, 654 ], "spans": [ { "bbox": [ 106, 642, 506, 654 ], "score": 1.0, "content": "Figure 1). To evaluate this new capability, we introduce Ferret-Bench, covering three new types of", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 652, 506, 666 ], "spans": [ { "bbox": [ 105, 652, 506, 666 ], "score": 1.0, "content": "tasks: Referring Description, Referring Reasoning, and Grounding in Conversation. We benchmark", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 663, 506, 677 ], "spans": [ { "bbox": [ 105, 663, 426, 677 ], "score": 1.0, "content": "existing MLLMs and observe that Ferret can outperform the best of them by", "type": "text" }, { "bbox": [ 426, 664, 454, 675 ], "score": 0.86, "content": "2 0 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 454, 663, 506, 677 ], "score": 1.0, "content": "on average.", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 674, 464, 687 ], "spans": [ { "bbox": [ 105, 674, 464, 687 ], "score": 1.0, "content": "Moreover, Ferret demonstrates an intriguing property of alleviating object hallucinations.", "type": "text" } ], "index": 51 } ], "index": 47.5 }, { "type": "text", "bbox": [ 109, 691, 503, 714 ], "lines": [ { "bbox": [ 105, 690, 504, 705 ], "spans": [ { "bbox": [ 105, 690, 504, 705 ], "score": 1.0, "content": "In summary, our contributions are threefold. (i) We propose Ferret, that uses a hybrid region rep-", "type": "text" } ], "index": 52 }, { "bbox": [ 106, 702, 504, 715 ], "spans": [ { "bbox": [ 106, 702, 504, 715 ], "score": 1.0, "content": "resentation equipped with a novel spatial-aware visual sampler, to enable fine-grained and open-", "type": "text" } ], "index": 53 } ], "index": 52.5 } ], "page_idx": 1, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 117, 721, 452, 731 ], "lines": [ { "bbox": [ 119, 720, 453, 733 ], "spans": [ { "bbox": [ 119, 720, 453, 733 ], "score": 1.0, "content": "1Note that there is no additional vocabulary or position encoders introduced in Ferret model.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 750, 310, 763 ], "spans": [ { "bbox": [ 302, 750, 310, 763 ], "score": 1.0, "content": "", "type": "text", "height": 13, "width": 8 } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 81, 206, 93 ], "lines": [ { "bbox": [ 105, 80, 208, 97 ], "spans": [ { "bbox": [ 105, 80, 208, 97 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 102, 505, 180 ], "lines": [ { "bbox": [ 105, 102, 505, 115 ], "spans": [ { "bbox": [ 105, 102, 505, 115 ], "score": 1.0, "content": "In vision-language learning, how to enable spatial understanding in models is a fundamental research", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 113, 505, 127 ], "spans": [ { "bbox": [ 105, 113, 505, 127 ], "score": 1.0, "content": "problem. Two desired capabilities stem from this problem: referring and grounding. Referring", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 124, 505, 137 ], "spans": [ { "bbox": [ 105, 124, 505, 137 ], "score": 1.0, "content": "demands that the model can accurately comprehend the semantics of specific given regions (Krahmer", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 136, 505, 148 ], "spans": [ { "bbox": [ 105, 136, 505, 148 ], "score": 1.0, "content": "& Van Deemter, 2012; Kazemzadeh et al., 2014; Mao et al., 2016; Yu et al., 2016; Zellers et al.,", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 146, 505, 159 ], "spans": [ { "bbox": [ 105, 146, 505, 159 ], "score": 1.0, "content": "2019), whereas grounding necessitates that the model to localize the region in accordance with the", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 157, 506, 171 ], "spans": [ { "bbox": [ 105, 157, 506, 171 ], "score": 1.0, "content": "given semantic description (Luo & Shakhnarovich, 2017; Nagaraja et al., 2016; Yu et al., 2017;", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 168, 194, 181 ], "spans": [ { "bbox": [ 105, 168, 194, 181 ], "score": 1.0, "content": "Kamath et al., 2021).", "type": "text" } ], "index": 7 } ], "index": 4, "bbox_fs": [ 105, 102, 506, 181 ] }, { "type": "text", "bbox": [ 107, 185, 505, 295 ], "lines": [ { "bbox": [ 105, 185, 506, 199 ], "spans": [ { "bbox": [ 105, 185, 506, 199 ], "score": 1.0, "content": "Essentially, referring and grounding demand the same type of knowledge: alignment of spatial", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 196, 505, 209 ], "spans": [ { "bbox": [ 105, 196, 505, 209 ], "score": 1.0, "content": "information and semantics. Despite this, existing works mostly learn referring and grounding indi-", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 207, 505, 220 ], "spans": [ { "bbox": [ 105, 207, 505, 220 ], "score": 1.0, "content": "vidually (Li et al., 2022; Wu et al., 2022; Yu et al., 2017). In comparison, humans can learn from one", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 218, 505, 231 ], "spans": [ { "bbox": [ 105, 218, 505, 231 ], "score": 1.0, "content": "task and generalize the shared knowledge to the other task effortlessly, and are able to seamlessly", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 230, 505, 242 ], "spans": [ { "bbox": [ 105, 230, 505, 242 ], "score": 1.0, "content": "integrate referring/grounding capabilities with daily dialogue and reasoning (Zellers et al., 2019).", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 239, 506, 254 ], "spans": [ { "bbox": [ 105, 239, 506, 254 ], "score": 1.0, "content": "Inspired by the above gap, in this paper, we study three main questions: (i) How to unify referring", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 251, 505, 264 ], "spans": [ { "bbox": [ 105, 251, 505, 264 ], "score": 1.0, "content": "and grounding in one framework, and will they benefit each other? (ii) How to represent versatile", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 262, 506, 275 ], "spans": [ { "bbox": [ 105, 262, 506, 275 ], "score": 1.0, "content": "types of regions that humans usually use for referring, such as point, box, scribble, and even free-", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 272, 506, 288 ], "spans": [ { "bbox": [ 105, 272, 506, 288 ], "score": 1.0, "content": "form shapes? (iii) How to make referring and grounding open-vocabulary, instruction-following,", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 285, 330, 296 ], "spans": [ { "bbox": [ 106, 285, 330, 296 ], "score": 1.0, "content": "and robust, which are crucial for practical applications?", "type": "text" } ], "index": 17 } ], "index": 12.5, "bbox_fs": [ 105, 185, 506, 296 ] }, { "type": "text", "bbox": [ 107, 302, 505, 465 ], "lines": [ { "bbox": [ 105, 300, 505, 314 ], "spans": [ { "bbox": [ 105, 300, 505, 314 ], "score": 1.0, "content": "Targeting these three questions, we introduce Ferret, a novel refer-and-ground Multimodal Large", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 312, 505, 325 ], "spans": [ { "bbox": [ 105, 312, 505, 325 ], "score": 1.0, "content": "Language Model (MLLM). First of all, we choose MLLM as the bedrock of Ferret to leverage their", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 323, 506, 336 ], "spans": [ { "bbox": [ 105, 323, 506, 336 ], "score": 1.0, "content": "powerful vision-language global understanding capability (Zhu et al., 2023a; Liu et al., 2023b; Li", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 334, 506, 347 ], "spans": [ { "bbox": [ 105, 334, 506, 347 ], "score": 1.0, "content": "et al., 2023c). To unify referring and grounding, Ferret first represents the coordinates of regions in", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 345, 505, 358 ], "spans": [ { "bbox": [ 105, 345, 505, 358 ], "score": 1.0, "content": "natural language numerical form,1 as illustrated in Figure 3. However, it is inefficient to use single", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 356, 505, 368 ], "spans": [ { "bbox": [ 105, 356, 505, 368 ], "score": 1.0, "content": "point or box coordinates to represent versatile shapes of regions, such as strokes, scribbles, or com-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 367, 504, 379 ], "spans": [ { "bbox": [ 105, 367, 504, 379 ], "score": 1.0, "content": "plex polygons. These shapes are essential for more universal and precise human-model interaction.", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 378, 504, 390 ], "spans": [ { "bbox": [ 106, 378, 504, 390 ], "score": 1.0, "content": "To solve this problem, we further propose a spatial-aware visual sampler to acquire the visual fea-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 388, 505, 402 ], "spans": [ { "bbox": [ 105, 388, 505, 402 ], "score": 1.0, "content": "tures for regions in any shape, taking care of the varying sparsity in those shapes. Then, the discrete", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 400, 505, 412 ], "spans": [ { "bbox": [ 105, 400, 505, 412 ], "score": 1.0, "content": "coordinates and the continuous visual features are combined together to represent the visual regions", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 411, 504, 423 ], "spans": [ { "bbox": [ 105, 411, 504, 423 ], "score": 1.0, "content": "in the input, composing a hybrid region representation in Ferret. Equipped with above methods,", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 422, 505, 434 ], "spans": [ { "bbox": [ 105, 422, 505, 434 ], "score": 1.0, "content": "Ferret can deal with input that mixes referred regions with free-form text, and is able to ground the", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 433, 505, 446 ], "spans": [ { "bbox": [ 105, 433, 505, 446 ], "score": 1.0, "content": "mentioned objects in its output by seamlessly generating the coordinates for each groundable object", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 444, 505, 457 ], "spans": [ { "bbox": [ 105, 444, 505, 457 ], "score": 1.0, "content": "along with generating text. To our best knowledge, Ferret is the first work that is able to process", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 455, 260, 466 ], "spans": [ { "bbox": [ 106, 455, 260, 466 ], "score": 1.0, "content": "free-formed region inputs in MLLMs.", "type": "text" } ], "index": 32 } ], "index": 25, "bbox_fs": [ 105, 300, 506, 466 ] }, { "type": "text", "bbox": [ 107, 471, 505, 592 ], "lines": [ { "bbox": [ 105, 470, 505, 486 ], "spans": [ { "bbox": [ 105, 470, 505, 486 ], "score": 1.0, "content": "In order to make the refer-and-ground capability in Ferret open-vocabulary, instruction-following,", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 482, 505, 495 ], "spans": [ { "bbox": [ 105, 482, 505, 495 ], "score": 1.0, "content": "and robust, we collect GRIT, a Ground-and-Refer Instruction-Tuning dataset with 1.1M samples.", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 493, 504, 506 ], "spans": [ { "bbox": [ 105, 493, 504, 506 ], "score": 1.0, "content": "GRIT contains multiple levels of spatial knowledge, covering objects, relationships, region descrip-", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 504, 505, 517 ], "spans": [ { "bbox": [ 105, 504, 505, 517 ], "score": 1.0, "content": "tions, and complex reasoning. It includes both text-in location-out (grounding) and location-in text-", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 514, 505, 529 ], "spans": [ { "bbox": [ 105, 514, 505, 529 ], "score": 1.0, "content": "out (referring) data, as well as data that mixes location and text in both input and output. The ma-", "type": "text" } ], "index": 37 }, { "bbox": [ 104, 526, 506, 539 ], "spans": [ { "bbox": [ 104, 526, 506, 539 ], "score": 1.0, "content": "jority of the dataset is converted from existing vision(-language) tasks like object detection (Krishna", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 537, 505, 550 ], "spans": [ { "bbox": [ 105, 537, 505, 550 ], "score": 1.0, "content": "et al., 2017) and phrase grounding (Yu et al., 2016; Plummer et al., 2015) with carefully designed", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 547, 506, 562 ], "spans": [ { "bbox": [ 105, 547, 506, 562 ], "score": 1.0, "content": "templates to make it instruction-following. Additionally, 34K refer-and-ground instruction-tuning", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 559, 506, 573 ], "spans": [ { "bbox": [ 105, 559, 506, 573 ], "score": 1.0, "content": "conversations are collected via the help of ChatGPT/GPT-4 (OpenAI, 2023b) to facilitate training", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 570, 506, 583 ], "spans": [ { "bbox": [ 105, 570, 506, 583 ], "score": 1.0, "content": "an instruction-following and open-vocabulary refer-and-ground generalist. Moreover, we conduct", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 582, 420, 594 ], "spans": [ { "bbox": [ 106, 582, 420, 594 ], "score": 1.0, "content": "spatial-aware negative data mining, which further promotes model robustness.", "type": "text" } ], "index": 43 } ], "index": 38, "bbox_fs": [ 104, 470, 506, 594 ] }, { "type": "text", "bbox": [ 106, 598, 505, 686 ], "lines": [ { "bbox": [ 105, 597, 505, 611 ], "spans": [ { "bbox": [ 105, 597, 505, 611 ], "score": 1.0, "content": "Ferret subsumes strong open-vocabulary capabilities of spatial understanding and localization.", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 608, 506, 622 ], "spans": [ { "bbox": [ 105, 608, 506, 622 ], "score": 1.0, "content": "When evaluated on conventional referring and grounding tasks, it achieves superior performance.", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 619, 505, 633 ], "spans": [ { "bbox": [ 105, 619, 505, 633 ], "score": 1.0, "content": "More than that, we believe refer-and-ground capabilities should be integrated into daily conversa-", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 631, 506, 644 ], "spans": [ { "bbox": [ 105, 631, 506, 644 ], "score": 1.0, "content": "tions of humans, e.g., people refer to something they don’t know and ask what it is used for (like", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 642, 506, 654 ], "spans": [ { "bbox": [ 106, 642, 506, 654 ], "score": 1.0, "content": "Figure 1). To evaluate this new capability, we introduce Ferret-Bench, covering three new types of", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 652, 506, 666 ], "spans": [ { "bbox": [ 105, 652, 506, 666 ], "score": 1.0, "content": "tasks: Referring Description, Referring Reasoning, and Grounding in Conversation. We benchmark", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 663, 506, 677 ], "spans": [ { "bbox": [ 105, 663, 426, 677 ], "score": 1.0, "content": "existing MLLMs and observe that Ferret can outperform the best of them by", "type": "text" }, { "bbox": [ 426, 664, 454, 675 ], "score": 0.86, "content": "2 0 . 4 \\%", "type": "inline_equation" }, { "bbox": [ 454, 663, 506, 677 ], "score": 1.0, "content": "on average.", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 674, 464, 687 ], "spans": [ { "bbox": [ 105, 674, 464, 687 ], "score": 1.0, "content": "Moreover, Ferret demonstrates an intriguing property of alleviating object hallucinations.", "type": "text" } ], "index": 51 } ], "index": 47.5, "bbox_fs": [ 105, 597, 506, 687 ] }, { "type": "text", "bbox": [ 109, 691, 503, 714 ], "lines": [ { "bbox": [ 105, 690, 504, 705 ], "spans": [ { "bbox": [ 105, 690, 504, 705 ], "score": 1.0, "content": "In summary, our contributions are threefold. (i) We propose Ferret, that uses a hybrid region rep-", "type": "text" } ], "index": 52 }, { "bbox": [ 106, 702, 504, 715 ], "spans": [ { "bbox": [ 106, 702, 504, 715 ], "score": 1.0, "content": "resentation equipped with a novel spatial-aware visual sampler, to enable fine-grained and open-", "type": "text" } ], "index": 53 }, { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "vocabulary referring and grounding in MLLM. (ii) We construct GRIT, a large-scale ground-and-", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 105, 93, 505, 107 ], "spans": [ { "bbox": [ 105, 93, 505, 107 ], "score": 1.0, "content": "refer instruction tuning dataset, for model training. It also contains additional spatial negative sam-", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 104, 505, 118 ], "spans": [ { "bbox": [ 105, 104, 505, 118 ], "score": 1.0, "content": "ples to enhance model robustness. (iii) We introduce Ferret-Bench, to evaluate tasks jointly re-", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 105, 115, 506, 129 ], "spans": [ { "bbox": [ 105, 115, 506, 129 ], "score": 1.0, "content": "quiring referring/grounding, semantics, knowledge, and reasoning. Our model exhibits superior", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 105, 127, 389, 139 ], "spans": [ { "bbox": [ 105, 127, 389, 139 ], "score": 1.0, "content": "performance in a wide range of tasks and reduces object hallucination.", "type": "text", "cross_page": true } ], "index": 4 } ], "index": 52.5, "bbox_fs": [ 105, 690, 504, 715 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 504, 138 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "vocabulary referring and grounding in MLLM. (ii) We construct GRIT, a large-scale ground-and-", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 505, 107 ], "spans": [ { "bbox": [ 105, 93, 505, 107 ], "score": 1.0, "content": "refer instruction tuning dataset, for model training. It also contains additional spatial negative sam-", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 505, 118 ], "spans": [ { "bbox": [ 105, 104, 505, 118 ], "score": 1.0, "content": "ples to enhance model robustness. (iii) We introduce Ferret-Bench, to evaluate tasks jointly re-", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 506, 129 ], "spans": [ { "bbox": [ 105, 115, 506, 129 ], "score": 1.0, "content": "quiring referring/grounding, semantics, knowledge, and reasoning. Our model exhibits superior", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 127, 389, 139 ], "spans": [ { "bbox": [ 105, 127, 389, 139 ], "score": 1.0, "content": "performance in a wide range of tasks and reduces object hallucination.", "type": "text" } ], "index": 4 } ], "index": 2 }, { "type": "title", "bbox": [ 108, 154, 173, 167 ], "lines": [ { "bbox": [ 104, 152, 175, 170 ], "spans": [ { "bbox": [ 104, 152, 175, 170 ], "score": 1.0, "content": "2 METHOD", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 107, 179, 504, 201 ], "lines": [ { "bbox": [ 105, 178, 505, 192 ], "spans": [ { "bbox": [ 105, 178, 505, 192 ], "score": 1.0, "content": "We start with detailing the proposed hybrid region representation to depict regions of various shapes", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 189, 361, 202 ], "spans": [ { "bbox": [ 106, 189, 361, 202 ], "score": 1.0, "content": "and formats. Then, we present the model architecture of Ferret.", "type": "text" } ], "index": 7 } ], "index": 6.5 }, { "type": "title", "bbox": [ 108, 216, 282, 227 ], "lines": [ { "bbox": [ 106, 215, 283, 228 ], "spans": [ { "bbox": [ 106, 215, 283, 228 ], "score": 1.0, "content": "2.1 HYBRID REGION REPRESENTATION", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 107, 232, 403, 352 ], "lines": [ { "bbox": [ 106, 232, 404, 244 ], "spans": [ { "bbox": [ 106, 232, 404, 244 ], "score": 1.0, "content": "When referring to specific regions, three primary formats are generally", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 243, 404, 255 ], "spans": [ { "bbox": [ 106, 243, 404, 255 ], "score": 1.0, "content": "used: point, box, and free-form shapes. While the point and box for-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 254, 404, 267 ], "spans": [ { "bbox": [ 105, 254, 332, 267 ], "score": 1.0, "content": "mats can be succinctly represented by coordinates (e.g.,", "type": "text" }, { "bbox": [ 332, 254, 356, 265 ], "score": 0.54, "content": "[ x , y ]", "type": "inline_equation" }, { "bbox": [ 356, 254, 404, 267 ], "score": 1.0, "content": "for a point,", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 264, 404, 277 ], "spans": [ { "bbox": [ 105, 264, 124, 277 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 125, 266, 214, 276 ], "score": 0.78, "content": "[ x _ { \\mathrm { m i n } } , y _ { \\mathrm { m i n } } , x _ { \\mathrm { m a x } } , y _ { \\mathrm { m a x } } ]", "type": "inline_equation" }, { "bbox": [ 214, 264, 404, 277 ], "score": 1.0, "content": "for a box) as in Peng et al. (2023); Chen et al.", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 275, 404, 289 ], "spans": [ { "bbox": [ 105, 275, 404, 289 ], "score": 1.0, "content": "(2023b), the free-form shape is more versatile, encompassing a variety", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 286, 404, 299 ], "spans": [ { "bbox": [ 105, 286, 404, 299 ], "score": 1.0, "content": "of region types such as scribbles, polygons, and masks. The advantage", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 297, 404, 310 ], "spans": [ { "bbox": [ 105, 297, 404, 310 ], "score": 1.0, "content": "of free-form shape is straightforwardly illustrated in Figure 2. Depict-", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 308, 405, 321 ], "spans": [ { "bbox": [ 106, 308, 405, 321 ], "score": 1.0, "content": "ing free-form shapes through coordinates is computationally expensive and", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 319, 404, 331 ], "spans": [ { "bbox": [ 106, 319, 404, 331 ], "score": 1.0, "content": "obscure, and its complexity hinders the model learning to establish a clear", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 329, 404, 343 ], "spans": [ { "bbox": [ 105, 329, 404, 343 ], "score": 1.0, "content": "correlation between the provided coordinates and the corresponding re-", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 341, 134, 354 ], "spans": [ { "bbox": [ 106, 341, 134, 354 ], "score": 1.0, "content": "gions.", "type": "text" } ], "index": 19 } ], "index": 14 }, { "type": "text", "bbox": [ 107, 358, 403, 461 ], "lines": [ { "bbox": [ 106, 357, 404, 371 ], "spans": [ { "bbox": [ 106, 357, 404, 371 ], "score": 1.0, "content": "To generalize across all three distinct formats, we propose a hybrid region", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 370, 403, 381 ], "spans": [ { "bbox": [ 106, 370, 403, 381 ], "score": 1.0, "content": "representation that synergizes discrete coordinates with continuous visual", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 379, 404, 394 ], "spans": [ { "bbox": [ 105, 379, 404, 394 ], "score": 1.0, "content": "features to refer to a particular region, which is shown in the top-left of Fig-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 390, 404, 405 ], "spans": [ { "bbox": [ 105, 390, 404, 405 ], "score": 1.0, "content": "ure 3. For coordinates, following Chen et al. (2021); Yang et al. (2022),", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 401, 405, 416 ], "spans": [ { "bbox": [ 105, 401, 280, 416 ], "score": 1.0, "content": "we quantize each coordinate into one of the", "type": "text" }, { "bbox": [ 280, 403, 300, 413 ], "score": 0.88, "content": "n _ { \\mathrm { b i n s } }", "type": "inline_equation" }, { "bbox": [ 300, 401, 405, 416 ], "score": 1.0, "content": "discrete bins.2 Regarding", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 414, 403, 425 ], "spans": [ { "bbox": [ 106, 414, 294, 425 ], "score": 1.0, "content": "continuous visual features, for a given region", "type": "text" }, { "bbox": [ 294, 414, 303, 424 ], "score": 0.51, "content": "\\mathbf { R }", "type": "inline_equation" }, { "bbox": [ 304, 414, 403, 425 ], "score": 1.0, "content": ", we first construct a 2D", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 424, 404, 436 ], "spans": [ { "bbox": [ 106, 424, 404, 436 ], "score": 1.0, "content": "binary mask M of the same size as the image, marking a value of 1 in-", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 434, 404, 448 ], "spans": [ { "bbox": [ 106, 434, 404, 448 ], "score": 1.0, "content": "side the targeted region and 0 outside of the region. Then, the binary mask", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 446, 404, 459 ], "spans": [ { "bbox": [ 106, 446, 307, 459 ], "score": 1.0, "content": "M, jointly with the extracted image feature map", "type": "text" }, { "bbox": [ 308, 446, 316, 456 ], "score": 0.52, "content": "\\mathbf { Z }", "type": "inline_equation" }, { "bbox": [ 316, 446, 404, 459 ], "score": 1.0, "content": ", is sent into our pro-", "type": "text" } ], "index": 29 } ], "index": 25 }, { "type": "image", "bbox": [ 412, 232, 503, 356 ], "blocks": [ { "type": "image_body", "bbox": [ 412, 232, 503, 356 ], "group_id": 0, "lines": [ { "bbox": [ 412, 232, 503, 356 ], "spans": [ { "bbox": [ 412, 232, 503, 356 ], "score": 0.945, "type": "image", "image_path": "814d338d9332a576dcba58c12b5035c6780590dbd09fe60291b13f10e93ffff4.jpg" } ] } ], "index": 20, "virtual_lines": [ { "bbox": [ 412, 232, 503, 356 ], "spans": [], "index": 20 } ] }, { "type": "image_caption", "bbox": [ 411, 358, 504, 457 ], "group_id": 0, "lines": [ { "bbox": [ 411, 357, 505, 369 ], "spans": [ { "bbox": [ 411, 357, 505, 369 ], "score": 1.0, "content": "Figure 2: Bounding box", "type": "text" } ], "index": 30 }, { "bbox": [ 411, 366, 505, 379 ], "spans": [ { "bbox": [ 411, 366, 505, 379 ], "score": 1.0, "content": "v.s. Free-from Shape.", "type": "text" } ], "index": 31 }, { "bbox": [ 410, 376, 505, 389 ], "spans": [ { "bbox": [ 410, 376, 505, 389 ], "score": 1.0, "content": "These two objects have", "type": "text" } ], "index": 32 }, { "bbox": [ 410, 387, 505, 399 ], "spans": [ { "bbox": [ 410, 387, 505, 399 ], "score": 1.0, "content": "almost the same bound-", "type": "text" } ], "index": 33 }, { "bbox": [ 410, 396, 505, 409 ], "spans": [ { "bbox": [ 410, 396, 505, 409 ], "score": 1.0, "content": "ing box, causing ambigu-", "type": "text" } ], "index": 34 }, { "bbox": [ 410, 407, 506, 419 ], "spans": [ { "bbox": [ 410, 407, 506, 419 ], "score": 1.0, "content": "ity when relying on the", "type": "text" } ], "index": 35 }, { "bbox": [ 410, 417, 506, 429 ], "spans": [ { "bbox": [ 410, 417, 506, 429 ], "score": 1.0, "content": "box to refer to. Equipped", "type": "text" } ], "index": 36 }, { "bbox": [ 410, 426, 506, 439 ], "spans": [ { "bbox": [ 410, 426, 506, 439 ], "score": 1.0, "content": "with hybrid region repre-", "type": "text" } ], "index": 37 }, { "bbox": [ 410, 436, 505, 449 ], "spans": [ { "bbox": [ 410, 436, 505, 449 ], "score": 1.0, "content": "sentation, Ferret can sep-", "type": "text" } ], "index": 38 }, { "bbox": [ 410, 446, 454, 458 ], "spans": [ { "bbox": [ 410, 446, 454, 458 ], "score": 1.0, "content": "arate them.", "type": "text" } ], "index": 39 } ], "index": 34.5 } ], "index": 27.25 }, { "type": "text", "bbox": [ 108, 457, 504, 479 ], "lines": [ { "bbox": [ 105, 456, 505, 469 ], "spans": [ { "bbox": [ 105, 456, 249, 469 ], "score": 1.0, "content": "posed spatial-aware visual sampler", "type": "text" }, { "bbox": [ 250, 457, 266, 469 ], "score": 0.9, "content": "s ( \\cdot )", "type": "inline_equation" }, { "bbox": [ 266, 456, 505, 469 ], "score": 1.0, "content": ", which will be detailed in Section 2.2, to extract the visual", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 468, 240, 480 ], "spans": [ { "bbox": [ 106, 468, 183, 480 ], "score": 1.0, "content": "continuous feature", "type": "text" }, { "bbox": [ 183, 468, 236, 480 ], "score": 0.92, "content": "\\mathbf { f } = s ( \\mathbf { M } , \\mathbf { Z } )", "type": "inline_equation" }, { "bbox": [ 237, 468, 240, 480 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 41 } ], "index": 40.5 }, { "type": "text", "bbox": [ 107, 484, 505, 531 ], "lines": [ { "bbox": [ 105, 484, 505, 498 ], "spans": [ { "bbox": [ 105, 484, 241, 498 ], "score": 1.0, "content": "Finally, we represent a point with", "type": "text" }, { "bbox": [ 241, 485, 285, 498 ], "score": 0.93, "content": "\\{ x , y , \\mathbf { f } _ { R _ { p } } \\}", "type": "inline_equation" }, { "bbox": [ 286, 484, 357, 498 ], "score": 1.0, "content": ", where the region", "type": "text" }, { "bbox": [ 357, 486, 371, 497 ], "score": 0.89, "content": "R _ { p }", "type": "inline_equation" }, { "bbox": [ 371, 484, 457, 498 ], "score": 1.0, "content": "is a circle centered in", "type": "text" }, { "bbox": [ 458, 485, 484, 497 ], "score": 0.93, "content": "\\{ x , y \\}", "type": "inline_equation" }, { "bbox": [ 484, 484, 505, 498 ], "score": 1.0, "content": "with", "type": "text" } ], "index": 42 }, { "bbox": [ 104, 496, 505, 511 ], "spans": [ { "bbox": [ 104, 496, 383, 511 ], "score": 1.0, "content": "a fixed radius.3 A box or a free-form shape can both be represented by", "type": "text" }, { "bbox": [ 383, 497, 501, 510 ], "score": 0.87, "content": "\\{ x _ { \\mathrm { m i n } } , y _ { \\mathrm { m i n } } , x _ { \\mathrm { m a x } } , y _ { \\mathrm { m a x } } , \\mathbf { f } _ { R _ { b o x } } \\}", "type": "inline_equation" }, { "bbox": [ 501, 496, 505, 511 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 509, 505, 521 ], "spans": [ { "bbox": [ 106, 509, 133, 521 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 509, 173, 520 ], "score": 0.91, "content": "{ x _ { \\operatorname* { m i n } } } / { x _ { \\operatorname* { m a x } } }", "type": "inline_equation" }, { "bbox": [ 173, 509, 308, 521 ], "score": 1.0, "content": "denotes the minimum/maximum", "type": "text" }, { "bbox": [ 309, 511, 316, 519 ], "score": 0.75, "content": "x", "type": "inline_equation" }, { "bbox": [ 316, 509, 505, 521 ], "score": 1.0, "content": "-axis coordinate of the region, and so forth for", "type": "text" } ], "index": 44 }, { "bbox": [ 107, 520, 259, 532 ], "spans": [ { "bbox": [ 107, 522, 113, 531 ], "score": 0.8, "content": "y", "type": "inline_equation" }, { "bbox": [ 113, 520, 136, 532 ], "score": 1.0, "content": "-axis.", "type": "text" }, { "bbox": [ 137, 520, 158, 531 ], "score": 0.9, "content": "R _ { b o x }", "type": "inline_equation" }, { "bbox": [ 158, 520, 259, 532 ], "score": 1.0, "content": "denotes the input region.", "type": "text" } ], "index": 45 } ], "index": 43.5 }, { "type": "title", "bbox": [ 108, 542, 235, 553 ], "lines": [ { "bbox": [ 106, 542, 236, 555 ], "spans": [ { "bbox": [ 106, 542, 236, 555 ], "score": 1.0, "content": "2.2 MODEL ARCHITECTURE", "type": "text" } ], "index": 46 } ], "index": 46 }, { "type": "text", "bbox": [ 107, 559, 503, 592 ], "lines": [ { "bbox": [ 105, 557, 505, 573 ], "spans": [ { "bbox": [ 105, 557, 345, 573 ], "score": 1.0, "content": "As illustrated in Figure 3, Ferret is mainly composed of", "type": "text" }, { "bbox": [ 345, 560, 356, 570 ], "score": 0.42, "content": "( i )", "type": "inline_equation" }, { "bbox": [ 356, 557, 505, 573 ], "score": 1.0, "content": "an image encoder to extract image", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 570, 505, 583 ], "spans": [ { "bbox": [ 106, 570, 160, 583 ], "score": 1.0, "content": "embeddings,", "type": "text" }, { "bbox": [ 160, 570, 174, 581 ], "score": 0.68, "content": "( i i )", "type": "inline_equation" }, { "bbox": [ 175, 570, 505, 583 ], "score": 1.0, "content": "the proposed spatial-aware visual sampler to extract regional continuous features,", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 581, 375, 594 ], "spans": [ { "bbox": [ 105, 581, 123, 594 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 124, 581, 141, 592 ], "score": 0.67, "content": "( i i i )", "type": "inline_equation" }, { "bbox": [ 142, 581, 375, 594 ], "score": 1.0, "content": "an LLM to jointly model image, text, and region features.", "type": "text" } ], "index": 49 } ], "index": 48 }, { "type": "text", "bbox": [ 106, 597, 505, 685 ], "lines": [ { "bbox": [ 105, 597, 505, 610 ], "spans": [ { "bbox": [ 105, 597, 505, 610 ], "score": 1.0, "content": "Input. We feed the image into a pre-trained visual encoder, CLIP-ViT-L/14 (Radford et al., 2021),", "type": "text" } ], "index": 50 }, { "bbox": [ 104, 605, 506, 622 ], "spans": [ { "bbox": [ 104, 605, 239, 622 ], "score": 1.0, "content": "to extract the image embeddings", "type": "text" }, { "bbox": [ 240, 608, 302, 619 ], "score": 0.92, "content": "\\mathbf { \\dot { Z } } \\in \\mathbb { R } ^ { H \\times W \\times C }", "type": "inline_equation" }, { "bbox": [ 303, 605, 506, 622 ], "score": 1.0, "content": ". For text, we tokenize the text sequence using the", "type": "text" } ], "index": 51 }, { "bbox": [ 104, 619, 506, 632 ], "spans": [ { "bbox": [ 104, 619, 384, 632 ], "score": 1.0, "content": "pre-trained LLM’s tokenizer and project them into text embeddings", "type": "text" }, { "bbox": [ 385, 619, 435, 630 ], "score": 0.92, "content": "\\mathbf { T } \\in \\mathbb { R } ^ { L \\times D }", "type": "inline_equation" }, { "bbox": [ 435, 619, 506, 632 ], "score": 1.0, "content": ". As for referred", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 631, 506, 643 ], "spans": [ { "bbox": [ 105, 631, 506, 643 ], "score": 1.0, "content": "regions, we append the coordinates and a special token as a placeholder for continuous features after", "type": "text" } ], "index": 53 }, { "bbox": [ 105, 641, 505, 654 ], "spans": [ { "bbox": [ 105, 641, 327, 654 ], "score": 1.0, "content": "the name of the region: “⟨region name⟩ ⟨coordinates⟩", "type": "text" }, { "bbox": [ 328, 641, 357, 654 ], "score": 0.8, "content": "\\langle { \\mathrm { S P E } } \\rangle ^ { \\mathrm { , , } }", "type": "inline_equation" }, { "bbox": [ 358, 641, 505, 654 ], "score": 1.0, "content": ". For example, “a cat [100, 50, 200,", "type": "text" } ], "index": 54 }, { "bbox": [ 105, 651, 506, 666 ], "spans": [ { "bbox": [ 105, 651, 128, 666 ], "score": 1.0, "content": "300]", "type": "text" }, { "bbox": [ 128, 652, 158, 664 ], "score": 0.72, "content": "\\langle \\mathrm { S P E } \\rangle ^ { \\mathrm { , , } }", "type": "inline_equation" }, { "bbox": [ 158, 651, 506, 666 ], "score": 1.0, "content": ". If the name is unknown or hard to describe because multiple objects are included, we", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 663, 506, 677 ], "spans": [ { "bbox": [ 105, 663, 506, 677 ], "score": 1.0, "content": "just use “region” or “area” as the “⟨region name⟩”. In this way, referred regions can be well mixed", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 674, 299, 687 ], "spans": [ { "bbox": [ 105, 674, 299, 687 ], "score": 1.0, "content": "with ordinary texts to form complete sentences.", "type": "text" } ], "index": 57 } ], "index": 53.5 } ], "page_idx": 2, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 700, 504, 731 ], "lines": [ { "bbox": [ 121, 698, 506, 714 ], "spans": [ { "bbox": [ 121, 700, 172, 710 ], "score": 0.84, "content": "^ 2 n _ { \\mathrm { b i n s } } = 1 0 0 0", "type": "inline_equation" }, { "bbox": [ 172, 698, 506, 714 ], "score": 1.0, "content": "by default. The value is input invariant, which means for any input image size, the original", "type": "text" } ] }, { "bbox": [ 105, 710, 502, 722 ], "spans": [ { "bbox": [ 105, 710, 502, 722 ], "score": 1.0, "content": "coordinate will be mapped to the new coordinates. This makes the model robust to different input resolutions.", "type": "text" } ] }, { "bbox": [ 118, 720, 227, 733 ], "spans": [ { "bbox": [ 118, 720, 227, 733 ], "score": 1.0, "content": "3Radius is set to 5 by default.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 760 ], "lines": [ { "bbox": [ 301, 750, 310, 762 ], "spans": [ { "bbox": [ 301, 750, 310, 762 ], "score": 1.0, "content": "3", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 504, 138 ], "lines": [], "index": 2, "bbox_fs": [ 105, 82, 506, 139 ], "lines_deleted": true }, { "type": "title", "bbox": [ 108, 154, 173, 167 ], "lines": [ { "bbox": [ 104, 152, 175, 170 ], "spans": [ { "bbox": [ 104, 152, 175, 170 ], "score": 1.0, "content": "2 METHOD", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 107, 179, 504, 201 ], "lines": [ { "bbox": [ 105, 178, 505, 192 ], "spans": [ { "bbox": [ 105, 178, 505, 192 ], "score": 1.0, "content": "We start with detailing the proposed hybrid region representation to depict regions of various shapes", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 189, 361, 202 ], "spans": [ { "bbox": [ 106, 189, 361, 202 ], "score": 1.0, "content": "and formats. Then, we present the model architecture of Ferret.", "type": "text" } ], "index": 7 } ], "index": 6.5, "bbox_fs": [ 105, 178, 505, 202 ] }, { "type": "title", "bbox": [ 108, 216, 282, 227 ], "lines": [ { "bbox": [ 106, 215, 283, 228 ], "spans": [ { "bbox": [ 106, 215, 283, 228 ], "score": 1.0, "content": "2.1 HYBRID REGION REPRESENTATION", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 107, 232, 403, 352 ], "lines": [ { "bbox": [ 106, 232, 404, 244 ], "spans": [ { "bbox": [ 106, 232, 404, 244 ], "score": 1.0, "content": "When referring to specific regions, three primary formats are generally", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 243, 404, 255 ], "spans": [ { "bbox": [ 106, 243, 404, 255 ], "score": 1.0, "content": "used: point, box, and free-form shapes. While the point and box for-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 254, 404, 267 ], "spans": [ { "bbox": [ 105, 254, 332, 267 ], "score": 1.0, "content": "mats can be succinctly represented by coordinates (e.g.,", "type": "text" }, { "bbox": [ 332, 254, 356, 265 ], "score": 0.54, "content": "[ x , y ]", "type": "inline_equation" }, { "bbox": [ 356, 254, 404, 267 ], "score": 1.0, "content": "for a point,", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 264, 404, 277 ], "spans": [ { "bbox": [ 105, 264, 124, 277 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 125, 266, 214, 276 ], "score": 0.78, "content": "[ x _ { \\mathrm { m i n } } , y _ { \\mathrm { m i n } } , x _ { \\mathrm { m a x } } , y _ { \\mathrm { m a x } } ]", "type": "inline_equation" }, { "bbox": [ 214, 264, 404, 277 ], "score": 1.0, "content": "for a box) as in Peng et al. (2023); Chen et al.", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 275, 404, 289 ], "spans": [ { "bbox": [ 105, 275, 404, 289 ], "score": 1.0, "content": "(2023b), the free-form shape is more versatile, encompassing a variety", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 286, 404, 299 ], "spans": [ { "bbox": [ 105, 286, 404, 299 ], "score": 1.0, "content": "of region types such as scribbles, polygons, and masks. The advantage", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 297, 404, 310 ], "spans": [ { "bbox": [ 105, 297, 404, 310 ], "score": 1.0, "content": "of free-form shape is straightforwardly illustrated in Figure 2. Depict-", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 308, 405, 321 ], "spans": [ { "bbox": [ 106, 308, 405, 321 ], "score": 1.0, "content": "ing free-form shapes through coordinates is computationally expensive and", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 319, 404, 331 ], "spans": [ { "bbox": [ 106, 319, 404, 331 ], "score": 1.0, "content": "obscure, and its complexity hinders the model learning to establish a clear", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 329, 404, 343 ], "spans": [ { "bbox": [ 105, 329, 404, 343 ], "score": 1.0, "content": "correlation between the provided coordinates and the corresponding re-", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 341, 134, 354 ], "spans": [ { "bbox": [ 106, 341, 134, 354 ], "score": 1.0, "content": "gions.", "type": "text" } ], "index": 19 } ], "index": 14, "bbox_fs": [ 105, 232, 405, 354 ] }, { "type": "text", "bbox": [ 107, 358, 403, 461 ], "lines": [ { "bbox": [ 106, 357, 404, 371 ], "spans": [ { "bbox": [ 106, 357, 404, 371 ], "score": 1.0, "content": "To generalize across all three distinct formats, we propose a hybrid region", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 370, 403, 381 ], "spans": [ { "bbox": [ 106, 370, 403, 381 ], "score": 1.0, "content": "representation that synergizes discrete coordinates with continuous visual", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 379, 404, 394 ], "spans": [ { "bbox": [ 105, 379, 404, 394 ], "score": 1.0, "content": "features to refer to a particular region, which is shown in the top-left of Fig-", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 390, 404, 405 ], "spans": [ { "bbox": [ 105, 390, 404, 405 ], "score": 1.0, "content": "ure 3. For coordinates, following Chen et al. (2021); Yang et al. (2022),", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 401, 405, 416 ], "spans": [ { "bbox": [ 105, 401, 280, 416 ], "score": 1.0, "content": "we quantize each coordinate into one of the", "type": "text" }, { "bbox": [ 280, 403, 300, 413 ], "score": 0.88, "content": "n _ { \\mathrm { b i n s } }", "type": "inline_equation" }, { "bbox": [ 300, 401, 405, 416 ], "score": 1.0, "content": "discrete bins.2 Regarding", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 414, 403, 425 ], "spans": [ { "bbox": [ 106, 414, 294, 425 ], "score": 1.0, "content": "continuous visual features, for a given region", "type": "text" }, { "bbox": [ 294, 414, 303, 424 ], "score": 0.51, "content": "\\mathbf { R }", "type": "inline_equation" }, { "bbox": [ 304, 414, 403, 425 ], "score": 1.0, "content": ", we first construct a 2D", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 424, 404, 436 ], "spans": [ { "bbox": [ 106, 424, 404, 436 ], "score": 1.0, "content": "binary mask M of the same size as the image, marking a value of 1 in-", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 434, 404, 448 ], "spans": [ { "bbox": [ 106, 434, 404, 448 ], "score": 1.0, "content": "side the targeted region and 0 outside of the region. Then, the binary mask", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 446, 404, 459 ], "spans": [ { "bbox": [ 106, 446, 307, 459 ], "score": 1.0, "content": "M, jointly with the extracted image feature map", "type": "text" }, { "bbox": [ 308, 446, 316, 456 ], "score": 0.52, "content": "\\mathbf { Z }", "type": "inline_equation" }, { "bbox": [ 316, 446, 404, 459 ], "score": 1.0, "content": ", is sent into our pro-", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 456, 505, 469 ], "spans": [ { "bbox": [ 105, 456, 249, 469 ], "score": 1.0, "content": "posed spatial-aware visual sampler", "type": "text" }, { "bbox": [ 250, 457, 266, 469 ], "score": 0.9, "content": "s ( \\cdot )", "type": "inline_equation" }, { "bbox": [ 266, 456, 505, 469 ], "score": 1.0, "content": ", which will be detailed in Section 2.2, to extract the visual", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 468, 240, 480 ], "spans": [ { "bbox": [ 106, 468, 183, 480 ], "score": 1.0, "content": "continuous feature", "type": "text" }, { "bbox": [ 183, 468, 236, 480 ], "score": 0.92, "content": "\\mathbf { f } = s ( \\mathbf { M } , \\mathbf { Z } )", "type": "inline_equation" }, { "bbox": [ 237, 468, 240, 480 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 41 } ], "index": 25, "bbox_fs": [ 105, 357, 405, 459 ] }, { "type": "image", "bbox": [ 412, 232, 503, 356 ], "blocks": [ { "type": "image_body", "bbox": [ 412, 232, 503, 356 ], "group_id": 0, "lines": [ { "bbox": [ 412, 232, 503, 356 ], "spans": [ { "bbox": [ 412, 232, 503, 356 ], "score": 0.945, "type": "image", "image_path": "814d338d9332a576dcba58c12b5035c6780590dbd09fe60291b13f10e93ffff4.jpg" } ] } ], "index": 20, "virtual_lines": [ { "bbox": [ 412, 232, 503, 356 ], "spans": [], "index": 20 } ] }, { "type": "image_caption", "bbox": [ 411, 358, 504, 457 ], "group_id": 0, "lines": [ { "bbox": [ 411, 357, 505, 369 ], "spans": [ { "bbox": [ 411, 357, 505, 369 ], "score": 1.0, "content": "Figure 2: Bounding box", "type": "text" } ], "index": 30 }, { "bbox": [ 411, 366, 505, 379 ], "spans": [ { "bbox": [ 411, 366, 505, 379 ], "score": 1.0, "content": "v.s. Free-from Shape.", "type": "text" } ], "index": 31 }, { "bbox": [ 410, 376, 505, 389 ], "spans": [ { "bbox": [ 410, 376, 505, 389 ], "score": 1.0, "content": "These two objects have", "type": "text" } ], "index": 32 }, { "bbox": [ 410, 387, 505, 399 ], "spans": [ { "bbox": [ 410, 387, 505, 399 ], "score": 1.0, "content": "almost the same bound-", "type": "text" } ], "index": 33 }, { "bbox": [ 410, 396, 505, 409 ], "spans": [ { "bbox": [ 410, 396, 505, 409 ], "score": 1.0, "content": "ing box, causing ambigu-", "type": "text" } ], "index": 34 }, { "bbox": [ 410, 407, 506, 419 ], "spans": [ { "bbox": [ 410, 407, 506, 419 ], "score": 1.0, "content": "ity when relying on the", "type": "text" } ], "index": 35 }, { "bbox": [ 410, 417, 506, 429 ], "spans": [ { "bbox": [ 410, 417, 506, 429 ], "score": 1.0, "content": "box to refer to. Equipped", "type": "text" } ], "index": 36 }, { "bbox": [ 410, 426, 506, 439 ], "spans": [ { "bbox": [ 410, 426, 506, 439 ], "score": 1.0, "content": "with hybrid region repre-", "type": "text" } ], "index": 37 }, { "bbox": [ 410, 436, 505, 449 ], "spans": [ { "bbox": [ 410, 436, 505, 449 ], "score": 1.0, "content": "sentation, Ferret can sep-", "type": "text" } ], "index": 38 }, { "bbox": [ 410, 446, 454, 458 ], "spans": [ { "bbox": [ 410, 446, 454, 458 ], "score": 1.0, "content": "arate them.", "type": "text" } ], "index": 39 } ], "index": 34.5 } ], "index": 27.25 }, { "type": "text", "bbox": [ 108, 457, 504, 479 ], "lines": [], "index": 40.5, "bbox_fs": [ 105, 456, 505, 480 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 484, 505, 531 ], "lines": [ { "bbox": [ 105, 484, 505, 498 ], "spans": [ { "bbox": [ 105, 484, 241, 498 ], "score": 1.0, "content": "Finally, we represent a point with", "type": "text" }, { "bbox": [ 241, 485, 285, 498 ], "score": 0.93, "content": "\\{ x , y , \\mathbf { f } _ { R _ { p } } \\}", "type": "inline_equation" }, { "bbox": [ 286, 484, 357, 498 ], "score": 1.0, "content": ", where the region", "type": "text" }, { "bbox": [ 357, 486, 371, 497 ], "score": 0.89, "content": "R _ { p }", "type": "inline_equation" }, { "bbox": [ 371, 484, 457, 498 ], "score": 1.0, "content": "is a circle centered in", "type": "text" }, { "bbox": [ 458, 485, 484, 497 ], "score": 0.93, "content": "\\{ x , y \\}", "type": "inline_equation" }, { "bbox": [ 484, 484, 505, 498 ], "score": 1.0, "content": "with", "type": "text" } ], "index": 42 }, { "bbox": [ 104, 496, 505, 511 ], "spans": [ { "bbox": [ 104, 496, 383, 511 ], "score": 1.0, "content": "a fixed radius.3 A box or a free-form shape can both be represented by", "type": "text" }, { "bbox": [ 383, 497, 501, 510 ], "score": 0.87, "content": "\\{ x _ { \\mathrm { m i n } } , y _ { \\mathrm { m i n } } , x _ { \\mathrm { m a x } } , y _ { \\mathrm { m a x } } , \\mathbf { f } _ { R _ { b o x } } \\}", "type": "inline_equation" }, { "bbox": [ 501, 496, 505, 511 ], "score": 1.0, "content": ",", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 509, 505, 521 ], "spans": [ { "bbox": [ 106, 509, 133, 521 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 134, 509, 173, 520 ], "score": 0.91, "content": "{ x _ { \\operatorname* { m i n } } } / { x _ { \\operatorname* { m a x } } }", "type": "inline_equation" }, { "bbox": [ 173, 509, 308, 521 ], "score": 1.0, "content": "denotes the minimum/maximum", "type": "text" }, { "bbox": [ 309, 511, 316, 519 ], "score": 0.75, "content": "x", "type": "inline_equation" }, { "bbox": [ 316, 509, 505, 521 ], "score": 1.0, "content": "-axis coordinate of the region, and so forth for", "type": "text" } ], "index": 44 }, { "bbox": [ 107, 520, 259, 532 ], "spans": [ { "bbox": [ 107, 522, 113, 531 ], "score": 0.8, "content": "y", "type": "inline_equation" }, { "bbox": [ 113, 520, 136, 532 ], "score": 1.0, "content": "-axis.", "type": "text" }, { "bbox": [ 137, 520, 158, 531 ], "score": 0.9, "content": "R _ { b o x }", "type": "inline_equation" }, { "bbox": [ 158, 520, 259, 532 ], "score": 1.0, "content": "denotes the input region.", "type": "text" } ], "index": 45 } ], "index": 43.5, "bbox_fs": [ 104, 484, 505, 532 ] }, { "type": "title", "bbox": [ 108, 542, 235, 553 ], "lines": [ { "bbox": [ 106, 542, 236, 555 ], "spans": [ { "bbox": [ 106, 542, 236, 555 ], "score": 1.0, "content": "2.2 MODEL ARCHITECTURE", "type": "text" } ], "index": 46 } ], "index": 46 }, { "type": "text", "bbox": [ 107, 559, 503, 592 ], "lines": [ { "bbox": [ 105, 557, 505, 573 ], "spans": [ { "bbox": [ 105, 557, 345, 573 ], "score": 1.0, "content": "As illustrated in Figure 3, Ferret is mainly composed of", "type": "text" }, { "bbox": [ 345, 560, 356, 570 ], "score": 0.42, "content": "( i )", "type": "inline_equation" }, { "bbox": [ 356, 557, 505, 573 ], "score": 1.0, "content": "an image encoder to extract image", "type": "text" } ], "index": 47 }, { "bbox": [ 106, 570, 505, 583 ], "spans": [ { "bbox": [ 106, 570, 160, 583 ], "score": 1.0, "content": "embeddings,", "type": "text" }, { "bbox": [ 160, 570, 174, 581 ], "score": 0.68, "content": "( i i )", "type": "inline_equation" }, { "bbox": [ 175, 570, 505, 583 ], "score": 1.0, "content": "the proposed spatial-aware visual sampler to extract regional continuous features,", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 581, 375, 594 ], "spans": [ { "bbox": [ 105, 581, 123, 594 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 124, 581, 141, 592 ], "score": 0.67, "content": "( i i i )", "type": "inline_equation" }, { "bbox": [ 142, 581, 375, 594 ], "score": 1.0, "content": "an LLM to jointly model image, text, and region features.", "type": "text" } ], "index": 49 } ], "index": 48, "bbox_fs": [ 105, 557, 505, 594 ] }, { "type": "text", "bbox": [ 106, 597, 505, 685 ], "lines": [ { "bbox": [ 105, 597, 505, 610 ], "spans": [ { "bbox": [ 105, 597, 505, 610 ], "score": 1.0, "content": "Input. We feed the image into a pre-trained visual encoder, CLIP-ViT-L/14 (Radford et al., 2021),", "type": "text" } ], "index": 50 }, { "bbox": [ 104, 605, 506, 622 ], "spans": [ { "bbox": [ 104, 605, 239, 622 ], "score": 1.0, "content": "to extract the image embeddings", "type": "text" }, { "bbox": [ 240, 608, 302, 619 ], "score": 0.92, "content": "\\mathbf { \\dot { Z } } \\in \\mathbb { R } ^ { H \\times W \\times C }", "type": "inline_equation" }, { "bbox": [ 303, 605, 506, 622 ], "score": 1.0, "content": ". For text, we tokenize the text sequence using the", "type": "text" } ], "index": 51 }, { "bbox": [ 104, 619, 506, 632 ], "spans": [ { "bbox": [ 104, 619, 384, 632 ], "score": 1.0, "content": "pre-trained LLM’s tokenizer and project them into text embeddings", "type": "text" }, { "bbox": [ 385, 619, 435, 630 ], "score": 0.92, "content": "\\mathbf { T } \\in \\mathbb { R } ^ { L \\times D }", "type": "inline_equation" }, { "bbox": [ 435, 619, 506, 632 ], "score": 1.0, "content": ". As for referred", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 631, 506, 643 ], "spans": [ { "bbox": [ 105, 631, 506, 643 ], "score": 1.0, "content": "regions, we append the coordinates and a special token as a placeholder for continuous features after", "type": "text" } ], "index": 53 }, { "bbox": [ 105, 641, 505, 654 ], "spans": [ { "bbox": [ 105, 641, 327, 654 ], "score": 1.0, "content": "the name of the region: “⟨region name⟩ ⟨coordinates⟩", "type": "text" }, { "bbox": [ 328, 641, 357, 654 ], "score": 0.8, "content": "\\langle { \\mathrm { S P E } } \\rangle ^ { \\mathrm { , , } }", "type": "inline_equation" }, { "bbox": [ 358, 641, 505, 654 ], "score": 1.0, "content": ". For example, “a cat [100, 50, 200,", "type": "text" } ], "index": 54 }, { "bbox": [ 105, 651, 506, 666 ], "spans": [ { "bbox": [ 105, 651, 128, 666 ], "score": 1.0, "content": "300]", "type": "text" }, { "bbox": [ 128, 652, 158, 664 ], "score": 0.72, "content": "\\langle \\mathrm { S P E } \\rangle ^ { \\mathrm { , , } }", "type": "inline_equation" }, { "bbox": [ 158, 651, 506, 666 ], "score": 1.0, "content": ". If the name is unknown or hard to describe because multiple objects are included, we", "type": "text" } ], "index": 55 }, { "bbox": [ 105, 663, 506, 677 ], "spans": [ { "bbox": [ 105, 663, 506, 677 ], "score": 1.0, "content": "just use “region” or “area” as the “⟨region name⟩”. In this way, referred regions can be well mixed", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 674, 299, 687 ], "spans": [ { "bbox": [ 105, 674, 299, 687 ], "score": 1.0, "content": "with ordinary texts to form complete sentences.", "type": "text" } ], "index": 57 } ], "index": 53.5, "bbox_fs": [ 104, 597, 506, 687 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 106, 83, 503, 235 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 83, 503, 235 ], "group_id": 0, "lines": [ { "bbox": [ 106, 83, 503, 235 ], "spans": [ { "bbox": [ 106, 83, 503, 235 ], "score": 0.974, "type": "image", "image_path": "2e74618c412c130b6ef8a3756501208aa3e5f026e792a59a88eb0a93bbd2a57f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 83, 503, 133.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 106, 133.66666666666666, 503, 184.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 106, 184.33333333333331, 503, 234.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 239, 504, 271 ], "group_id": 0, "lines": [ { "bbox": [ 105, 237, 505, 251 ], "spans": [ { "bbox": [ 105, 237, 505, 251 ], "score": 1.0, "content": "Figure 3: Overview of the proposed Ferret model architecture. (Left) The proposed hybrid region", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 249, 505, 262 ], "spans": [ { "bbox": [ 105, 249, 505, 262 ], "score": 1.0, "content": "representation and spatial-aware visual sampler. (Right) Overall model architecture. All parameters", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 261, 268, 272 ], "spans": [ { "bbox": [ 106, 261, 268, 272 ], "score": 1.0, "content": "besides the image encoder are trainable.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 106, 275, 505, 330 ], "lines": [ { "bbox": [ 105, 274, 505, 287 ], "spans": [ { "bbox": [ 105, 274, 505, 287 ], "score": 1.0, "content": "Spatial-aware Visual Sampler. The shape of the referred regions can be quite varied, not limited", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 286, 505, 298 ], "spans": [ { "bbox": [ 105, 286, 505, 298 ], "score": 1.0, "content": "to just points or rectangle boxes. Grid-based processing like convolution or patch attention cannot", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 297, 505, 309 ], "spans": [ { "bbox": [ 106, 297, 505, 309 ], "score": 1.0, "content": "handle irregular shapes. Similar to our cases, 3D point clouds are also in irregular shape and show", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 308, 506, 321 ], "spans": [ { "bbox": [ 106, 308, 506, 321 ], "score": 1.0, "content": "varied sparsity in the 3D space. Inspired by existing works in 3D point cloud learning (Qi et al.,", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 317, 452, 332 ], "spans": [ { "bbox": [ 105, 317, 452, 332 ], "score": 1.0, "content": "2017a; Ma et al., 2022; Wang et al., 2019), we propose a spatial-aware visual sampler.", "type": "text" } ], "index": 10 } ], "index": 8 }, { "type": "text", "bbox": [ 106, 334, 505, 426 ], "lines": [ { "bbox": [ 105, 332, 506, 349 ], "spans": [ { "bbox": [ 105, 332, 246, 349 ], "score": 1.0, "content": "Given extracted image feature map", "type": "text" }, { "bbox": [ 246, 334, 308, 346 ], "score": 0.92, "content": "\\mathbf { Z } \\in \\mathbb { R } ^ { H \\times W \\times C }", "type": "inline_equation" }, { "bbox": [ 308, 332, 506, 349 ], "score": 1.0, "content": "and the binary region mask M, we first randomly", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 347, 505, 359 ], "spans": [ { "bbox": [ 106, 347, 137, 359 ], "score": 1.0, "content": "sample", "type": "text" }, { "bbox": [ 137, 347, 148, 357 ], "score": 0.81, "content": "N", "type": "inline_equation" }, { "bbox": [ 148, 347, 505, 359 ], "score": 1.0, "content": "positive points inside M. For each point, its feature is obtained by bilinear interpolation.", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 356, 505, 371 ], "spans": [ { "bbox": [ 105, 356, 124, 371 ], "score": 1.0, "content": "The", "type": "text" }, { "bbox": [ 125, 358, 135, 368 ], "score": 0.81, "content": "N", "type": "inline_equation" }, { "bbox": [ 136, 356, 505, 371 ], "score": 1.0, "content": "points are fed into a cascade of blocks, where each of them includes three steps: sampling,", "type": "text" } ], "index": 13 }, { "bbox": [ 104, 367, 506, 384 ], "spans": [ { "bbox": [ 104, 367, 245, 384 ], "score": 1.0, "content": "gathering, pooling. (1) Sampling:", "type": "text" }, { "bbox": [ 245, 368, 255, 381 ], "score": 0.87, "content": "\\textstyle { \\frac { N } { r } }", "type": "inline_equation" }, { "bbox": [ 256, 367, 357, 384 ], "score": 1.0, "content": "points are sampled from", "type": "text" }, { "bbox": [ 357, 369, 367, 379 ], "score": 0.8, "content": "N", "type": "inline_equation" }, { "bbox": [ 367, 367, 506, 384 ], "score": 1.0, "content": "points via farthest point sampling", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 381, 506, 394 ], "spans": [ { "bbox": [ 105, 381, 506, 394 ], "score": 1.0, "content": "(FPS) algorithm (Qi et al., 2017b),4 which can guarantee sufficient coverage. (2) Gathering: For", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 392, 505, 405 ], "spans": [ { "bbox": [ 105, 392, 214, 405 ], "score": 1.0, "content": "each of the sampled points", "type": "text" }, { "bbox": [ 214, 394, 224, 403 ], "score": 0.85, "content": "x _ { i }", "type": "inline_equation" }, { "bbox": [ 225, 392, 281, 405 ], "score": 1.0, "content": ", we search its", "type": "text" }, { "bbox": [ 281, 393, 288, 402 ], "score": 0.81, "content": "k", "type": "inline_equation" }, { "bbox": [ 289, 392, 464, 405 ], "score": 1.0, "content": "nearest neighbors from the pool of previous", "type": "text" }, { "bbox": [ 465, 393, 474, 402 ], "score": 0.82, "content": "N", "type": "inline_equation" }, { "bbox": [ 475, 392, 505, 405 ], "score": 1.0, "content": "points,", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 402, 506, 417 ], "spans": [ { "bbox": [ 105, 402, 223, 417 ], "score": 1.0, "content": "and obtain a group of points", "type": "text" }, { "bbox": [ 223, 403, 295, 415 ], "score": 0.92, "content": "\\{ x _ { i 1 } , x _ { i 2 } , . . . , x _ { i k } \\}", "type": "inline_equation" }, { "bbox": [ 295, 402, 506, 417 ], "score": 1.0, "content": ". Then, inspired by PointMLP (Ma et al., 2022), for", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 414, 423, 427 ], "spans": [ { "bbox": [ 105, 414, 307, 427 ], "score": 1.0, "content": "each group, we fuse the features of sampled point", "type": "text" }, { "bbox": [ 307, 416, 317, 425 ], "score": 0.87, "content": "x _ { i }", "type": "inline_equation" }, { "bbox": [ 317, 414, 423, 427 ], "score": 1.0, "content": "and it neighbor points by:", "type": "text" } ], "index": 18 } ], "index": 14.5 }, { "type": "interline_equation", "bbox": [ 169, 429, 431, 442 ], "lines": [ { "bbox": [ 169, 429, 431, 442 ], "spans": [ { "bbox": [ 169, 429, 431, 442 ], "score": 0.87, "content": "h _ { i k } = \\sigma ( [ \\theta ( [ \\mathbf { Z } ( x _ { i k } ) - \\mathbf { Z } ( x _ { i } ) ; C ( x _ { i k } ) - C ( x _ { i } ) ] ) ; \\mathbf { Z } ( x _ { i } ) ; C ( x _ { i } ) ] ) ,", "type": "interline_equation", "image_path": "059228fb9168f9566faf7b42d37a3d5182dbf8b08139a407c1d89cc7e45e3e15.jpg" } ] } ], "index": 19, "virtual_lines": [ { "bbox": [ 169, 429, 431, 442 ], "spans": [], "index": 19 } ] }, { "type": "text", "bbox": [ 106, 446, 505, 523 ], "lines": [ { "bbox": [ 106, 446, 505, 459 ], "spans": [ { "bbox": [ 106, 446, 135, 459 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 135, 448, 149, 458 ], "score": 0.86, "content": "x _ { i k }", "type": "inline_equation" }, { "bbox": [ 150, 446, 264, 459 ], "score": 1.0, "content": "is one of the neighbors of", "type": "text" }, { "bbox": [ 264, 448, 275, 457 ], "score": 0.68, "content": "x _ { i }", "type": "inline_equation" }, { "bbox": [ 275, 446, 280, 459 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 281, 446, 302, 459 ], "score": 0.83, "content": "\\mathbf { Z } ( x )", "type": "inline_equation" }, { "bbox": [ 302, 446, 379, 459 ], "score": 1.0, "content": "denotes the point", "type": "text" }, { "bbox": [ 379, 448, 387, 456 ], "score": 0.78, "content": "x", "type": "inline_equation" }, { "bbox": [ 387, 446, 505, 459 ], "score": 1.0, "content": "’s feature (in the first block,", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 457, 505, 470 ], "spans": [ { "bbox": [ 105, 457, 251, 470 ], "score": 1.0, "content": "it is interpolated from feature map", "type": "text" }, { "bbox": [ 252, 458, 260, 468 ], "score": 0.68, "content": "\\mathbf { Z }", "type": "inline_equation" }, { "bbox": [ 260, 457, 505, 470 ], "score": 1.0, "content": "; in the succeeding blocks, it is the output feature from the", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 468, 505, 481 ], "spans": [ { "bbox": [ 105, 468, 173, 481 ], "score": 1.0, "content": "previous block),", "type": "text" }, { "bbox": [ 174, 469, 195, 480 ], "score": 0.91, "content": "C ( x )", "type": "inline_equation" }, { "bbox": [ 196, 468, 340, 481 ], "score": 1.0, "content": "denotes the 2D coordinates of point", "type": "text" }, { "bbox": [ 341, 468, 363, 480 ], "score": 0.28, "content": "x , [ ; ]", "type": "inline_equation" }, { "bbox": [ 363, 468, 505, 481 ], "score": 1.0, "content": "means channel-wise concatenation", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 480, 505, 491 ], "spans": [ { "bbox": [ 106, 480, 187, 491 ], "score": 1.0, "content": "of multiple vectors,", "type": "text" }, { "bbox": [ 188, 480, 194, 489 ], "score": 0.81, "content": "\\theta", "type": "inline_equation" }, { "bbox": [ 194, 480, 487, 491 ], "score": 1.0, "content": "is implemented by a linear layer to adapt the relative local features, and", "type": "text" }, { "bbox": [ 487, 481, 495, 489 ], "score": 0.77, "content": "\\sigma", "type": "inline_equation" }, { "bbox": [ 495, 480, 505, 491 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 489, 505, 504 ], "spans": [ { "bbox": [ 105, 489, 505, 504 ], "score": 1.0, "content": "also a linear layer to fuse each local feature from neighbors with sampled point feature. (3) Pooling:", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 501, 505, 513 ], "spans": [ { "bbox": [ 105, 501, 250, 513 ], "score": 1.0, "content": "A max pooling is conducted to fuse", "type": "text" }, { "bbox": [ 250, 502, 257, 511 ], "score": 0.82, "content": "k", "type": "inline_equation" }, { "bbox": [ 258, 501, 505, 513 ], "score": 1.0, "content": "neighbor features into one feature as the representation of the", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 513, 168, 525 ], "spans": [ { "bbox": [ 105, 513, 168, 525 ], "score": 1.0, "content": "sampled point:", "type": "text" } ], "index": 26 } ], "index": 23 }, { "type": "interline_equation", "bbox": [ 250, 522, 360, 542 ], "lines": [ { "bbox": [ 250, 522, 360, 542 ], "spans": [ { "bbox": [ 250, 522, 360, 542 ], "score": 0.93, "content": "h _ { i } = \\operatorname* { m a x } _ { k : ( x _ { i k } ) \\in \\mathrm { K N N s o f } x _ { i } } h _ { i k } .", "type": "interline_equation", "image_path": "69fd2d4e808ada90534a632eefe211e4667be67e8b1b868f5e6c5d7e7e185e2e.jpg" } ] } ], "index": 27, "virtual_lines": [ { "bbox": [ 250, 522, 360, 542 ], "spans": [], "index": 27 } ] }, { "type": "text", "bbox": [ 107, 543, 505, 599 ], "lines": [ { "bbox": [ 106, 544, 505, 556 ], "spans": [ { "bbox": [ 106, 544, 505, 556 ], "score": 1.0, "content": "After the three steps, we obtain fewer points but a more dense feature space since it incorporates the", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 554, 505, 567 ], "spans": [ { "bbox": [ 105, 554, 434, 567 ], "score": 1.0, "content": "local neighbor features as well as their relative positions. In experiments, we set", "type": "text" }, { "bbox": [ 435, 555, 465, 565 ], "score": 0.85, "content": "N { = } 5 1 2", "type": "inline_equation" }, { "bbox": [ 466, 554, 469, 567 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 470, 555, 487, 565 ], "score": 0.82, "content": "r { = } 4", "type": "inline_equation" }, { "bbox": [ 487, 554, 505, 567 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 565, 505, 577 ], "spans": [ { "bbox": [ 106, 565, 129, 576 ], "score": 0.85, "content": "k { = } 2 4", "type": "inline_equation" }, { "bbox": [ 129, 565, 505, 577 ], "score": 1.0, "content": ", and cascade two such blocks, which in the end outputs 32 points with their features. Similar", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 577, 505, 589 ], "spans": [ { "bbox": [ 105, 577, 505, 589 ], "score": 1.0, "content": "to ROIAlign (He et al., 2017), we flatten the point features into a single vector and project it to the", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 587, 495, 600 ], "spans": [ { "bbox": [ 105, 587, 393, 600 ], "score": 1.0, "content": "dimension of LLM embeddings. The final feature is used to replace the", "type": "text" }, { "bbox": [ 394, 587, 419, 599 ], "score": 0.61, "content": "\\langle { \\mathrm { S P E } } \\rangle", "type": "inline_equation" }, { "bbox": [ 419, 587, 495, 600 ], "score": 1.0, "content": "token in the input.", "type": "text" } ], "index": 32 } ], "index": 30 }, { "type": "text", "bbox": [ 107, 604, 505, 660 ], "lines": [ { "bbox": [ 106, 605, 504, 617 ], "spans": [ { "bbox": [ 106, 605, 504, 617 ], "score": 1.0, "content": "Output. The above region denotations are used in Ferret input to refer to specific regions. In", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 615, 505, 628 ], "spans": [ { "bbox": [ 105, 615, 505, 628 ], "score": 1.0, "content": "Ferret output, to achieve grounding, we generate the box coordinates right after the corresponding", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 626, 504, 639 ], "spans": [ { "bbox": [ 105, 626, 504, 639 ], "score": 1.0, "content": "regions/nouns in the text response. For instance, “There is a dog [100, 150, 300, 200] in the figure.”", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 636, 505, 650 ], "spans": [ { "bbox": [ 105, 636, 505, 650 ], "score": 1.0, "content": "With this data format, our model is expected to implicitly learn what is groundable in the current", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 647, 249, 662 ], "spans": [ { "bbox": [ 105, 647, 249, 662 ], "score": 1.0, "content": "image and what their locations are.", "type": "text" } ], "index": 37 } ], "index": 35 }, { "type": "text", "bbox": [ 107, 664, 505, 709 ], "lines": [ { "bbox": [ 105, 664, 505, 677 ], "spans": [ { "bbox": [ 105, 664, 505, 677 ], "score": 1.0, "content": "LLM. We consider Vicuna (Chiang et al., 2023) as our language model, a decoder-only", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 676, 505, 688 ], "spans": [ { "bbox": [ 105, 676, 505, 688 ], "score": 1.0, "content": "LLM (Brown et al., 2020) that is instruction-tuned on top of LLaMA (Touvron et al., 2023a). Prior", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "to being fed into the LLM, the image embeddings undergo transformation via an additional linear", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 699, 345, 710 ], "spans": [ { "bbox": [ 106, 699, 345, 710 ], "score": 1.0, "content": "layer to match the embedding dimension of the text tokens.", "type": "text" } ], "index": 41 } ], "index": 39.5 }, { "type": "text", "bbox": [ 106, 711, 505, 732 ], "lines": [ { "bbox": [ 118, 709, 505, 724 ], "spans": [ { "bbox": [ 118, 709, 312, 724 ], "score": 1.0, "content": "4FPS starts from a random single point sampled from", "type": "text" }, { "bbox": [ 312, 714, 321, 720 ], "score": 0.83, "content": "N", "type": "inline_equation" }, { "bbox": [ 321, 709, 505, 724 ], "score": 1.0, "content": "points. In each iteration, it samples one point from", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 721, 504, 732 ], "spans": [ { "bbox": [ 106, 721, 504, 732 ], "score": 1.0, "content": "the rest points such that it is the farthest from the set of already sampled points. See detail in Qi et al. (2017b).", "type": "text" } ], "index": 43 } ], "index": 42.5 } ], "page_idx": 3, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 293, 38 ], "spans": [ { "bbox": [ 106, 26, 293, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 759 ], "lines": [] } ], "para_blocks": [ { "type": "image", "bbox": [ 106, 83, 503, 235 ], "blocks": [ { "type": "image_body", "bbox": [ 106, 83, 503, 235 ], "group_id": 0, "lines": [ { "bbox": [ 106, 83, 503, 235 ], "spans": [ { "bbox": [ 106, 83, 503, 235 ], "score": 0.974, "type": "image", "image_path": "2e74618c412c130b6ef8a3756501208aa3e5f026e792a59a88eb0a93bbd2a57f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 106, 83, 503, 133.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 106, 133.66666666666666, 503, 184.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 106, 184.33333333333331, 503, 234.99999999999997 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 239, 504, 271 ], "group_id": 0, "lines": [ { "bbox": [ 105, 237, 505, 251 ], "spans": [ { "bbox": [ 105, 237, 505, 251 ], "score": 1.0, "content": "Figure 3: Overview of the proposed Ferret model architecture. (Left) The proposed hybrid region", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 249, 505, 262 ], "spans": [ { "bbox": [ 105, 249, 505, 262 ], "score": 1.0, "content": "representation and spatial-aware visual sampler. (Right) Overall model architecture. All parameters", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 261, 268, 272 ], "spans": [ { "bbox": [ 106, 261, 268, 272 ], "score": 1.0, "content": "besides the image encoder are trainable.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "text", "bbox": [ 106, 275, 505, 330 ], "lines": [ { "bbox": [ 105, 274, 505, 287 ], "spans": [ { "bbox": [ 105, 274, 505, 287 ], "score": 1.0, "content": "Spatial-aware Visual Sampler. The shape of the referred regions can be quite varied, not limited", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 286, 505, 298 ], "spans": [ { "bbox": [ 105, 286, 505, 298 ], "score": 1.0, "content": "to just points or rectangle boxes. Grid-based processing like convolution or patch attention cannot", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 297, 505, 309 ], "spans": [ { "bbox": [ 106, 297, 505, 309 ], "score": 1.0, "content": "handle irregular shapes. Similar to our cases, 3D point clouds are also in irregular shape and show", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 308, 506, 321 ], "spans": [ { "bbox": [ 106, 308, 506, 321 ], "score": 1.0, "content": "varied sparsity in the 3D space. Inspired by existing works in 3D point cloud learning (Qi et al.,", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 317, 452, 332 ], "spans": [ { "bbox": [ 105, 317, 452, 332 ], "score": 1.0, "content": "2017a; Ma et al., 2022; Wang et al., 2019), we propose a spatial-aware visual sampler.", "type": "text" } ], "index": 10 } ], "index": 8, "bbox_fs": [ 105, 274, 506, 332 ] }, { "type": "text", "bbox": [ 106, 334, 505, 426 ], "lines": [ { "bbox": [ 105, 332, 506, 349 ], "spans": [ { "bbox": [ 105, 332, 246, 349 ], "score": 1.0, "content": "Given extracted image feature map", "type": "text" }, { "bbox": [ 246, 334, 308, 346 ], "score": 0.92, "content": "\\mathbf { Z } \\in \\mathbb { R } ^ { H \\times W \\times C }", "type": "inline_equation" }, { "bbox": [ 308, 332, 506, 349 ], "score": 1.0, "content": "and the binary region mask M, we first randomly", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 347, 505, 359 ], "spans": [ { "bbox": [ 106, 347, 137, 359 ], "score": 1.0, "content": "sample", "type": "text" }, { "bbox": [ 137, 347, 148, 357 ], "score": 0.81, "content": "N", "type": "inline_equation" }, { "bbox": [ 148, 347, 505, 359 ], "score": 1.0, "content": "positive points inside M. For each point, its feature is obtained by bilinear interpolation.", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 356, 505, 371 ], "spans": [ { "bbox": [ 105, 356, 124, 371 ], "score": 1.0, "content": "The", "type": "text" }, { "bbox": [ 125, 358, 135, 368 ], "score": 0.81, "content": "N", "type": "inline_equation" }, { "bbox": [ 136, 356, 505, 371 ], "score": 1.0, "content": "points are fed into a cascade of blocks, where each of them includes three steps: sampling,", "type": "text" } ], "index": 13 }, { "bbox": [ 104, 367, 506, 384 ], "spans": [ { "bbox": [ 104, 367, 245, 384 ], "score": 1.0, "content": "gathering, pooling. (1) Sampling:", "type": "text" }, { "bbox": [ 245, 368, 255, 381 ], "score": 0.87, "content": "\\textstyle { \\frac { N } { r } }", "type": "inline_equation" }, { "bbox": [ 256, 367, 357, 384 ], "score": 1.0, "content": "points are sampled from", "type": "text" }, { "bbox": [ 357, 369, 367, 379 ], "score": 0.8, "content": "N", "type": "inline_equation" }, { "bbox": [ 367, 367, 506, 384 ], "score": 1.0, "content": "points via farthest point sampling", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 381, 506, 394 ], "spans": [ { "bbox": [ 105, 381, 506, 394 ], "score": 1.0, "content": "(FPS) algorithm (Qi et al., 2017b),4 which can guarantee sufficient coverage. (2) Gathering: For", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 392, 505, 405 ], "spans": [ { "bbox": [ 105, 392, 214, 405 ], "score": 1.0, "content": "each of the sampled points", "type": "text" }, { "bbox": [ 214, 394, 224, 403 ], "score": 0.85, "content": "x _ { i }", "type": "inline_equation" }, { "bbox": [ 225, 392, 281, 405 ], "score": 1.0, "content": ", we search its", "type": "text" }, { "bbox": [ 281, 393, 288, 402 ], "score": 0.81, "content": "k", "type": "inline_equation" }, { "bbox": [ 289, 392, 464, 405 ], "score": 1.0, "content": "nearest neighbors from the pool of previous", "type": "text" }, { "bbox": [ 465, 393, 474, 402 ], "score": 0.82, "content": "N", "type": "inline_equation" }, { "bbox": [ 475, 392, 505, 405 ], "score": 1.0, "content": "points,", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 402, 506, 417 ], "spans": [ { "bbox": [ 105, 402, 223, 417 ], "score": 1.0, "content": "and obtain a group of points", "type": "text" }, { "bbox": [ 223, 403, 295, 415 ], "score": 0.92, "content": "\\{ x _ { i 1 } , x _ { i 2 } , . . . , x _ { i k } \\}", "type": "inline_equation" }, { "bbox": [ 295, 402, 506, 417 ], "score": 1.0, "content": ". Then, inspired by PointMLP (Ma et al., 2022), for", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 414, 423, 427 ], "spans": [ { "bbox": [ 105, 414, 307, 427 ], "score": 1.0, "content": "each group, we fuse the features of sampled point", "type": "text" }, { "bbox": [ 307, 416, 317, 425 ], "score": 0.87, "content": "x _ { i }", "type": "inline_equation" }, { "bbox": [ 317, 414, 423, 427 ], "score": 1.0, "content": "and it neighbor points by:", "type": "text" } ], "index": 18 } ], "index": 14.5, "bbox_fs": [ 104, 332, 506, 427 ] }, { "type": "interline_equation", "bbox": [ 169, 429, 431, 442 ], "lines": [ { "bbox": [ 169, 429, 431, 442 ], "spans": [ { "bbox": [ 169, 429, 431, 442 ], "score": 0.87, "content": "h _ { i k } = \\sigma ( [ \\theta ( [ \\mathbf { Z } ( x _ { i k } ) - \\mathbf { Z } ( x _ { i } ) ; C ( x _ { i k } ) - C ( x _ { i } ) ] ) ; \\mathbf { Z } ( x _ { i } ) ; C ( x _ { i } ) ] ) ,", "type": "interline_equation", "image_path": "059228fb9168f9566faf7b42d37a3d5182dbf8b08139a407c1d89cc7e45e3e15.jpg" } ] } ], "index": 19, "virtual_lines": [ { "bbox": [ 169, 429, 431, 442 ], "spans": [], "index": 19 } ] }, { "type": "text", "bbox": [ 106, 446, 505, 523 ], "lines": [ { "bbox": [ 106, 446, 505, 459 ], "spans": [ { "bbox": [ 106, 446, 135, 459 ], "score": 1.0, "content": "where", "type": "text" }, { "bbox": [ 135, 448, 149, 458 ], "score": 0.86, "content": "x _ { i k }", "type": "inline_equation" }, { "bbox": [ 150, 446, 264, 459 ], "score": 1.0, "content": "is one of the neighbors of", "type": "text" }, { "bbox": [ 264, 448, 275, 457 ], "score": 0.68, "content": "x _ { i }", "type": "inline_equation" }, { "bbox": [ 275, 446, 280, 459 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 281, 446, 302, 459 ], "score": 0.83, "content": "\\mathbf { Z } ( x )", "type": "inline_equation" }, { "bbox": [ 302, 446, 379, 459 ], "score": 1.0, "content": "denotes the point", "type": "text" }, { "bbox": [ 379, 448, 387, 456 ], "score": 0.78, "content": "x", "type": "inline_equation" }, { "bbox": [ 387, 446, 505, 459 ], "score": 1.0, "content": "’s feature (in the first block,", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 457, 505, 470 ], "spans": [ { "bbox": [ 105, 457, 251, 470 ], "score": 1.0, "content": "it is interpolated from feature map", "type": "text" }, { "bbox": [ 252, 458, 260, 468 ], "score": 0.68, "content": "\\mathbf { Z }", "type": "inline_equation" }, { "bbox": [ 260, 457, 505, 470 ], "score": 1.0, "content": "; in the succeeding blocks, it is the output feature from the", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 468, 505, 481 ], "spans": [ { "bbox": [ 105, 468, 173, 481 ], "score": 1.0, "content": "previous block),", "type": "text" }, { "bbox": [ 174, 469, 195, 480 ], "score": 0.91, "content": "C ( x )", "type": "inline_equation" }, { "bbox": [ 196, 468, 340, 481 ], "score": 1.0, "content": "denotes the 2D coordinates of point", "type": "text" }, { "bbox": [ 341, 468, 363, 480 ], "score": 0.28, "content": "x , [ ; ]", "type": "inline_equation" }, { "bbox": [ 363, 468, 505, 481 ], "score": 1.0, "content": "means channel-wise concatenation", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 480, 505, 491 ], "spans": [ { "bbox": [ 106, 480, 187, 491 ], "score": 1.0, "content": "of multiple vectors,", "type": "text" }, { "bbox": [ 188, 480, 194, 489 ], "score": 0.81, "content": "\\theta", "type": "inline_equation" }, { "bbox": [ 194, 480, 487, 491 ], "score": 1.0, "content": "is implemented by a linear layer to adapt the relative local features, and", "type": "text" }, { "bbox": [ 487, 481, 495, 489 ], "score": 0.77, "content": "\\sigma", "type": "inline_equation" }, { "bbox": [ 495, 480, 505, 491 ], "score": 1.0, "content": "is", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 489, 505, 504 ], "spans": [ { "bbox": [ 105, 489, 505, 504 ], "score": 1.0, "content": "also a linear layer to fuse each local feature from neighbors with sampled point feature. (3) Pooling:", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 501, 505, 513 ], "spans": [ { "bbox": [ 105, 501, 250, 513 ], "score": 1.0, "content": "A max pooling is conducted to fuse", "type": "text" }, { "bbox": [ 250, 502, 257, 511 ], "score": 0.82, "content": "k", "type": "inline_equation" }, { "bbox": [ 258, 501, 505, 513 ], "score": 1.0, "content": "neighbor features into one feature as the representation of the", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 513, 168, 525 ], "spans": [ { "bbox": [ 105, 513, 168, 525 ], "score": 1.0, "content": "sampled point:", "type": "text" } ], "index": 26 } ], "index": 23, "bbox_fs": [ 105, 446, 505, 525 ] }, { "type": "interline_equation", "bbox": [ 250, 522, 360, 542 ], "lines": [ { "bbox": [ 250, 522, 360, 542 ], "spans": [ { "bbox": [ 250, 522, 360, 542 ], "score": 0.93, "content": "h _ { i } = \\operatorname* { m a x } _ { k : ( x _ { i k } ) \\in \\mathrm { K N N s o f } x _ { i } } h _ { i k } .", "type": "interline_equation", "image_path": "69fd2d4e808ada90534a632eefe211e4667be67e8b1b868f5e6c5d7e7e185e2e.jpg" } ] } ], "index": 27, "virtual_lines": [ { "bbox": [ 250, 522, 360, 542 ], "spans": [], "index": 27 } ] }, { "type": "text", "bbox": [ 107, 543, 505, 599 ], "lines": [ { "bbox": [ 106, 544, 505, 556 ], "spans": [ { "bbox": [ 106, 544, 505, 556 ], "score": 1.0, "content": "After the three steps, we obtain fewer points but a more dense feature space since it incorporates the", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 554, 505, 567 ], "spans": [ { "bbox": [ 105, 554, 434, 567 ], "score": 1.0, "content": "local neighbor features as well as their relative positions. In experiments, we set", "type": "text" }, { "bbox": [ 435, 555, 465, 565 ], "score": 0.85, "content": "N { = } 5 1 2", "type": "inline_equation" }, { "bbox": [ 466, 554, 469, 567 ], "score": 1.0, "content": ",", "type": "text" }, { "bbox": [ 470, 555, 487, 565 ], "score": 0.82, "content": "r { = } 4", "type": "inline_equation" }, { "bbox": [ 487, 554, 505, 567 ], "score": 1.0, "content": "and", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 565, 505, 577 ], "spans": [ { "bbox": [ 106, 565, 129, 576 ], "score": 0.85, "content": "k { = } 2 4", "type": "inline_equation" }, { "bbox": [ 129, 565, 505, 577 ], "score": 1.0, "content": ", and cascade two such blocks, which in the end outputs 32 points with their features. Similar", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 577, 505, 589 ], "spans": [ { "bbox": [ 105, 577, 505, 589 ], "score": 1.0, "content": "to ROIAlign (He et al., 2017), we flatten the point features into a single vector and project it to the", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 587, 495, 600 ], "spans": [ { "bbox": [ 105, 587, 393, 600 ], "score": 1.0, "content": "dimension of LLM embeddings. The final feature is used to replace the", "type": "text" }, { "bbox": [ 394, 587, 419, 599 ], "score": 0.61, "content": "\\langle { \\mathrm { S P E } } \\rangle", "type": "inline_equation" }, { "bbox": [ 419, 587, 495, 600 ], "score": 1.0, "content": "token in the input.", "type": "text" } ], "index": 32 } ], "index": 30, "bbox_fs": [ 105, 544, 505, 600 ] }, { "type": "text", "bbox": [ 107, 604, 505, 660 ], "lines": [ { "bbox": [ 106, 605, 504, 617 ], "spans": [ { "bbox": [ 106, 605, 504, 617 ], "score": 1.0, "content": "Output. The above region denotations are used in Ferret input to refer to specific regions. In", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 615, 505, 628 ], "spans": [ { "bbox": [ 105, 615, 505, 628 ], "score": 1.0, "content": "Ferret output, to achieve grounding, we generate the box coordinates right after the corresponding", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 626, 504, 639 ], "spans": [ { "bbox": [ 105, 626, 504, 639 ], "score": 1.0, "content": "regions/nouns in the text response. For instance, “There is a dog [100, 150, 300, 200] in the figure.”", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 636, 505, 650 ], "spans": [ { "bbox": [ 105, 636, 505, 650 ], "score": 1.0, "content": "With this data format, our model is expected to implicitly learn what is groundable in the current", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 647, 249, 662 ], "spans": [ { "bbox": [ 105, 647, 249, 662 ], "score": 1.0, "content": "image and what their locations are.", "type": "text" } ], "index": 37 } ], "index": 35, "bbox_fs": [ 105, 605, 505, 662 ] }, { "type": "text", "bbox": [ 107, 664, 505, 709 ], "lines": [ { "bbox": [ 105, 664, 505, 677 ], "spans": [ { "bbox": [ 105, 664, 505, 677 ], "score": 1.0, "content": "LLM. We consider Vicuna (Chiang et al., 2023) as our language model, a decoder-only", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 676, 505, 688 ], "spans": [ { "bbox": [ 105, 676, 505, 688 ], "score": 1.0, "content": "LLM (Brown et al., 2020) that is instruction-tuned on top of LLaMA (Touvron et al., 2023a). Prior", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "to being fed into the LLM, the image embeddings undergo transformation via an additional linear", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 699, 345, 710 ], "spans": [ { "bbox": [ 106, 699, 345, 710 ], "score": 1.0, "content": "layer to match the embedding dimension of the text tokens.", "type": "text" } ], "index": 41 } ], "index": 39.5, "bbox_fs": [ 105, 664, 506, 710 ] }, { "type": "text", "bbox": [ 106, 711, 505, 732 ], "lines": [ { "bbox": [ 118, 709, 505, 724 ], "spans": [ { "bbox": [ 118, 709, 312, 724 ], "score": 1.0, "content": "4FPS starts from a random single point sampled from", "type": "text" }, { "bbox": [ 312, 714, 321, 720 ], "score": 0.83, "content": "N", "type": "inline_equation" }, { "bbox": [ 321, 709, 505, 724 ], "score": 1.0, "content": "points. In each iteration, it samples one point from", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 721, 504, 732 ], "spans": [ { "bbox": [ 106, 721, 504, 732 ], "score": 1.0, "content": "the rest points such that it is the farthest from the set of already sampled points. See detail in Qi et al. (2017b).", "type": "text" } ], "index": 43 } ], "index": 42.5, "bbox_fs": [ 106, 709, 505, 732 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 107, 82, 500, 238 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 82, 500, 238 ], "group_id": 0, "lines": [ { "bbox": [ 107, 82, 500, 238 ], "spans": [ { "bbox": [ 107, 82, 500, 238 ], "score": 0.97, "type": "image", "image_path": "0f82847302401b88ae5e2072fa55c0a80e7982cdd90fe826e6eeacb28ce9c4ed.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 82, 500, 134.0 ], "spans": [], "index": 0 }, { "bbox": [ 107, 134.0, 500, 186.0 ], "spans": [], "index": 1 }, { "bbox": [ 107, 186.0, 500, 238.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 246, 505, 277 ], "group_id": 0, "lines": [ { "bbox": [ 105, 245, 505, 258 ], "spans": [ { "bbox": [ 105, 245, 505, 258 ], "score": 1.0, "content": "Figure 4: Overview of the GRIT dataset for Ferret model training. It contains three types of data: (i) public", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 255, 505, 268 ], "spans": [ { "bbox": [ 105, 255, 393, 268 ], "score": 1.0, "content": "datasets that are converted into an instruction-following format (the top-3 rows);", "type": "text" }, { "bbox": [ 394, 257, 406, 267 ], "score": 0.25, "content": "( i i )", "type": "inline_equation" }, { "bbox": [ 407, 255, 505, 268 ], "score": 1.0, "content": "data generated via prompt-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 266, 488, 278 ], "spans": [ { "bbox": [ 105, 266, 488, 278 ], "score": 1.0, "content": "ing ChatGPT and GPT-4 (the 4th row); and (iii) negative data to enhance model robustness (the last row).", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "title", "bbox": [ 107, 289, 441, 302 ], "lines": [ { "bbox": [ 104, 288, 443, 304 ], "spans": [ { "bbox": [ 104, 288, 443, 304 ], "score": 1.0, "content": "3 GRIT: GROUND-AND-REFER INSTRUCTION-TUNING DATASET", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 107, 311, 505, 366 ], "lines": [ { "bbox": [ 105, 311, 504, 322 ], "spans": [ { "bbox": [ 105, 311, 504, 322 ], "score": 1.0, "content": "In this section, we present GRIT, a Ground-and-Refer Instruction-Tuning dataset containing around", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 321, 505, 335 ], "spans": [ { "bbox": [ 105, 321, 505, 335 ], "score": 1.0, "content": "1.1M multimodal dialogues for model training. GRIT consists of three types of data: (i) public", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 331, 505, 347 ], "spans": [ { "bbox": [ 105, 331, 505, 347 ], "score": 1.0, "content": "datasets that are converted into an instruction-following format (Section 3.1); (ii) instruction-tuning", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 343, 505, 357 ], "spans": [ { "bbox": [ 105, 343, 505, 357 ], "score": 1.0, "content": "data generated via ChatGPT and GPT-4 (Section 3.2); and (iii) additional data from spatial negative", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 355, 322, 367 ], "spans": [ { "bbox": [ 106, 355, 322, 367 ], "score": 1.0, "content": "mining for enhancing model robustness (Section 3.3).", "type": "text" } ], "index": 11 } ], "index": 9 }, { "type": "title", "bbox": [ 107, 378, 184, 389 ], "lines": [ { "bbox": [ 105, 377, 185, 391 ], "spans": [ { "bbox": [ 105, 377, 185, 391 ], "score": 1.0, "content": "3.1 HIERARCHY", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 394, 503, 417 ], "lines": [ { "bbox": [ 105, 393, 505, 408 ], "spans": [ { "bbox": [ 105, 393, 505, 408 ], "score": 1.0, "content": "Spatial understanding can be characterized by varying levels of granularity and task formats. During", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 406, 445, 418 ], "spans": [ { "bbox": [ 106, 406, 445, 418 ], "score": 1.0, "content": "our dataset creation, we look into the following categories based on two dimensions:", "type": "text" } ], "index": 14 } ], "index": 13.5 }, { "type": "text", "bbox": [ 106, 422, 505, 471 ], "lines": [ { "bbox": [ 105, 422, 504, 436 ], "spans": [ { "bbox": [ 105, 422, 347, 436 ], "score": 1.0, "content": "• In terms of granularity, we identify four main categories:", "type": "text" }, { "bbox": [ 348, 423, 358, 434 ], "score": 0.43, "content": "( i )", "type": "inline_equation" }, { "bbox": [ 358, 422, 504, 436 ], "score": 1.0, "content": "individual objects, (ii) relationships", "type": "text" } ], "index": 15 }, { "bbox": [ 113, 433, 494, 448 ], "spans": [ { "bbox": [ 113, 433, 343, 448 ], "score": 1.0, "content": "among objects, (iii) descriptions of specific regions, and", "type": "text" }, { "bbox": [ 343, 435, 358, 444 ], "score": 0.43, "content": "( i v )", "type": "inline_equation" }, { "bbox": [ 358, 433, 494, 448 ], "score": 1.0, "content": "region-based complex reasoning.", "type": "text" } ], "index": 16 }, { "bbox": [ 104, 448, 506, 463 ], "spans": [ { "bbox": [ 104, 448, 506, 463 ], "score": 1.0, "content": "• In terms of task format, we further divide the data into three distinct types: (i) Region-in Text-out", "type": "text" } ], "index": 17 }, { "bbox": [ 113, 460, 408, 473 ], "spans": [ { "bbox": [ 113, 460, 408, 473 ], "score": 1.0, "content": "data, (ii) Text-in Region-out data, and (iii) Text-Region combined data.5", "type": "text" } ], "index": 18 } ], "index": 16.5 }, { "type": "text", "bbox": [ 108, 477, 502, 511 ], "lines": [ { "bbox": [ 105, 476, 504, 491 ], "spans": [ { "bbox": [ 105, 476, 504, 491 ], "score": 1.0, "content": "We compiled an extensive set of public data focusing on the aforementioned dimensions and con-", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 488, 504, 501 ], "spans": [ { "bbox": [ 106, 488, 504, 501 ], "score": 1.0, "content": "verted them into an instruction-following format using carefully designed templates. A more in-", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 500, 345, 512 ], "spans": [ { "bbox": [ 106, 500, 345, 512 ], "score": 1.0, "content": "depth view of these templates is available in Appendix C.1.", "type": "text" } ], "index": 21 } ], "index": 20 }, { "type": "text", "bbox": [ 106, 516, 505, 615 ], "lines": [ { "bbox": [ 106, 516, 505, 528 ], "spans": [ { "bbox": [ 106, 516, 505, 528 ], "score": 1.0, "content": "Individual objects. To achieve visual understanding at the object level, we select object detection", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 527, 505, 539 ], "spans": [ { "bbox": [ 106, 527, 505, 539 ], "score": 1.0, "content": "datasets such as Visual Genome (Krishna et al., 2017), Object365 (Shao et al., 2019), and visual", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 538, 506, 551 ], "spans": [ { "bbox": [ 105, 538, 506, 551 ], "score": 1.0, "content": "grounding datasets including RefCOCOs (Yu et al., 2016; Lin et al., 2014; Nagaraja et al., 2016)", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 549, 506, 562 ], "spans": [ { "bbox": [ 105, 549, 506, 562 ], "score": 1.0, "content": "and Flickr30k-Entities (Plummer et al., 2015). The converted Visual Genome object data follow a", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 559, 505, 573 ], "spans": [ { "bbox": [ 105, 559, 505, 573 ], "score": 1.0, "content": "Region-in Text-out format. Additionally, to enable Ferret to understand free-form shapes, we apply", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 570, 505, 583 ], "spans": [ { "bbox": [ 105, 570, 505, 583 ], "score": 1.0, "content": "SAM (Kirillov et al., 2023) to Visual Genome object data to obtain a segmentation mask for each", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 580, 505, 596 ], "spans": [ { "bbox": [ 105, 580, 505, 596 ], "score": 1.0, "content": "object, which is fed into the spatial-aware visual sampler to extract continuous region feature during", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 593, 505, 606 ], "spans": [ { "bbox": [ 105, 593, 505, 606 ], "score": 1.0, "content": "training. The visual grounding datasets and Object365 data adhere to a Text-in Region-out format.", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 603, 247, 616 ], "spans": [ { "bbox": [ 106, 603, 247, 616 ], "score": 1.0, "content": "This section has in total 678k data.", "type": "text" } ], "index": 30 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 621, 505, 676 ], "lines": [ { "bbox": [ 105, 620, 505, 633 ], "spans": [ { "bbox": [ 105, 620, 505, 633 ], "score": 1.0, "content": "Relationships among objects & descriptions of regions. We selected data pertaining to object", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 632, 505, 644 ], "spans": [ { "bbox": [ 106, 632, 505, 644 ], "score": 1.0, "content": "relationships and region captions from Visual Genome (Krishna et al., 2017) to address these two", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 642, 505, 655 ], "spans": [ { "bbox": [ 105, 642, 505, 655 ], "score": 1.0, "content": "facets, respectively. Both datasets employ a Region-in Text-out format and 177k data are obtained.", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 654, 505, 667 ], "spans": [ { "bbox": [ 105, 654, 505, 667 ], "score": 1.0, "content": "Similar to Visual Genome object data, we also extract segmentation masks of objects in Visual", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 664, 252, 676 ], "spans": [ { "bbox": [ 105, 664, 252, 676 ], "score": 1.0, "content": "Genome relationship data via SAM.", "type": "text" } ], "index": 35 } ], "index": 33 } ], "page_idx": 4, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 691, 505, 732 ], "lines": [ { "bbox": [ 118, 689, 506, 704 ], "spans": [ { "bbox": [ 118, 689, 506, 704 ], "score": 1.0, "content": "5For Region-in Text-out data, the input highlights a specific region, prompting queries about it. For Text-in", "type": "text" } ] }, { "bbox": [ 106, 701, 505, 713 ], "spans": [ { "bbox": [ 106, 701, 505, 713 ], "score": 1.0, "content": "Region-out data, the input comprises textual descriptions, and the task is to pinpoint or ground the relevant", "type": "text" } ] }, { "bbox": [ 105, 710, 505, 724 ], "spans": [ { "bbox": [ 105, 710, 505, 724 ], "score": 1.0, "content": "region in its response. The combined Text-Region data integrates both text and region within a single sequence,", "type": "text" } ] }, { "bbox": [ 106, 721, 286, 732 ], "spans": [ { "bbox": [ 106, 721, 286, 732 ], "score": 1.0, "content": "which can be present in the input, output, or both.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 108, 27, 292, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 308, 760 ], "lines": [ { "bbox": [ 302, 750, 309, 763 ], "spans": [ { "bbox": [ 302, 750, 309, 763 ], "score": 1.0, "content": "5", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 107, 82, 500, 238 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 82, 500, 238 ], "group_id": 0, "lines": [ { "bbox": [ 107, 82, 500, 238 ], "spans": [ { "bbox": [ 107, 82, 500, 238 ], "score": 0.97, "type": "image", "image_path": "0f82847302401b88ae5e2072fa55c0a80e7982cdd90fe826e6eeacb28ce9c4ed.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 82, 500, 134.0 ], "spans": [], "index": 0 }, { "bbox": [ 107, 134.0, 500, 186.0 ], "spans": [], "index": 1 }, { "bbox": [ 107, 186.0, 500, 238.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 246, 505, 277 ], "group_id": 0, "lines": [ { "bbox": [ 105, 245, 505, 258 ], "spans": [ { "bbox": [ 105, 245, 505, 258 ], "score": 1.0, "content": "Figure 4: Overview of the GRIT dataset for Ferret model training. It contains three types of data: (i) public", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 255, 505, 268 ], "spans": [ { "bbox": [ 105, 255, 393, 268 ], "score": 1.0, "content": "datasets that are converted into an instruction-following format (the top-3 rows);", "type": "text" }, { "bbox": [ 394, 257, 406, 267 ], "score": 0.25, "content": "( i i )", "type": "inline_equation" }, { "bbox": [ 407, 255, 505, 268 ], "score": 1.0, "content": "data generated via prompt-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 266, 488, 278 ], "spans": [ { "bbox": [ 105, 266, 488, 278 ], "score": 1.0, "content": "ing ChatGPT and GPT-4 (the 4th row); and (iii) negative data to enhance model robustness (the last row).", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "title", "bbox": [ 107, 289, 441, 302 ], "lines": [ { "bbox": [ 104, 288, 443, 304 ], "spans": [ { "bbox": [ 104, 288, 443, 304 ], "score": 1.0, "content": "3 GRIT: GROUND-AND-REFER INSTRUCTION-TUNING DATASET", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 107, 311, 505, 366 ], "lines": [ { "bbox": [ 105, 311, 504, 322 ], "spans": [ { "bbox": [ 105, 311, 504, 322 ], "score": 1.0, "content": "In this section, we present GRIT, a Ground-and-Refer Instruction-Tuning dataset containing around", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 321, 505, 335 ], "spans": [ { "bbox": [ 105, 321, 505, 335 ], "score": 1.0, "content": "1.1M multimodal dialogues for model training. GRIT consists of three types of data: (i) public", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 331, 505, 347 ], "spans": [ { "bbox": [ 105, 331, 505, 347 ], "score": 1.0, "content": "datasets that are converted into an instruction-following format (Section 3.1); (ii) instruction-tuning", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 343, 505, 357 ], "spans": [ { "bbox": [ 105, 343, 505, 357 ], "score": 1.0, "content": "data generated via ChatGPT and GPT-4 (Section 3.2); and (iii) additional data from spatial negative", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 355, 322, 367 ], "spans": [ { "bbox": [ 106, 355, 322, 367 ], "score": 1.0, "content": "mining for enhancing model robustness (Section 3.3).", "type": "text" } ], "index": 11 } ], "index": 9, "bbox_fs": [ 105, 311, 505, 367 ] }, { "type": "title", "bbox": [ 107, 378, 184, 389 ], "lines": [ { "bbox": [ 105, 377, 185, 391 ], "spans": [ { "bbox": [ 105, 377, 185, 391 ], "score": 1.0, "content": "3.1 HIERARCHY", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 394, 503, 417 ], "lines": [ { "bbox": [ 105, 393, 505, 408 ], "spans": [ { "bbox": [ 105, 393, 505, 408 ], "score": 1.0, "content": "Spatial understanding can be characterized by varying levels of granularity and task formats. During", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 406, 445, 418 ], "spans": [ { "bbox": [ 106, 406, 445, 418 ], "score": 1.0, "content": "our dataset creation, we look into the following categories based on two dimensions:", "type": "text" } ], "index": 14 } ], "index": 13.5, "bbox_fs": [ 105, 393, 505, 418 ] }, { "type": "text", "bbox": [ 106, 422, 505, 471 ], "lines": [ { "bbox": [ 105, 422, 504, 436 ], "spans": [ { "bbox": [ 105, 422, 347, 436 ], "score": 1.0, "content": "• In terms of granularity, we identify four main categories:", "type": "text" }, { "bbox": [ 348, 423, 358, 434 ], "score": 0.43, "content": "( i )", "type": "inline_equation" }, { "bbox": [ 358, 422, 504, 436 ], "score": 1.0, "content": "individual objects, (ii) relationships", "type": "text" } ], "index": 15 }, { "bbox": [ 113, 433, 494, 448 ], "spans": [ { "bbox": [ 113, 433, 343, 448 ], "score": 1.0, "content": "among objects, (iii) descriptions of specific regions, and", "type": "text" }, { "bbox": [ 343, 435, 358, 444 ], "score": 0.43, "content": "( i v )", "type": "inline_equation" }, { "bbox": [ 358, 433, 494, 448 ], "score": 1.0, "content": "region-based complex reasoning.", "type": "text" } ], "index": 16 }, { "bbox": [ 104, 448, 506, 463 ], "spans": [ { "bbox": [ 104, 448, 506, 463 ], "score": 1.0, "content": "• In terms of task format, we further divide the data into three distinct types: (i) Region-in Text-out", "type": "text" } ], "index": 17 }, { "bbox": [ 113, 460, 408, 473 ], "spans": [ { "bbox": [ 113, 460, 408, 473 ], "score": 1.0, "content": "data, (ii) Text-in Region-out data, and (iii) Text-Region combined data.5", "type": "text" } ], "index": 18 } ], "index": 16.5, "bbox_fs": [ 104, 422, 506, 473 ] }, { "type": "text", "bbox": [ 108, 477, 502, 511 ], "lines": [ { "bbox": [ 105, 476, 504, 491 ], "spans": [ { "bbox": [ 105, 476, 504, 491 ], "score": 1.0, "content": "We compiled an extensive set of public data focusing on the aforementioned dimensions and con-", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 488, 504, 501 ], "spans": [ { "bbox": [ 106, 488, 504, 501 ], "score": 1.0, "content": "verted them into an instruction-following format using carefully designed templates. A more in-", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 500, 345, 512 ], "spans": [ { "bbox": [ 106, 500, 345, 512 ], "score": 1.0, "content": "depth view of these templates is available in Appendix C.1.", "type": "text" } ], "index": 21 } ], "index": 20, "bbox_fs": [ 105, 476, 504, 512 ] }, { "type": "text", "bbox": [ 106, 516, 505, 615 ], "lines": [ { "bbox": [ 106, 516, 505, 528 ], "spans": [ { "bbox": [ 106, 516, 505, 528 ], "score": 1.0, "content": "Individual objects. To achieve visual understanding at the object level, we select object detection", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 527, 505, 539 ], "spans": [ { "bbox": [ 106, 527, 505, 539 ], "score": 1.0, "content": "datasets such as Visual Genome (Krishna et al., 2017), Object365 (Shao et al., 2019), and visual", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 538, 506, 551 ], "spans": [ { "bbox": [ 105, 538, 506, 551 ], "score": 1.0, "content": "grounding datasets including RefCOCOs (Yu et al., 2016; Lin et al., 2014; Nagaraja et al., 2016)", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 549, 506, 562 ], "spans": [ { "bbox": [ 105, 549, 506, 562 ], "score": 1.0, "content": "and Flickr30k-Entities (Plummer et al., 2015). The converted Visual Genome object data follow a", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 559, 505, 573 ], "spans": [ { "bbox": [ 105, 559, 505, 573 ], "score": 1.0, "content": "Region-in Text-out format. Additionally, to enable Ferret to understand free-form shapes, we apply", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 570, 505, 583 ], "spans": [ { "bbox": [ 105, 570, 505, 583 ], "score": 1.0, "content": "SAM (Kirillov et al., 2023) to Visual Genome object data to obtain a segmentation mask for each", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 580, 505, 596 ], "spans": [ { "bbox": [ 105, 580, 505, 596 ], "score": 1.0, "content": "object, which is fed into the spatial-aware visual sampler to extract continuous region feature during", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 593, 505, 606 ], "spans": [ { "bbox": [ 105, 593, 505, 606 ], "score": 1.0, "content": "training. The visual grounding datasets and Object365 data adhere to a Text-in Region-out format.", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 603, 247, 616 ], "spans": [ { "bbox": [ 106, 603, 247, 616 ], "score": 1.0, "content": "This section has in total 678k data.", "type": "text" } ], "index": 30 } ], "index": 26, "bbox_fs": [ 105, 516, 506, 616 ] }, { "type": "text", "bbox": [ 107, 621, 505, 676 ], "lines": [ { "bbox": [ 105, 620, 505, 633 ], "spans": [ { "bbox": [ 105, 620, 505, 633 ], "score": 1.0, "content": "Relationships among objects & descriptions of regions. We selected data pertaining to object", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 632, 505, 644 ], "spans": [ { "bbox": [ 106, 632, 505, 644 ], "score": 1.0, "content": "relationships and region captions from Visual Genome (Krishna et al., 2017) to address these two", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 642, 505, 655 ], "spans": [ { "bbox": [ 105, 642, 505, 655 ], "score": 1.0, "content": "facets, respectively. Both datasets employ a Region-in Text-out format and 177k data are obtained.", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 654, 505, 667 ], "spans": [ { "bbox": [ 105, 654, 505, 667 ], "score": 1.0, "content": "Similar to Visual Genome object data, we also extract segmentation masks of objects in Visual", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 664, 252, 676 ], "spans": [ { "bbox": [ 105, 664, 252, 676 ], "score": 1.0, "content": "Genome relationship data via SAM.", "type": "text" } ], "index": 35 } ], "index": 33, "bbox_fs": [ 105, 620, 505, 676 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 504, 115 ], "lines": [ { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "Region-based complex reasoning. Regarding complex reasoning centered on specific regions, we", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 505, 106 ], "spans": [ { "bbox": [ 105, 93, 505, 106 ], "score": 1.0, "content": "constructed a novel dataset with the help of ChatGPT/GPT-4. It adopts a combined Text-Region", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 303, 117 ], "spans": [ { "bbox": [ 105, 104, 303, 117 ], "score": 1.0, "content": "format, and is detailed in the subsequent section.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "title", "bbox": [ 107, 126, 379, 137 ], "lines": [ { "bbox": [ 105, 124, 380, 138 ], "spans": [ { "bbox": [ 105, 124, 380, 138 ], "score": 1.0, "content": "3.2 GPT-ASSISTED VISUAL INSTRUCTION DATA GENERATION", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 107, 141, 505, 207 ], "lines": [ { "bbox": [ 105, 140, 505, 153 ], "spans": [ { "bbox": [ 105, 140, 505, 153 ], "score": 1.0, "content": "Besides converting existing datasets by templates, dialogue instruction tuning data is proved to be", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 152, 505, 165 ], "spans": [ { "bbox": [ 106, 152, 505, 165 ], "score": 1.0, "content": "critical for MLLM to understand human intention and generate fluent, natural, and long-form re-", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 162, 506, 176 ], "spans": [ { "bbox": [ 105, 162, 506, 176 ], "score": 1.0, "content": "sponses (Liu et al., 2023b; Zhu et al., 2023a; Li et al., 2023d). Few-shot prompting is widely used", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 175, 505, 186 ], "spans": [ { "bbox": [ 106, 175, 505, 186 ], "score": 1.0, "content": "to obtain visual instruction tuning data, where textual scene descriptions of images and human-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 185, 505, 198 ], "spans": [ { "bbox": [ 105, 185, 505, 198 ], "score": 1.0, "content": "annotated dialogues are provided as few-shot demonstrations, and ChatGPT/GPT4 are prompted to", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 196, 411, 209 ], "spans": [ { "bbox": [ 105, 196, 411, 209 ], "score": 1.0, "content": "generate new dialogue based on the new image’s textual scene descriptions.", "type": "text" } ], "index": 9 } ], "index": 6.5 }, { "type": "text", "bbox": [ 107, 213, 505, 366 ], "lines": [ { "bbox": [ 105, 213, 505, 226 ], "spans": [ { "bbox": [ 105, 213, 505, 226 ], "score": 1.0, "content": "However, previous instruction tuning data mainly focus on describing the entire image without ex-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 223, 505, 238 ], "spans": [ { "bbox": [ 105, 223, 505, 238 ], "score": 1.0, "content": "plicitly specifying spatial-related information. To collect refer-and-ground instruction tuning data,", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 235, 506, 247 ], "spans": [ { "bbox": [ 105, 235, 506, 247 ], "score": 1.0, "content": "we emphasize region-based spatial knowledge in the following three steps. (i) Besides objects and", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 246, 505, 259 ], "spans": [ { "bbox": [ 105, 246, 505, 259 ], "score": 1.0, "content": "global captions usually used as before, our symbolic scene description additionally includes physical", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 257, 505, 270 ], "spans": [ { "bbox": [ 105, 257, 505, 270 ], "score": 1.0, "content": "relationships between objects and region captions along with coordinates of them. (ii) In human-", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 268, 505, 281 ], "spans": [ { "bbox": [ 105, 268, 505, 281 ], "score": 1.0, "content": "annotated dialogues, we add coordinates after the groundable regions or objects either in input or", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 278, 505, 293 ], "spans": [ { "bbox": [ 105, 278, 505, 293 ], "score": 1.0, "content": "output or both, and the dialogues are typically focused on specific regions. It helps to implicitly", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 290, 505, 303 ], "spans": [ { "bbox": [ 105, 290, 505, 303 ], "score": 1.0, "content": "prompt ChatGPT/GPT4 to follow similar patterns when generating new dialogues. (iii) The gen-", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 300, 506, 313 ], "spans": [ { "bbox": [ 105, 300, 506, 313 ], "score": 1.0, "content": "erated dialogues sometimes cannot follow the rules and patterns we wrote in system prompts and", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 312, 506, 325 ], "spans": [ { "bbox": [ 105, 312, 506, 325 ], "score": 1.0, "content": "few-shot examples, which might be due to that the context of LLM input is too long to handle all", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 322, 505, 335 ], "spans": [ { "bbox": [ 105, 322, 505, 335 ], "score": 1.0, "content": "the details. To alleviate it, we propose to use ChatGPT/GPT-4 again to refine the initially generated", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 332, 505, 348 ], "spans": [ { "bbox": [ 105, 332, 267, 348 ], "score": 1.0, "content": "dialogues, whose context length is only", "type": "text" }, { "bbox": [ 268, 334, 287, 344 ], "score": 0.86, "content": "10 \\%", "type": "inline_equation" }, { "bbox": [ 288, 332, 505, 348 ], "score": 1.0, "content": "of the data generated from the first round on average.", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 344, 505, 357 ], "spans": [ { "bbox": [ 105, 344, 505, 357 ], "score": 1.0, "content": "To save cost, we use ChatGPT in the first round of generation and GPT-4 for refining. 34k dialogues", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 356, 192, 367 ], "spans": [ { "bbox": [ 105, 356, 192, 367 ], "score": 1.0, "content": "in total are collected.", "type": "text" } ], "index": 23 } ], "index": 16.5 }, { "type": "text", "bbox": [ 107, 372, 505, 416 ], "lines": [ { "bbox": [ 105, 372, 505, 385 ], "spans": [ { "bbox": [ 105, 372, 505, 385 ], "score": 1.0, "content": "Additionally, to exploit existing instruction-tuning data such as those in LLaVA (Liu et al., 2023b),", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 383, 506, 396 ], "spans": [ { "bbox": [ 105, 383, 506, 396 ], "score": 1.0, "content": "we apply an open-vocabulary object detector, GLIPv2 (Zhang et al., 2022), on LLaVA-158k data to", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 394, 505, 407 ], "spans": [ { "bbox": [ 105, 394, 505, 407 ], "score": 1.0, "content": "localize groundable nouns in the text. Then, we append the bounding boxes after the corresponding", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 405, 492, 418 ], "spans": [ { "bbox": [ 105, 405, 492, 418 ], "score": 1.0, "content": "nouns, forming a pseudo-grounded LLaVA instruction data that are also used for training Ferret.", "type": "text" } ], "index": 27 } ], "index": 25.5 }, { "type": "title", "bbox": [ 108, 426, 252, 438 ], "lines": [ { "bbox": [ 105, 425, 253, 439 ], "spans": [ { "bbox": [ 105, 425, 253, 439 ], "score": 1.0, "content": "3.3 SPATIAL NEGATIVE MINING", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 107, 442, 505, 562 ], "lines": [ { "bbox": [ 105, 441, 506, 455 ], "spans": [ { "bbox": [ 105, 441, 506, 455 ], "score": 1.0, "content": "As highlighted in prior studies (Li et al., 2023e; Liu et al., 2023a), MLLM exhibits a propensity to", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 453, 505, 466 ], "spans": [ { "bbox": [ 105, 453, 505, 466 ], "score": 1.0, "content": "hallucinate in response to yes/no questions. We observed a similar occurrence when inquiring about", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 464, 505, 477 ], "spans": [ { "bbox": [ 105, 464, 505, 477 ], "score": 1.0, "content": "detailed regions. To address this, we also conduct negative sample mining by following two ways: (i)", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 475, 505, 488 ], "spans": [ { "bbox": [ 105, 475, 505, 488 ], "score": 1.0, "content": "Image-conditioned Category Localization, and (ii) Semantics-conditioned Category Localization.", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 485, 505, 500 ], "spans": [ { "bbox": [ 105, 485, 505, 500 ], "score": 1.0, "content": "They both ask the model to localize specific object categories, thereby enabling the model’s ability", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 497, 505, 510 ], "spans": [ { "bbox": [ 105, 497, 505, 510 ], "score": 1.0, "content": "to discern and potentially recognize the absence of certain objects. They differ in how to select the", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 508, 505, 520 ], "spans": [ { "bbox": [ 105, 508, 505, 520 ], "score": 1.0, "content": "negative category. For (i), Object365 data are employed and we randomly select the object class", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 519, 505, 531 ], "spans": [ { "bbox": [ 105, 519, 505, 531 ], "score": 1.0, "content": "from the vocabulary that is not shown in the given image. For (ii), Flickr30k data are used and", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 530, 505, 542 ], "spans": [ { "bbox": [ 105, 530, 505, 542 ], "score": 1.0, "content": "negative categories are sourced by utilizing ChatGPT/GPT4 to find entities that are most analogous", "type": "text" } ], "index": 37 }, { "bbox": [ 104, 540, 506, 554 ], "spans": [ { "bbox": [ 104, 540, 506, 554 ], "score": 1.0, "content": "to the original class, attribute, or quantity, e.g., ‘man’ vs. ‘woman’, ‘blue’ vs. ‘yellow’, ‘two’ vs.", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 551, 140, 563 ], "spans": [ { "bbox": [ 106, 551, 140, 563 ], "score": 1.0, "content": "‘three’.", "type": "text" } ], "index": 39 } ], "index": 34 }, { "type": "text", "bbox": [ 107, 568, 504, 591 ], "lines": [ { "bbox": [ 106, 568, 505, 581 ], "spans": [ { "bbox": [ 106, 568, 505, 581 ], "score": 1.0, "content": "We curate the data to maintain an equilibrium between positive and negative samples for each of the", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 579, 504, 592 ], "spans": [ { "bbox": [ 105, 579, 504, 592 ], "score": 1.0, "content": "two types.6 95k data are collected. A more comprehensive elaboration is provided in Appendix C.2.", "type": "text" } ], "index": 41 } ], "index": 40.5 }, { "type": "title", "bbox": [ 108, 604, 200, 616 ], "lines": [ { "bbox": [ 105, 603, 201, 618 ], "spans": [ { "bbox": [ 105, 603, 201, 618 ], "score": 1.0, "content": "4 EXPERIMENTS", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "text", "bbox": [ 107, 624, 505, 690 ], "lines": [ { "bbox": [ 105, 622, 506, 638 ], "spans": [ { "bbox": [ 105, 622, 506, 638 ], "score": 1.0, "content": "First of all, we illustrate the training details of Ferret. Then in evaluation, we start with evaluating", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 635, 505, 647 ], "spans": [ { "bbox": [ 105, 635, 505, 647 ], "score": 1.0, "content": "Ferret on conventional referring and grounding benchmarks (Sec. 4.1 and 4.2). Then, we demon-", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 645, 505, 659 ], "spans": [ { "bbox": [ 105, 645, 505, 659 ], "score": 1.0, "content": "strate the power of Ferret in more complex multimodal chatting with refer-and-ground capability", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 656, 505, 669 ], "spans": [ { "bbox": [ 105, 656, 505, 669 ], "score": 1.0, "content": "in Sec. 4.3. For a detailed visualization of each, kindly check Appendix E. We further ablate key", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 668, 505, 680 ], "spans": [ { "bbox": [ 105, 668, 505, 680 ], "score": 1.0, "content": "components in Ferret (Sec. 4.4), analyze the object hallucination of Ferret (Sec. 4.5) and discuss", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 678, 226, 691 ], "spans": [ { "bbox": [ 105, 678, 226, 691 ], "score": 1.0, "content": "Ferret v.s. GPT-4V (Sec. ??).", "type": "text" } ], "index": 48 } ], "index": 45.5 } ], "page_idx": 5, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 701, 504, 732 ], "lines": [ { "bbox": [ 118, 699, 505, 714 ], "spans": [ { "bbox": [ 118, 699, 505, 714 ], "score": 1.0, "content": "6We observed that even though we don’t collect other data specifically for training, Ferret demonstrates the", "type": "text" } ] }, { "bbox": [ 106, 712, 505, 723 ], "spans": [ { "bbox": [ 106, 712, 505, 723 ], "score": 1.0, "content": "capability to generalize robustness across diverse categories like relationships, events, etc. We attribute this", "type": "text" } ] }, { "bbox": [ 105, 721, 352, 732 ], "spans": [ { "bbox": [ 105, 721, 352, 732 ], "score": 1.0, "content": "versatility to the potent compositional capabilities inherent to LLM.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "6", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 504, 115 ], "lines": [ { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "Region-based complex reasoning. Regarding complex reasoning centered on specific regions, we", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 505, 106 ], "spans": [ { "bbox": [ 105, 93, 505, 106 ], "score": 1.0, "content": "constructed a novel dataset with the help of ChatGPT/GPT-4. It adopts a combined Text-Region", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 303, 117 ], "spans": [ { "bbox": [ 105, 104, 303, 117 ], "score": 1.0, "content": "format, and is detailed in the subsequent section.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 105, 82, 505, 117 ] }, { "type": "title", "bbox": [ 107, 126, 379, 137 ], "lines": [ { "bbox": [ 105, 124, 380, 138 ], "spans": [ { "bbox": [ 105, 124, 380, 138 ], "score": 1.0, "content": "3.2 GPT-ASSISTED VISUAL INSTRUCTION DATA GENERATION", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 107, 141, 505, 207 ], "lines": [ { "bbox": [ 105, 140, 505, 153 ], "spans": [ { "bbox": [ 105, 140, 505, 153 ], "score": 1.0, "content": "Besides converting existing datasets by templates, dialogue instruction tuning data is proved to be", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 152, 505, 165 ], "spans": [ { "bbox": [ 106, 152, 505, 165 ], "score": 1.0, "content": "critical for MLLM to understand human intention and generate fluent, natural, and long-form re-", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 162, 506, 176 ], "spans": [ { "bbox": [ 105, 162, 506, 176 ], "score": 1.0, "content": "sponses (Liu et al., 2023b; Zhu et al., 2023a; Li et al., 2023d). Few-shot prompting is widely used", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 175, 505, 186 ], "spans": [ { "bbox": [ 106, 175, 505, 186 ], "score": 1.0, "content": "to obtain visual instruction tuning data, where textual scene descriptions of images and human-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 185, 505, 198 ], "spans": [ { "bbox": [ 105, 185, 505, 198 ], "score": 1.0, "content": "annotated dialogues are provided as few-shot demonstrations, and ChatGPT/GPT4 are prompted to", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 196, 411, 209 ], "spans": [ { "bbox": [ 105, 196, 411, 209 ], "score": 1.0, "content": "generate new dialogue based on the new image’s textual scene descriptions.", "type": "text" } ], "index": 9 } ], "index": 6.5, "bbox_fs": [ 105, 140, 506, 209 ] }, { "type": "text", "bbox": [ 107, 213, 505, 366 ], "lines": [ { "bbox": [ 105, 213, 505, 226 ], "spans": [ { "bbox": [ 105, 213, 505, 226 ], "score": 1.0, "content": "However, previous instruction tuning data mainly focus on describing the entire image without ex-", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 223, 505, 238 ], "spans": [ { "bbox": [ 105, 223, 505, 238 ], "score": 1.0, "content": "plicitly specifying spatial-related information. To collect refer-and-ground instruction tuning data,", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 235, 506, 247 ], "spans": [ { "bbox": [ 105, 235, 506, 247 ], "score": 1.0, "content": "we emphasize region-based spatial knowledge in the following three steps. (i) Besides objects and", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 246, 505, 259 ], "spans": [ { "bbox": [ 105, 246, 505, 259 ], "score": 1.0, "content": "global captions usually used as before, our symbolic scene description additionally includes physical", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 257, 505, 270 ], "spans": [ { "bbox": [ 105, 257, 505, 270 ], "score": 1.0, "content": "relationships between objects and region captions along with coordinates of them. (ii) In human-", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 268, 505, 281 ], "spans": [ { "bbox": [ 105, 268, 505, 281 ], "score": 1.0, "content": "annotated dialogues, we add coordinates after the groundable regions or objects either in input or", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 278, 505, 293 ], "spans": [ { "bbox": [ 105, 278, 505, 293 ], "score": 1.0, "content": "output or both, and the dialogues are typically focused on specific regions. It helps to implicitly", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 290, 505, 303 ], "spans": [ { "bbox": [ 105, 290, 505, 303 ], "score": 1.0, "content": "prompt ChatGPT/GPT4 to follow similar patterns when generating new dialogues. (iii) The gen-", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 300, 506, 313 ], "spans": [ { "bbox": [ 105, 300, 506, 313 ], "score": 1.0, "content": "erated dialogues sometimes cannot follow the rules and patterns we wrote in system prompts and", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 312, 506, 325 ], "spans": [ { "bbox": [ 105, 312, 506, 325 ], "score": 1.0, "content": "few-shot examples, which might be due to that the context of LLM input is too long to handle all", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 322, 505, 335 ], "spans": [ { "bbox": [ 105, 322, 505, 335 ], "score": 1.0, "content": "the details. To alleviate it, we propose to use ChatGPT/GPT-4 again to refine the initially generated", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 332, 505, 348 ], "spans": [ { "bbox": [ 105, 332, 267, 348 ], "score": 1.0, "content": "dialogues, whose context length is only", "type": "text" }, { "bbox": [ 268, 334, 287, 344 ], "score": 0.86, "content": "10 \\%", "type": "inline_equation" }, { "bbox": [ 288, 332, 505, 348 ], "score": 1.0, "content": "of the data generated from the first round on average.", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 344, 505, 357 ], "spans": [ { "bbox": [ 105, 344, 505, 357 ], "score": 1.0, "content": "To save cost, we use ChatGPT in the first round of generation and GPT-4 for refining. 34k dialogues", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 356, 192, 367 ], "spans": [ { "bbox": [ 105, 356, 192, 367 ], "score": 1.0, "content": "in total are collected.", "type": "text" } ], "index": 23 } ], "index": 16.5, "bbox_fs": [ 105, 213, 506, 367 ] }, { "type": "text", "bbox": [ 107, 372, 505, 416 ], "lines": [ { "bbox": [ 105, 372, 505, 385 ], "spans": [ { "bbox": [ 105, 372, 505, 385 ], "score": 1.0, "content": "Additionally, to exploit existing instruction-tuning data such as those in LLaVA (Liu et al., 2023b),", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 383, 506, 396 ], "spans": [ { "bbox": [ 105, 383, 506, 396 ], "score": 1.0, "content": "we apply an open-vocabulary object detector, GLIPv2 (Zhang et al., 2022), on LLaVA-158k data to", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 394, 505, 407 ], "spans": [ { "bbox": [ 105, 394, 505, 407 ], "score": 1.0, "content": "localize groundable nouns in the text. Then, we append the bounding boxes after the corresponding", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 405, 492, 418 ], "spans": [ { "bbox": [ 105, 405, 492, 418 ], "score": 1.0, "content": "nouns, forming a pseudo-grounded LLaVA instruction data that are also used for training Ferret.", "type": "text" } ], "index": 27 } ], "index": 25.5, "bbox_fs": [ 105, 372, 506, 418 ] }, { "type": "title", "bbox": [ 108, 426, 252, 438 ], "lines": [ { "bbox": [ 105, 425, 253, 439 ], "spans": [ { "bbox": [ 105, 425, 253, 439 ], "score": 1.0, "content": "3.3 SPATIAL NEGATIVE MINING", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 107, 442, 505, 562 ], "lines": [ { "bbox": [ 105, 441, 506, 455 ], "spans": [ { "bbox": [ 105, 441, 506, 455 ], "score": 1.0, "content": "As highlighted in prior studies (Li et al., 2023e; Liu et al., 2023a), MLLM exhibits a propensity to", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 453, 505, 466 ], "spans": [ { "bbox": [ 105, 453, 505, 466 ], "score": 1.0, "content": "hallucinate in response to yes/no questions. We observed a similar occurrence when inquiring about", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 464, 505, 477 ], "spans": [ { "bbox": [ 105, 464, 505, 477 ], "score": 1.0, "content": "detailed regions. To address this, we also conduct negative sample mining by following two ways: (i)", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 475, 505, 488 ], "spans": [ { "bbox": [ 105, 475, 505, 488 ], "score": 1.0, "content": "Image-conditioned Category Localization, and (ii) Semantics-conditioned Category Localization.", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 485, 505, 500 ], "spans": [ { "bbox": [ 105, 485, 505, 500 ], "score": 1.0, "content": "They both ask the model to localize specific object categories, thereby enabling the model’s ability", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 497, 505, 510 ], "spans": [ { "bbox": [ 105, 497, 505, 510 ], "score": 1.0, "content": "to discern and potentially recognize the absence of certain objects. They differ in how to select the", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 508, 505, 520 ], "spans": [ { "bbox": [ 105, 508, 505, 520 ], "score": 1.0, "content": "negative category. For (i), Object365 data are employed and we randomly select the object class", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 519, 505, 531 ], "spans": [ { "bbox": [ 105, 519, 505, 531 ], "score": 1.0, "content": "from the vocabulary that is not shown in the given image. For (ii), Flickr30k data are used and", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 530, 505, 542 ], "spans": [ { "bbox": [ 105, 530, 505, 542 ], "score": 1.0, "content": "negative categories are sourced by utilizing ChatGPT/GPT4 to find entities that are most analogous", "type": "text" } ], "index": 37 }, { "bbox": [ 104, 540, 506, 554 ], "spans": [ { "bbox": [ 104, 540, 506, 554 ], "score": 1.0, "content": "to the original class, attribute, or quantity, e.g., ‘man’ vs. ‘woman’, ‘blue’ vs. ‘yellow’, ‘two’ vs.", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 551, 140, 563 ], "spans": [ { "bbox": [ 106, 551, 140, 563 ], "score": 1.0, "content": "‘three’.", "type": "text" } ], "index": 39 } ], "index": 34, "bbox_fs": [ 104, 441, 506, 563 ] }, { "type": "text", "bbox": [ 107, 568, 504, 591 ], "lines": [ { "bbox": [ 106, 568, 505, 581 ], "spans": [ { "bbox": [ 106, 568, 505, 581 ], "score": 1.0, "content": "We curate the data to maintain an equilibrium between positive and negative samples for each of the", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 579, 504, 592 ], "spans": [ { "bbox": [ 105, 579, 504, 592 ], "score": 1.0, "content": "two types.6 95k data are collected. A more comprehensive elaboration is provided in Appendix C.2.", "type": "text" } ], "index": 41 } ], "index": 40.5, "bbox_fs": [ 105, 568, 505, 592 ] }, { "type": "title", "bbox": [ 108, 604, 200, 616 ], "lines": [ { "bbox": [ 105, 603, 201, 618 ], "spans": [ { "bbox": [ 105, 603, 201, 618 ], "score": 1.0, "content": "4 EXPERIMENTS", "type": "text" } ], "index": 42 } ], "index": 42 }, { "type": "text", "bbox": [ 107, 624, 505, 690 ], "lines": [ { "bbox": [ 105, 622, 506, 638 ], "spans": [ { "bbox": [ 105, 622, 506, 638 ], "score": 1.0, "content": "First of all, we illustrate the training details of Ferret. Then in evaluation, we start with evaluating", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 635, 505, 647 ], "spans": [ { "bbox": [ 105, 635, 505, 647 ], "score": 1.0, "content": "Ferret on conventional referring and grounding benchmarks (Sec. 4.1 and 4.2). Then, we demon-", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 645, 505, 659 ], "spans": [ { "bbox": [ 105, 645, 505, 659 ], "score": 1.0, "content": "strate the power of Ferret in more complex multimodal chatting with refer-and-ground capability", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 656, 505, 669 ], "spans": [ { "bbox": [ 105, 656, 505, 669 ], "score": 1.0, "content": "in Sec. 4.3. For a detailed visualization of each, kindly check Appendix E. We further ablate key", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 668, 505, 680 ], "spans": [ { "bbox": [ 105, 668, 505, 680 ], "score": 1.0, "content": "components in Ferret (Sec. 4.4), analyze the object hallucination of Ferret (Sec. 4.5) and discuss", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 678, 226, 691 ], "spans": [ { "bbox": [ 105, 678, 226, 691 ], "score": 1.0, "content": "Ferret v.s. GPT-4V (Sec. ??).", "type": "text" } ], "index": 48 } ], "index": 45.5, "bbox_fs": [ 105, 622, 506, 691 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 107, 124, 273, 213 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 81, 274, 123 ], "group_id": 0, "lines": [ { "bbox": [ 106, 81, 275, 93 ], "spans": [ { "bbox": [ 106, 81, 275, 93 ], "score": 1.0, "content": "Table 1: Results of referring object classi-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 92, 275, 104 ], "spans": [ { "bbox": [ 106, 92, 275, 104 ], "score": 1.0, "content": "fication on three different referring types, in-", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 102, 273, 114 ], "spans": [ { "bbox": [ 106, 102, 260, 114 ], "score": 1.0, "content": "cluding point, box, and free-form shape.", "type": "text" }, { "bbox": [ 260, 102, 273, 112 ], "score": 0.41, "content": "\\mathbf { \\vec { \\mathbf { \\rho } } } \\mathbf { \\times } \\mathbf { \\vec { \\mathbf { \\rho } } }", "type": "inline_equation" } ], "index": 2 }, { "bbox": [ 106, 112, 201, 123 ], "spans": [ { "bbox": [ 106, 112, 201, 123 ], "score": 1.0, "content": "means no such capability.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "table_body", "bbox": [ 107, 124, 273, 213 ], "group_id": 0, "lines": [ { "bbox": [ 107, 124, 273, 213 ], "spans": [ { "bbox": [ 107, 124, 273, 213 ], "score": 0.979, "html": "
ModelsLVIS (Acc %)
PointBoxFree-form
Random Guess505050
LLaVA50.150.3×
Kosmos-2 (Peng et al.,2023)×60.25×
Shikra-7B(Chen et al.,2023b)57.82 67.71×
GPT4-ROI (Zhang et al.,2023)×61.76×
Ferret-7B67.9479.4269.77
Ferret-13B68.3580.4670.98
", "type": "table", "image_path": "d6bb85e1180177a92096c416e7a449b143cb3b05a6e56804b17bee6fbcb06413.jpg" } ] } ], "index": 13.0, "virtual_lines": [ { "bbox": [ 107, 124, 273, 138.83333333333334 ], "spans": [], "index": 8 }, { "bbox": [ 107, 138.83333333333334, 273, 153.66666666666669 ], "spans": [], "index": 10 }, { "bbox": [ 107, 153.66666666666669, 273, 168.50000000000003 ], "spans": [], "index": 12 }, { "bbox": [ 107, 168.50000000000003, 273, 183.33333333333337 ], "spans": [], "index": 14 }, { "bbox": [ 107, 183.33333333333337, 273, 198.1666666666667 ], "spans": [], "index": 16 }, { "bbox": [ 107, 198.1666666666667, 273, 213.00000000000006 ], "spans": [], "index": 18 } ] } ], "index": 7.25 }, { "type": "table", "bbox": [ 280, 125, 501, 213 ], "blocks": [ { "type": "table_caption", "bbox": [ 277, 81, 502, 122 ], "group_id": 1, "lines": [ { "bbox": [ 277, 80, 503, 93 ], "spans": [ { "bbox": [ 277, 80, 503, 93 ], "score": 1.0, "content": "Table 2: Results of grounded image captioning on the test", "type": "text" } ], "index": 4 }, { "bbox": [ 278, 91, 502, 101 ], "spans": [ { "bbox": [ 278, 91, 398, 101 ], "score": 1.0, "content": "set of Flickr30k Entities. BLEU", "type": "text" }, { "bbox": [ 399, 91, 411, 100 ], "score": 0.63, "content": "@ 4", "type": "inline_equation" }, { "bbox": [ 412, 91, 502, 101 ], "score": 1.0, "content": ", METEOR, CIDEr, and", "type": "text" } ], "index": 5 }, { "bbox": [ 277, 100, 501, 113 ], "spans": [ { "bbox": [ 277, 100, 438, 113 ], "score": 1.0, "content": "SPICE are used for the caption evaluation.", "type": "text" }, { "bbox": [ 439, 101, 461, 111 ], "score": 0.9, "content": "F 1 _ { a l l }", "type": "inline_equation" }, { "bbox": [ 461, 100, 478, 113 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 478, 101, 501, 111 ], "score": 0.9, "content": "F 1 _ { l o c }", "type": "inline_equation" } ], "index": 6 }, { "bbox": [ 277, 111, 490, 122 ], "spans": [ { "bbox": [ 277, 111, 404, 122 ], "score": 1.0, "content": "are used for grounding evaluation.", "type": "text" }, { "bbox": [ 405, 111, 415, 120 ], "score": 0.26, "content": "\" - \"", "type": "inline_equation" }, { "bbox": [ 416, 111, 490, 122 ], "score": 1.0, "content": "means not reported.", "type": "text" } ], "index": 7 } ], "index": 5.5 }, { "type": "table_body", "bbox": [ 280, 125, 501, 213 ], "group_id": 1, "lines": [ { "bbox": [ 280, 125, 501, 213 ], "spans": [ { "bbox": [ 280, 125, 501, 213 ], "score": 0.98, "html": "
ModelsCaption Eval.Grounding Eval.
B@4MCSF1allF1loc
GVD (Zhou et al.,2019)27.3 22.5 62.3 16.57.5522.2
Cyclical (Ma et al.,2020)26.8 22.4 61.1 16.88.4422.78
POS-SCAN (Zhou et al., 2020)30.1 22.6 69.3 16.87.1717.49
UniTAB (Yang et al., 2022)30.1 23.7 69.7 17.412.9534.79
Shikra-13B (Chen et al.,2023b)1173.9111
Ferret-7B35.1 24.6 74.8 18.015.0237.62
Ferret-13B37.0 25.5 76.1 18.315.1238.03
", "type": "table", "image_path": "bafe86a4352ddc05a3896487ffd7534bf2f5780b5f85b54c1f3dd76763561a93.jpg" } ] } ], "index": 14.0, "virtual_lines": [ { "bbox": [ 280, 125, 501, 139.66666666666666 ], "spans": [], "index": 9 }, { "bbox": [ 280, 139.66666666666666, 501, 154.33333333333331 ], "spans": [], "index": 11 }, { "bbox": [ 280, 154.33333333333331, 501, 168.99999999999997 ], "spans": [], "index": 13 }, { "bbox": [ 280, 168.99999999999997, 501, 183.66666666666663 ], "spans": [], "index": 15 }, { "bbox": [ 280, 183.66666666666663, 501, 198.3333333333333 ], "spans": [], "index": 17 }, { "bbox": [ 280, 198.3333333333333, 501, 212.99999999999994 ], "spans": [], "index": 19 } ] } ], "index": 9.75 }, { "type": "text", "bbox": [ 107, 217, 505, 306 ], "lines": [ { "bbox": [ 105, 217, 505, 230 ], "spans": [ { "bbox": [ 105, 217, 376, 230 ], "score": 1.0, "content": "Training Details. We initialize the image encoder with CLIP-ViT-", "type": "text" }, { "bbox": [ 377, 218, 426, 229 ], "score": 0.4, "content": "\\mathrm { L } / 1 4 @ 3 3 6 \\mathrm { p }", "type": "inline_equation" }, { "bbox": [ 426, 217, 505, 230 ], "score": 1.0, "content": ", the LLM with Vi-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 228, 505, 240 ], "spans": [ { "bbox": [ 105, 228, 505, 240 ], "score": 1.0, "content": "cuna, and the projection layer with LLaVA’s first-stage weights, leaving the visual sampler ran-", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 240, 505, 250 ], "spans": [ { "bbox": [ 106, 240, 505, 250 ], "score": 1.0, "content": "domly initialized. After the initialization, Ferret is trained on the aforementioned GRIT data for", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 250, 505, 263 ], "spans": [ { "bbox": [ 106, 250, 426, 263 ], "score": 1.0, "content": "three epochs, optimized by Loshchilov & Hutter (2017) with a learning rate of", "type": "text" }, { "bbox": [ 426, 250, 455, 261 ], "score": 0.88, "content": "2 e - 5", "type": "inline_equation" }, { "bbox": [ 455, 250, 505, 263 ], "score": 1.0, "content": "and a batch", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 259, 505, 275 ], "spans": [ { "bbox": [ 105, 259, 232, 275 ], "score": 1.0, "content": "size of 128. The training takes", "type": "text" }, { "bbox": [ 232, 261, 262, 272 ], "score": 0.91, "content": "{ \\sim } 5 / 2 . 5", "type": "inline_equation" }, { "bbox": [ 263, 259, 505, 275 ], "score": 1.0, "content": "days on 8 A100 GPU for a Ferret-13B/7B. During training,", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 272, 505, 285 ], "spans": [ { "bbox": [ 106, 272, 505, 285 ], "score": 1.0, "content": "when input refers to regions, we randomly choose either the center points or the bounding boxes", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 283, 505, 296 ], "spans": [ { "bbox": [ 105, 283, 505, 296 ], "score": 1.0, "content": "(or segmentation masks if available) to represent the regions. We perform de-duplication in training", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 294, 360, 306 ], "spans": [ { "bbox": [ 106, 294, 360, 306 ], "score": 1.0, "content": "data to remove the samples that are in downstream evaluations.", "type": "text" } ], "index": 27 } ], "index": 23.5 }, { "type": "title", "bbox": [ 107, 313, 211, 324 ], "lines": [ { "bbox": [ 105, 312, 212, 325 ], "spans": [ { "bbox": [ 105, 312, 212, 325 ], "score": 1.0, "content": "4.1 INPUT REFERRING", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 106, 327, 505, 438 ], "lines": [ { "bbox": [ 106, 328, 505, 340 ], "spans": [ { "bbox": [ 106, 328, 505, 340 ], "score": 1.0, "content": "The model’s capability of understanding referring is reflected in that, given a referred region in the", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 339, 506, 353 ], "spans": [ { "bbox": [ 105, 339, 506, 353 ], "score": 1.0, "content": "question, how accurately the model can understand the semantics of the referred region. To measure", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 350, 505, 363 ], "spans": [ { "bbox": [ 105, 350, 505, 363 ], "score": 1.0, "content": "it, we start with the most basic semantics, object, as it is fundamental and clear to define. To be more", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 361, 505, 374 ], "spans": [ { "bbox": [ 105, 361, 505, 374 ], "score": 1.0, "content": "specific, the task we evaluate on is Referring Object Classification: the question refers to a specific", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 372, 505, 385 ], "spans": [ { "bbox": [ 105, 372, 505, 385 ], "score": 1.0, "content": "region in the image, and the model needs to classify the object in the region. Since Ferret and", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 383, 505, 396 ], "spans": [ { "bbox": [ 105, 383, 505, 396 ], "score": 1.0, "content": "MLLMs usually generate free-form text responses, it is inaccurate to match the predicted class with", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 393, 506, 407 ], "spans": [ { "bbox": [ 105, 393, 506, 407 ], "score": 1.0, "content": "the ground-truth class if directly asking the model to classify without constraints. Alternatively,", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 404, 506, 417 ], "spans": [ { "bbox": [ 105, 404, 506, 417 ], "score": 1.0, "content": "we make it a binary-choice question in the format of “Is the object ⟨location⟩ a ⟨class A⟩ or a", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 414, 506, 430 ], "spans": [ { "bbox": [ 106, 414, 506, 430 ], "score": 1.0, "content": "⟨class B⟩?”. We feed the binary-choice question and image into the MLLMs to obtain the response,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 426, 437, 439 ], "spans": [ { "bbox": [ 105, 426, 437, 439 ], "score": 1.0, "content": "and then detect if the response matches the ground-truth (GT) class by some rule.7", "type": "text" } ], "index": 38 } ], "index": 33.5 }, { "type": "text", "bbox": [ 107, 443, 505, 565 ], "lines": [ { "bbox": [ 105, 442, 505, 457 ], "spans": [ { "bbox": [ 105, 442, 505, 457 ], "score": 1.0, "content": "To prepare the data, we used the validation split of LVIS dataset (Gupta et al., 2019) covering", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 455, 505, 468 ], "spans": [ { "bbox": [ 105, 455, 505, 468 ], "score": 1.0, "content": "over 1000 object categories, and sampled 2667 objects as the GT objects. Then, we randomly", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 466, 505, 478 ], "spans": [ { "bbox": [ 105, 466, 505, 478 ], "score": 1.0, "content": "choose a different object category in the same image whose central point is close to the GT object", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 477, 505, 489 ], "spans": [ { "bbox": [ 105, 477, 248, 488 ], "score": 1.0, "content": "as the negative object, and replace", "type": "text" }, { "bbox": [ 249, 477, 287, 489 ], "score": 0.59, "content": "\\langle \\mathrm { c l a s s \\_ A } \\rangle", "type": "inline_equation" }, { "bbox": [ 287, 477, 306, 488 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 307, 477, 344, 488 ], "score": 0.77, "content": "\\left. \\mathrm { c l a s s \\mathbf { \\mathbf { B } } } \\right.", "type": "inline_equation" }, { "bbox": [ 345, 477, 505, 488 ], "score": 1.0, "content": "with those two randomly to form 2667", "type": "text" } ], "index": 42 }, { "bbox": [ 104, 486, 505, 501 ], "spans": [ { "bbox": [ 104, 486, 505, 501 ], "score": 1.0, "content": "questions. Additionally, to mimic the versatility of referring in human life, we replace the ⟨location⟩", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 499, 506, 511 ], "spans": [ { "bbox": [ 105, 499, 506, 511 ], "score": 1.0, "content": "with three different types: point, box, and free-form shape. For point, we randomly sample a point", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 509, 505, 522 ], "spans": [ { "bbox": [ 105, 509, 505, 522 ], "score": 1.0, "content": "inside the GT object that is also near the GT object’s boundary. For box, we use the GT bounding", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 520, 505, 533 ], "spans": [ { "bbox": [ 105, 520, 505, 533 ], "score": 1.0, "content": "box provided by LVIS. For the free-form shape, we randomly generate some strokes inside the GT", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 531, 505, 544 ], "spans": [ { "bbox": [ 105, 531, 505, 544 ], "score": 1.0, "content": "object to simulate that. Results on all three types of referring are summarized in Table 1. Ferret can", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 542, 505, 555 ], "spans": [ { "bbox": [ 105, 542, 505, 555 ], "score": 1.0, "content": "significantly outperform previous models (Peng et al., 2023; Chen et al., 2023b) and handle all types", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 554, 342, 566 ], "spans": [ { "bbox": [ 105, 554, 342, 566 ], "score": 1.0, "content": "of referring, a capability notably absent in previous works.", "type": "text" } ], "index": 49 } ], "index": 44 }, { "type": "title", "bbox": [ 107, 572, 224, 583 ], "lines": [ { "bbox": [ 105, 571, 225, 585 ], "spans": [ { "bbox": [ 105, 571, 225, 585 ], "score": 1.0, "content": "4.2 OUTPUT GROUNDING", "type": "text" } ], "index": 50 } ], "index": 50 }, { "type": "text", "bbox": [ 107, 587, 505, 632 ], "lines": [ { "bbox": [ 105, 586, 505, 600 ], "spans": [ { "bbox": [ 105, 586, 505, 600 ], "score": 1.0, "content": "Ferret performs well in referential dialogue, allowing for its integration into various VL tasks, no-", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 597, 506, 611 ], "spans": [ { "bbox": [ 105, 597, 506, 611 ], "score": 1.0, "content": "tably those with grounding outputs. To rigorously assess the grounding capability, we first subject", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 609, 505, 622 ], "spans": [ { "bbox": [ 105, 609, 505, 622 ], "score": 1.0, "content": "Ferret to benchmark visual grounding tasks in a generative paradigm. Then, to measure the align-", "type": "text" } ], "index": 53 }, { "bbox": [ 105, 621, 469, 632 ], "spans": [ { "bbox": [ 105, 621, 469, 632 ], "score": 1.0, "content": "ments between words and regions, we further evaluate Ferret on grounded captioning task.", "type": "text" } ], "index": 54 } ], "index": 52.5 }, { "type": "text", "bbox": [ 107, 637, 505, 704 ], "lines": [ { "bbox": [ 105, 636, 505, 650 ], "spans": [ { "bbox": [ 105, 636, 505, 650 ], "score": 1.0, "content": "Visual grounding. Visual grounding aims to ground language queries into aligned image regions.", "type": "text" } ], "index": 55 }, { "bbox": [ 106, 648, 505, 661 ], "spans": [ { "bbox": [ 106, 648, 505, 661 ], "score": 1.0, "content": "We experiment on the sub-tasks of referring expression comprehension (REC) with three renowned", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 658, 506, 672 ], "spans": [ { "bbox": [ 105, 658, 284, 672 ], "score": 1.0, "content": "benchmarks: RefCOCO (Lin et al., 2014),", "type": "text" }, { "bbox": [ 284, 659, 334, 670 ], "score": 0.29, "content": "\\operatorname { R e f C O C O + }", "type": "inline_equation" }, { "bbox": [ 335, 658, 426, 672 ], "score": 1.0, "content": "(Yu et al., 2016), and", "type": "text" }, { "bbox": [ 427, 659, 476, 671 ], "score": 0.36, "content": "\\operatorname { R e f C O C O g }", "type": "inline_equation" }, { "bbox": [ 477, 658, 506, 672 ], "score": 1.0, "content": "(Mao", "type": "text" } ], "index": 57 }, { "bbox": [ 105, 670, 505, 682 ], "spans": [ { "bbox": [ 105, 670, 505, 682 ], "score": 1.0, "content": "et al., 2016), and phrase grounding with Flickr30k Entities dataset (Plummer et al., 2015). REC", "type": "text" } ], "index": 58 }, { "bbox": [ 105, 681, 505, 694 ], "spans": [ { "bbox": [ 105, 681, 505, 694 ], "score": 1.0, "content": "task involves a question or description about a specific area in an image, with the model expected", "type": "text" } ], "index": 59 }, { "bbox": [ 105, 693, 505, 704 ], "spans": [ { "bbox": [ 105, 693, 505, 704 ], "score": 1.0, "content": "to predict just one bounding box. Phrase grounding, conversely, seeks to associate all the noun", "type": "text" } ], "index": 60 } ], "index": 57.5 } ], "page_idx": 6, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 712, 505, 732 ], "lines": [ { "bbox": [ 118, 709, 506, 724 ], "spans": [ { "bbox": [ 118, 709, 446, 724 ], "score": 1.0, "content": "7Sometimes both GT class and negative class appear in the answer, e.g., “The object is", "type": "text" }, { "bbox": [ 446, 711, 486, 722 ], "score": 0.67, "content": "\\langle \\mathrm { c l a s s \\mathrm { \\bf . G T } } \\rangle", "type": "inline_equation" }, { "bbox": [ 487, 709, 506, 724 ], "score": 1.0, "content": ", not", "type": "text" } ] }, { "bbox": [ 107, 721, 503, 733 ], "spans": [ { "bbox": [ 107, 721, 503, 733 ], "score": 1.0, "content": "⟨class Neg⟩”. Our rule removes the substring in-between “not” and comma/period, and then detects GT class.", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 759 ], "lines": [ { "bbox": [ 302, 750, 309, 762 ], "spans": [ { "bbox": [ 302, 750, 309, 762 ], "score": 1.0, "content": "7", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 107, 124, 273, 213 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 81, 274, 123 ], "group_id": 0, "lines": [ { "bbox": [ 106, 81, 275, 93 ], "spans": [ { "bbox": [ 106, 81, 275, 93 ], "score": 1.0, "content": "Table 1: Results of referring object classi-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 92, 275, 104 ], "spans": [ { "bbox": [ 106, 92, 275, 104 ], "score": 1.0, "content": "fication on three different referring types, in-", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 102, 273, 114 ], "spans": [ { "bbox": [ 106, 102, 260, 114 ], "score": 1.0, "content": "cluding point, box, and free-form shape.", "type": "text" }, { "bbox": [ 260, 102, 273, 112 ], "score": 0.41, "content": "\\mathbf { \\vec { \\mathbf { \\rho } } } \\mathbf { \\times } \\mathbf { \\vec { \\mathbf { \\rho } } }", "type": "inline_equation" } ], "index": 2 }, { "bbox": [ 106, 112, 201, 123 ], "spans": [ { "bbox": [ 106, 112, 201, 123 ], "score": 1.0, "content": "means no such capability.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "table_body", "bbox": [ 107, 124, 273, 213 ], "group_id": 0, "lines": [ { "bbox": [ 107, 124, 273, 213 ], "spans": [ { "bbox": [ 107, 124, 273, 213 ], "score": 0.979, "html": "
ModelsLVIS (Acc %)
PointBoxFree-form
Random Guess505050
LLaVA50.150.3×
Kosmos-2 (Peng et al.,2023)×60.25×
Shikra-7B(Chen et al.,2023b)57.82 67.71×
GPT4-ROI (Zhang et al.,2023)×61.76×
Ferret-7B67.9479.4269.77
Ferret-13B68.3580.4670.98
", "type": "table", "image_path": "d6bb85e1180177a92096c416e7a449b143cb3b05a6e56804b17bee6fbcb06413.jpg" } ] } ], "index": 13.0, "virtual_lines": [ { "bbox": [ 107, 124, 273, 138.83333333333334 ], "spans": [], "index": 8 }, { "bbox": [ 107, 138.83333333333334, 273, 153.66666666666669 ], "spans": [], "index": 10 }, { "bbox": [ 107, 153.66666666666669, 273, 168.50000000000003 ], "spans": [], "index": 12 }, { "bbox": [ 107, 168.50000000000003, 273, 183.33333333333337 ], "spans": [], "index": 14 }, { "bbox": [ 107, 183.33333333333337, 273, 198.1666666666667 ], "spans": [], "index": 16 }, { "bbox": [ 107, 198.1666666666667, 273, 213.00000000000006 ], "spans": [], "index": 18 } ] } ], "index": 7.25 }, { "type": "table", "bbox": [ 280, 125, 501, 213 ], "blocks": [ { "type": "table_caption", "bbox": [ 277, 81, 502, 122 ], "group_id": 1, "lines": [ { "bbox": [ 277, 80, 503, 93 ], "spans": [ { "bbox": [ 277, 80, 503, 93 ], "score": 1.0, "content": "Table 2: Results of grounded image captioning on the test", "type": "text" } ], "index": 4 }, { "bbox": [ 278, 91, 502, 101 ], "spans": [ { "bbox": [ 278, 91, 398, 101 ], "score": 1.0, "content": "set of Flickr30k Entities. BLEU", "type": "text" }, { "bbox": [ 399, 91, 411, 100 ], "score": 0.63, "content": "@ 4", "type": "inline_equation" }, { "bbox": [ 412, 91, 502, 101 ], "score": 1.0, "content": ", METEOR, CIDEr, and", "type": "text" } ], "index": 5 }, { "bbox": [ 277, 100, 501, 113 ], "spans": [ { "bbox": [ 277, 100, 438, 113 ], "score": 1.0, "content": "SPICE are used for the caption evaluation.", "type": "text" }, { "bbox": [ 439, 101, 461, 111 ], "score": 0.9, "content": "F 1 _ { a l l }", "type": "inline_equation" }, { "bbox": [ 461, 100, 478, 113 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 478, 101, 501, 111 ], "score": 0.9, "content": "F 1 _ { l o c }", "type": "inline_equation" } ], "index": 6 }, { "bbox": [ 277, 111, 490, 122 ], "spans": [ { "bbox": [ 277, 111, 404, 122 ], "score": 1.0, "content": "are used for grounding evaluation.", "type": "text" }, { "bbox": [ 405, 111, 415, 120 ], "score": 0.26, "content": "\" - \"", "type": "inline_equation" }, { "bbox": [ 416, 111, 490, 122 ], "score": 1.0, "content": "means not reported.", "type": "text" } ], "index": 7 } ], "index": 5.5 }, { "type": "table_body", "bbox": [ 280, 125, 501, 213 ], "group_id": 1, "lines": [ { "bbox": [ 280, 125, 501, 213 ], "spans": [ { "bbox": [ 280, 125, 501, 213 ], "score": 0.98, "html": "
ModelsCaption Eval.Grounding Eval.
B@4MCSF1allF1loc
GVD (Zhou et al.,2019)27.3 22.5 62.3 16.57.5522.2
Cyclical (Ma et al.,2020)26.8 22.4 61.1 16.88.4422.78
POS-SCAN (Zhou et al., 2020)30.1 22.6 69.3 16.87.1717.49
UniTAB (Yang et al., 2022)30.1 23.7 69.7 17.412.9534.79
Shikra-13B (Chen et al.,2023b)1173.9111
Ferret-7B35.1 24.6 74.8 18.015.0237.62
Ferret-13B37.0 25.5 76.1 18.315.1238.03
", "type": "table", "image_path": "bafe86a4352ddc05a3896487ffd7534bf2f5780b5f85b54c1f3dd76763561a93.jpg" } ] } ], "index": 14.0, "virtual_lines": [ { "bbox": [ 280, 125, 501, 139.66666666666666 ], "spans": [], "index": 9 }, { "bbox": [ 280, 139.66666666666666, 501, 154.33333333333331 ], "spans": [], "index": 11 }, { "bbox": [ 280, 154.33333333333331, 501, 168.99999999999997 ], "spans": [], "index": 13 }, { "bbox": [ 280, 168.99999999999997, 501, 183.66666666666663 ], "spans": [], "index": 15 }, { "bbox": [ 280, 183.66666666666663, 501, 198.3333333333333 ], "spans": [], "index": 17 }, { "bbox": [ 280, 198.3333333333333, 501, 212.99999999999994 ], "spans": [], "index": 19 } ] } ], "index": 9.75 }, { "type": "text", "bbox": [ 107, 217, 505, 306 ], "lines": [ { "bbox": [ 105, 217, 505, 230 ], "spans": [ { "bbox": [ 105, 217, 376, 230 ], "score": 1.0, "content": "Training Details. We initialize the image encoder with CLIP-ViT-", "type": "text" }, { "bbox": [ 377, 218, 426, 229 ], "score": 0.4, "content": "\\mathrm { L } / 1 4 @ 3 3 6 \\mathrm { p }", "type": "inline_equation" }, { "bbox": [ 426, 217, 505, 230 ], "score": 1.0, "content": ", the LLM with Vi-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 228, 505, 240 ], "spans": [ { "bbox": [ 105, 228, 505, 240 ], "score": 1.0, "content": "cuna, and the projection layer with LLaVA’s first-stage weights, leaving the visual sampler ran-", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 240, 505, 250 ], "spans": [ { "bbox": [ 106, 240, 505, 250 ], "score": 1.0, "content": "domly initialized. After the initialization, Ferret is trained on the aforementioned GRIT data for", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 250, 505, 263 ], "spans": [ { "bbox": [ 106, 250, 426, 263 ], "score": 1.0, "content": "three epochs, optimized by Loshchilov & Hutter (2017) with a learning rate of", "type": "text" }, { "bbox": [ 426, 250, 455, 261 ], "score": 0.88, "content": "2 e - 5", "type": "inline_equation" }, { "bbox": [ 455, 250, 505, 263 ], "score": 1.0, "content": "and a batch", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 259, 505, 275 ], "spans": [ { "bbox": [ 105, 259, 232, 275 ], "score": 1.0, "content": "size of 128. The training takes", "type": "text" }, { "bbox": [ 232, 261, 262, 272 ], "score": 0.91, "content": "{ \\sim } 5 / 2 . 5", "type": "inline_equation" }, { "bbox": [ 263, 259, 505, 275 ], "score": 1.0, "content": "days on 8 A100 GPU for a Ferret-13B/7B. During training,", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 272, 505, 285 ], "spans": [ { "bbox": [ 106, 272, 505, 285 ], "score": 1.0, "content": "when input refers to regions, we randomly choose either the center points or the bounding boxes", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 283, 505, 296 ], "spans": [ { "bbox": [ 105, 283, 505, 296 ], "score": 1.0, "content": "(or segmentation masks if available) to represent the regions. We perform de-duplication in training", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 294, 360, 306 ], "spans": [ { "bbox": [ 106, 294, 360, 306 ], "score": 1.0, "content": "data to remove the samples that are in downstream evaluations.", "type": "text" } ], "index": 27 } ], "index": 23.5, "bbox_fs": [ 105, 217, 505, 306 ] }, { "type": "title", "bbox": [ 107, 313, 211, 324 ], "lines": [ { "bbox": [ 105, 312, 212, 325 ], "spans": [ { "bbox": [ 105, 312, 212, 325 ], "score": 1.0, "content": "4.1 INPUT REFERRING", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 106, 327, 505, 438 ], "lines": [ { "bbox": [ 106, 328, 505, 340 ], "spans": [ { "bbox": [ 106, 328, 505, 340 ], "score": 1.0, "content": "The model’s capability of understanding referring is reflected in that, given a referred region in the", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 339, 506, 353 ], "spans": [ { "bbox": [ 105, 339, 506, 353 ], "score": 1.0, "content": "question, how accurately the model can understand the semantics of the referred region. To measure", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 350, 505, 363 ], "spans": [ { "bbox": [ 105, 350, 505, 363 ], "score": 1.0, "content": "it, we start with the most basic semantics, object, as it is fundamental and clear to define. To be more", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 361, 505, 374 ], "spans": [ { "bbox": [ 105, 361, 505, 374 ], "score": 1.0, "content": "specific, the task we evaluate on is Referring Object Classification: the question refers to a specific", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 372, 505, 385 ], "spans": [ { "bbox": [ 105, 372, 505, 385 ], "score": 1.0, "content": "region in the image, and the model needs to classify the object in the region. Since Ferret and", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 383, 505, 396 ], "spans": [ { "bbox": [ 105, 383, 505, 396 ], "score": 1.0, "content": "MLLMs usually generate free-form text responses, it is inaccurate to match the predicted class with", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 393, 506, 407 ], "spans": [ { "bbox": [ 105, 393, 506, 407 ], "score": 1.0, "content": "the ground-truth class if directly asking the model to classify without constraints. Alternatively,", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 404, 506, 417 ], "spans": [ { "bbox": [ 105, 404, 506, 417 ], "score": 1.0, "content": "we make it a binary-choice question in the format of “Is the object ⟨location⟩ a ⟨class A⟩ or a", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 414, 506, 430 ], "spans": [ { "bbox": [ 106, 414, 506, 430 ], "score": 1.0, "content": "⟨class B⟩?”. We feed the binary-choice question and image into the MLLMs to obtain the response,", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 426, 437, 439 ], "spans": [ { "bbox": [ 105, 426, 437, 439 ], "score": 1.0, "content": "and then detect if the response matches the ground-truth (GT) class by some rule.7", "type": "text" } ], "index": 38 } ], "index": 33.5, "bbox_fs": [ 105, 328, 506, 439 ] }, { "type": "text", "bbox": [ 107, 443, 505, 565 ], "lines": [ { "bbox": [ 105, 442, 505, 457 ], "spans": [ { "bbox": [ 105, 442, 505, 457 ], "score": 1.0, "content": "To prepare the data, we used the validation split of LVIS dataset (Gupta et al., 2019) covering", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 455, 505, 468 ], "spans": [ { "bbox": [ 105, 455, 505, 468 ], "score": 1.0, "content": "over 1000 object categories, and sampled 2667 objects as the GT objects. Then, we randomly", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 466, 505, 478 ], "spans": [ { "bbox": [ 105, 466, 505, 478 ], "score": 1.0, "content": "choose a different object category in the same image whose central point is close to the GT object", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 477, 505, 489 ], "spans": [ { "bbox": [ 105, 477, 248, 488 ], "score": 1.0, "content": "as the negative object, and replace", "type": "text" }, { "bbox": [ 249, 477, 287, 489 ], "score": 0.59, "content": "\\langle \\mathrm { c l a s s \\_ A } \\rangle", "type": "inline_equation" }, { "bbox": [ 287, 477, 306, 488 ], "score": 1.0, "content": "and", "type": "text" }, { "bbox": [ 307, 477, 344, 488 ], "score": 0.77, "content": "\\left. \\mathrm { c l a s s \\mathbf { \\mathbf { B } } } \\right.", "type": "inline_equation" }, { "bbox": [ 345, 477, 505, 488 ], "score": 1.0, "content": "with those two randomly to form 2667", "type": "text" } ], "index": 42 }, { "bbox": [ 104, 486, 505, 501 ], "spans": [ { "bbox": [ 104, 486, 505, 501 ], "score": 1.0, "content": "questions. Additionally, to mimic the versatility of referring in human life, we replace the ⟨location⟩", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 499, 506, 511 ], "spans": [ { "bbox": [ 105, 499, 506, 511 ], "score": 1.0, "content": "with three different types: point, box, and free-form shape. For point, we randomly sample a point", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 509, 505, 522 ], "spans": [ { "bbox": [ 105, 509, 505, 522 ], "score": 1.0, "content": "inside the GT object that is also near the GT object’s boundary. For box, we use the GT bounding", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 520, 505, 533 ], "spans": [ { "bbox": [ 105, 520, 505, 533 ], "score": 1.0, "content": "box provided by LVIS. For the free-form shape, we randomly generate some strokes inside the GT", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 531, 505, 544 ], "spans": [ { "bbox": [ 105, 531, 505, 544 ], "score": 1.0, "content": "object to simulate that. Results on all three types of referring are summarized in Table 1. Ferret can", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 542, 505, 555 ], "spans": [ { "bbox": [ 105, 542, 505, 555 ], "score": 1.0, "content": "significantly outperform previous models (Peng et al., 2023; Chen et al., 2023b) and handle all types", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 554, 342, 566 ], "spans": [ { "bbox": [ 105, 554, 342, 566 ], "score": 1.0, "content": "of referring, a capability notably absent in previous works.", "type": "text" } ], "index": 49 } ], "index": 44, "bbox_fs": [ 104, 442, 506, 566 ] }, { "type": "title", "bbox": [ 107, 572, 224, 583 ], "lines": [ { "bbox": [ 105, 571, 225, 585 ], "spans": [ { "bbox": [ 105, 571, 225, 585 ], "score": 1.0, "content": "4.2 OUTPUT GROUNDING", "type": "text" } ], "index": 50 } ], "index": 50 }, { "type": "text", "bbox": [ 107, 587, 505, 632 ], "lines": [ { "bbox": [ 105, 586, 505, 600 ], "spans": [ { "bbox": [ 105, 586, 505, 600 ], "score": 1.0, "content": "Ferret performs well in referential dialogue, allowing for its integration into various VL tasks, no-", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 597, 506, 611 ], "spans": [ { "bbox": [ 105, 597, 506, 611 ], "score": 1.0, "content": "tably those with grounding outputs. To rigorously assess the grounding capability, we first subject", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 609, 505, 622 ], "spans": [ { "bbox": [ 105, 609, 505, 622 ], "score": 1.0, "content": "Ferret to benchmark visual grounding tasks in a generative paradigm. Then, to measure the align-", "type": "text" } ], "index": 53 }, { "bbox": [ 105, 621, 469, 632 ], "spans": [ { "bbox": [ 105, 621, 469, 632 ], "score": 1.0, "content": "ments between words and regions, we further evaluate Ferret on grounded captioning task.", "type": "text" } ], "index": 54 } ], "index": 52.5, "bbox_fs": [ 105, 586, 506, 632 ] }, { "type": "text", "bbox": [ 107, 637, 505, 704 ], "lines": [ { "bbox": [ 105, 636, 505, 650 ], "spans": [ { "bbox": [ 105, 636, 505, 650 ], "score": 1.0, "content": "Visual grounding. Visual grounding aims to ground language queries into aligned image regions.", "type": "text" } ], "index": 55 }, { "bbox": [ 106, 648, 505, 661 ], "spans": [ { "bbox": [ 106, 648, 505, 661 ], "score": 1.0, "content": "We experiment on the sub-tasks of referring expression comprehension (REC) with three renowned", "type": "text" } ], "index": 56 }, { "bbox": [ 105, 658, 506, 672 ], "spans": [ { "bbox": [ 105, 658, 284, 672 ], "score": 1.0, "content": "benchmarks: RefCOCO (Lin et al., 2014),", "type": "text" }, { "bbox": [ 284, 659, 334, 670 ], "score": 0.29, "content": "\\operatorname { R e f C O C O + }", "type": "inline_equation" }, { "bbox": [ 335, 658, 426, 672 ], "score": 1.0, "content": "(Yu et al., 2016), and", "type": "text" }, { "bbox": [ 427, 659, 476, 671 ], "score": 0.36, "content": "\\operatorname { R e f C O C O g }", "type": "inline_equation" }, { "bbox": [ 477, 658, 506, 672 ], "score": 1.0, "content": "(Mao", "type": "text" } ], "index": 57 }, { "bbox": [ 105, 670, 505, 682 ], "spans": [ { "bbox": [ 105, 670, 505, 682 ], "score": 1.0, "content": "et al., 2016), and phrase grounding with Flickr30k Entities dataset (Plummer et al., 2015). REC", "type": "text" } ], "index": 58 }, { "bbox": [ 105, 681, 505, 694 ], "spans": [ { "bbox": [ 105, 681, 505, 694 ], "score": 1.0, "content": "task involves a question or description about a specific area in an image, with the model expected", "type": "text" } ], "index": 59 }, { "bbox": [ 105, 693, 505, 704 ], "spans": [ { "bbox": [ 105, 693, 505, 704 ], "score": 1.0, "content": "to predict just one bounding box. Phrase grounding, conversely, seeks to associate all the noun", "type": "text" } ], "index": 60 }, { "bbox": [ 105, 263, 505, 276 ], "spans": [ { "bbox": [ 105, 263, 505, 276 ], "score": 1.0, "content": "phrases in the input sentence with corresponding boxes, requiring the model to predict these boxes", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 105, 274, 505, 286 ], "spans": [ { "bbox": [ 105, 274, 505, 286 ], "score": 1.0, "content": "and the word-box connections. For both tasks, we utilize uniform prompts, represented as “What are", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 105, 284, 505, 299 ], "spans": [ { "bbox": [ 105, 284, 172, 299 ], "score": 1.0, "content": "the locations of", "type": "text", "cross_page": true }, { "bbox": [ 172, 286, 268, 297 ], "score": 0.31, "content": "< q u e r y > / < p h r a s e s > ? ^ { , }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 269, 284, 300, 299 ], "score": 1.0, "content": ", where", "type": "text", "cross_page": true }, { "bbox": [ 301, 286, 340, 297 ], "score": 0.46, "content": "< q u e r y >", "type": "inline_equation", "cross_page": true }, { "bbox": [ 341, 284, 505, 299 ], "score": 1.0, "content": "denotes the textual referring expression,", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 105, 297, 506, 308 ], "spans": [ { "bbox": [ 105, 297, 132, 308 ], "score": 1.0, "content": "while", "type": "text", "cross_page": true }, { "bbox": [ 132, 297, 180, 308 ], "score": 0.57, "content": "< p h r a s e s >", "type": "inline_equation", "cross_page": true }, { "bbox": [ 180, 297, 506, 308 ], "score": 1.0, "content": "stands for a “comma-delimited” aggregation of the given phrases. The model is", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 105, 306, 506, 320 ], "spans": [ { "bbox": [ 105, 306, 506, 320 ], "score": 1.0, "content": "trained to output in “ [box].” format. The generated bounding box is considered correct if", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 105, 318, 506, 330 ], "spans": [ { "bbox": [ 105, 318, 506, 330 ], "score": 1.0, "content": "its intersection over union (IoU) with the GT box is greater than 0.5. As shown in Table 3, Ferret", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 105, 328, 505, 342 ], "spans": [ { "bbox": [ 105, 328, 505, 342 ], "score": 1.0, "content": "achieves an outstanding performance on all metrics, and is comparable to specialized fine-tuning", "type": "text", "cross_page": true } ], "index": 12 }, { "bbox": [ 105, 340, 244, 352 ], "spans": [ { "bbox": [ 105, 340, 244, 352 ], "score": 1.0, "content": "approaches (Kamath et al., 2021).", "type": "text", "cross_page": true } ], "index": 13 } ], "index": 57.5, "bbox_fs": [ 105, 636, 506, 704 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 115, 115, 492, 251 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 81, 502, 111 ], "group_id": 0, "lines": [ { "bbox": [ 106, 80, 504, 92 ], "spans": [ { "bbox": [ 106, 80, 239, 92 ], "score": 1.0, "content": "Table 3: Performance comparison", "type": "text" }, { "bbox": [ 239, 81, 279, 91 ], "score": 0.82, "content": "( \\operatorname { A c c } @ 0 . 5 )", "type": "inline_equation" }, { "bbox": [ 279, 80, 504, 92 ], "score": 1.0, "content": "on the referring expression comprehension (RefCOCO, Ref-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 91, 504, 102 ], "spans": [ { "bbox": [ 106, 91, 137, 101 ], "score": 0.43, "content": "\\mathrm { C O C O + }", "type": "inline_equation" }, { "bbox": [ 138, 91, 366, 102 ], "score": 1.0, "content": ", RefCOCOg) and phrase grounding (Flickr30k Entities) tasks.", "type": "text" }, { "bbox": [ 367, 92, 373, 100 ], "score": 0.7, "content": "^ *", "type": "inline_equation" }, { "bbox": [ 374, 91, 504, 102 ], "score": 1.0, "content": "indicates that the method is specifi-", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 100, 237, 113 ], "spans": [ { "bbox": [ 105, 100, 237, 113 ], "score": 1.0, "content": "cally fine-tuned in the second stage.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 115, 115, 492, 251 ], "group_id": 0, "lines": [ { "bbox": [ 115, 115, 492, 251 ], "spans": [ { "bbox": [ 115, 115, 492, 251 ], "score": 0.984, "html": "
Modelsval RefCOCOtestBvaRefCOCOtestBRefCOCOgFlickr30k Enties
MAttNet (Yu et al., 2018)76.40 80.43 69.28|64.93 70.26 56.00|66.67 67.011
OFA-L (Wang et al., 2022b)79.9683.67 76.3968.2976.00 61.7567.57 67.581
TransVG (Deng et al., 2021)81.0282.7278.3564.8270.70 56.9468.6767.7379.10
UNITER (Chen et al., 2020)81.4187.0474.1775.90 81.45 66.7074.02 68.6711
VILLA (Gan et al., 2020)82.3987.48 74.8476.17 81.54 66.8476.18 76.71/1
UniTAB (Yang et al., 2022)86.3288.8480.6178.70 83.2269.4879.96 79.9778.7679.58
MDETR (Kamath et al.,2021)86.7589.58 81.4179.52 84.09 70.6281.64 80.8982.3*83.8*
Shikra-7B (Chen et al.,2023b)87.01 90.6180.24|81.60 87.36 72.1282.27 82.1975.8476.54
Ferret-7B87.49 91.35 82.4580.78 87.38 73.1483.93 84.7680.3982.21
Shikra-13B(Chen et al.,2023b) 87.83 91.11 81.81|82.89 87.79 74.41|82.64 83.16[77.4178.44
Ferret-13B89.4892.41 84.3682.81 88.14 75.1785.83 86.3481.1384.76
", "type": "table", "image_path": "ab9a70b159c976205ce90d7e462cd4e3241a103b88c2c862e4ab4f63c696961e.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 115, 115, 492, 160.33333333333334 ], "spans": [], "index": 3 }, { "bbox": [ 115, 160.33333333333334, 492, 205.66666666666669 ], "spans": [], "index": 4 }, { "bbox": [ 115, 205.66666666666669, 492, 251.00000000000003 ], "spans": [], "index": 5 } ] } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 263, 505, 351 ], "lines": [ { "bbox": [ 105, 263, 505, 276 ], "spans": [ { "bbox": [ 105, 263, 505, 276 ], "score": 1.0, "content": "phrases in the input sentence with corresponding boxes, requiring the model to predict these boxes", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 274, 505, 286 ], "spans": [ { "bbox": [ 105, 274, 505, 286 ], "score": 1.0, "content": "and the word-box connections. For both tasks, we utilize uniform prompts, represented as “What are", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 284, 505, 299 ], "spans": [ { "bbox": [ 105, 284, 172, 299 ], "score": 1.0, "content": "the locations of", "type": "text" }, { "bbox": [ 172, 286, 268, 297 ], "score": 0.31, "content": "< q u e r y > / < p h r a s e s > ? ^ { , }", "type": "inline_equation" }, { "bbox": [ 269, 284, 300, 299 ], "score": 1.0, "content": ", where", "type": "text" }, { "bbox": [ 301, 286, 340, 297 ], "score": 0.46, "content": "< q u e r y >", "type": "inline_equation" }, { "bbox": [ 341, 284, 505, 299 ], "score": 1.0, "content": "denotes the textual referring expression,", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 297, 506, 308 ], "spans": [ { "bbox": [ 105, 297, 132, 308 ], "score": 1.0, "content": "while", "type": "text" }, { "bbox": [ 132, 297, 180, 308 ], "score": 0.57, "content": "< p h r a s e s >", "type": "inline_equation" }, { "bbox": [ 180, 297, 506, 308 ], "score": 1.0, "content": "stands for a “comma-delimited” aggregation of the given phrases. The model is", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 306, 506, 320 ], "spans": [ { "bbox": [ 105, 306, 506, 320 ], "score": 1.0, "content": "trained to output in “ [box].” format. The generated bounding box is considered correct if", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 318, 506, 330 ], "spans": [ { "bbox": [ 105, 318, 506, 330 ], "score": 1.0, "content": "its intersection over union (IoU) with the GT box is greater than 0.5. As shown in Table 3, Ferret", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 328, 505, 342 ], "spans": [ { "bbox": [ 105, 328, 505, 342 ], "score": 1.0, "content": "achieves an outstanding performance on all metrics, and is comparable to specialized fine-tuning", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 340, 244, 352 ], "spans": [ { "bbox": [ 105, 340, 244, 352 ], "score": 1.0, "content": "approaches (Kamath et al., 2021).", "type": "text" } ], "index": 13 } ], "index": 9.5 }, { "type": "text", "bbox": [ 106, 356, 505, 445 ], "lines": [ { "bbox": [ 106, 357, 505, 369 ], "spans": [ { "bbox": [ 106, 357, 505, 369 ], "score": 1.0, "content": "Grounded captioning. The grounded captioning task requires the model to generate a caption and", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 368, 506, 380 ], "spans": [ { "bbox": [ 105, 368, 506, 380 ], "score": 1.0, "content": "ground all generated noun phrases to image regions. The final predictions generally consist of three", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 379, 505, 391 ], "spans": [ { "bbox": [ 105, 379, 505, 391 ], "score": 1.0, "content": "parts, i.e., the text caption, visual regions as boxes, and the grounding alignments between words", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 390, 505, 402 ], "spans": [ { "bbox": [ 106, 390, 505, 402 ], "score": 1.0, "content": "and boxes. Following the established benchmarks on the Flickr30k Entities dataset, we evaluate", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 401, 505, 414 ], "spans": [ { "bbox": [ 106, 401, 505, 414 ], "score": 1.0, "content": "captioning and grounding separately with the captioning metrics and grounding F1 scores, respec-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 411, 506, 425 ], "spans": [ { "bbox": [ 105, 411, 133, 425 ], "score": 1.0, "content": "tively.", "type": "text" }, { "bbox": [ 134, 412, 158, 423 ], "score": 0.9, "content": "F 1 _ { a l l }", "type": "inline_equation" }, { "bbox": [ 159, 411, 461, 425 ], "score": 1.0, "content": "evaluates grounding as a multi-label classification problem. We also report", "type": "text" }, { "bbox": [ 462, 412, 486, 423 ], "score": 0.91, "content": "F 1 _ { l o c }", "type": "inline_equation" }, { "bbox": [ 487, 411, 506, 425 ], "score": 1.0, "content": "that", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 422, 506, 435 ], "spans": [ { "bbox": [ 105, 422, 506, 435 ], "score": 1.0, "content": "only computes the grounding score on correctly predicted object words. Results are summarized in", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 433, 287, 445 ], "spans": [ { "bbox": [ 106, 433, 287, 445 ], "score": 1.0, "content": "Table 2, and Ferret achieves state-of-the-art.", "type": "text" } ], "index": 21 } ], "index": 17.5 }, { "type": "title", "bbox": [ 110, 447, 469, 459 ], "lines": [ { "bbox": [ 108, 447, 470, 459 ], "spans": [ { "bbox": [ 108, 447, 470, 459 ], "score": 1.0, "content": "4.3 FERRET-BENCH: MULTIMODAL CHATTING WITH REFERRING AND GROUNDING", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 106, 462, 505, 594 ], "lines": [ { "bbox": [ 105, 461, 506, 475 ], "spans": [ { "bbox": [ 105, 461, 506, 475 ], "score": 1.0, "content": "Multimodal chatting has been an emergent ability of MLLMs. Previous benchmarks (Liu et al.,", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 473, 506, 487 ], "spans": [ { "bbox": [ 105, 473, 506, 487 ], "score": 1.0, "content": "2023b) mainly evaluate conversation, detailed description, and complex reasoning via GPT-4 as", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 485, 505, 497 ], "spans": [ { "bbox": [ 105, 485, 505, 497 ], "score": 1.0, "content": "a judge. Yet, a gap exists as no dataset currently evaluates multimodal chatting that necessitates", "type": "text" } ], "index": 25 }, { "bbox": [ 104, 495, 505, 508 ], "spans": [ { "bbox": [ 104, 495, 505, 508 ], "score": 1.0, "content": "referring or grounding actions, e.g., instances where individuals reference an unfamiliar object and", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 506, 505, 519 ], "spans": [ { "bbox": [ 106, 506, 505, 519 ], "score": 1.0, "content": "inquire about its purpose. To benchmark this intriguing and practical capability, we introduce Ferret-", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 517, 505, 530 ], "spans": [ { "bbox": [ 105, 517, 505, 530 ], "score": 1.0, "content": "Bench that covers three kinds of region-based questions evaluating referring and grounding capa-", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 528, 506, 541 ], "spans": [ { "bbox": [ 106, 528, 506, 541 ], "score": 1.0, "content": "bility: (i) Referring Description: models are asked to describe a referred region based on its in-", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 539, 506, 552 ], "spans": [ { "bbox": [ 105, 539, 506, 552 ], "score": 1.0, "content": "teraction with surrounding objects. (ii) Referring Reasoning: models need to reason on top of", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 550, 506, 563 ], "spans": [ { "bbox": [ 105, 550, 506, 563 ], "score": 1.0, "content": "one or more referred regions correctly. (iii) Grounding in Conversation: models are required to", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 561, 506, 575 ], "spans": [ { "bbox": [ 105, 561, 506, 575 ], "score": 1.0, "content": "reason correctly and accurately ground/localize the objects/regions necessary for the reasoning. For", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 572, 506, 586 ], "spans": [ { "bbox": [ 105, 572, 506, 586 ], "score": 1.0, "content": "the ease of benchmarking other methods, we represent the regions with boxes instead of points or", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 582, 178, 596 ], "spans": [ { "bbox": [ 105, 582, 178, 596 ], "score": 1.0, "content": "free-form shapes.", "type": "text" } ], "index": 34 } ], "index": 28.5 }, { "type": "text", "bbox": [ 106, 600, 505, 732 ], "lines": [ { "bbox": [ 106, 600, 505, 613 ], "spans": [ { "bbox": [ 106, 600, 505, 613 ], "score": 1.0, "content": "Specifically, we randomly sample 40 images from the COCO validation set for each type of question,", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 611, 505, 623 ], "spans": [ { "bbox": [ 106, 611, 505, 623 ], "score": 1.0, "content": "and generate the questions and GPT-4’s answers following the instruction generation pipeline in", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 622, 505, 634 ], "spans": [ { "bbox": [ 106, 622, 505, 634 ], "score": 1.0, "content": "Sec. 3.2. Following Liu et al. (2023b), we feed the question and image into MLLMs to obtain", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 633, 505, 645 ], "spans": [ { "bbox": [ 106, 633, 505, 645 ], "score": 1.0, "content": "the predicted answer, and then prompt GPT-4 to rate the predicted answer and pseudo answer from", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 643, 505, 657 ], "spans": [ { "bbox": [ 105, 643, 505, 657 ], "score": 1.0, "content": "GPT-4 based on the ground-truth textual scene description (object, relationship, region caption,", "type": "text" } ], "index": 39 }, { "bbox": [ 104, 654, 505, 668 ], "spans": [ { "bbox": [ 104, 654, 505, 668 ], "score": 1.0, "content": "global caption). GPT-4 evaluates both the precision of referring understanding, object grounding,", "type": "text" } ], "index": 40 }, { "bbox": [ 104, 664, 505, 679 ], "spans": [ { "bbox": [ 104, 664, 505, 679 ], "score": 1.0, "content": "and correctness of semantics. The rating score ranges from 1 to 10, in which higher means better.", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "We calculate the ratio of the predicted answer’s score and the GPT-4 answer’s score, which is then", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 687, 506, 701 ], "spans": [ { "bbox": [ 105, 687, 506, 701 ], "score": 1.0, "content": "presented as a percentage to measure the performance of MLLMs. We also asked GPT-4 to give a", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "comprehensive review for the rating and found that GPT-4 is good at measuring the degree of spatial", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "precision, such as how much the predicted bounding box diverges from the GT box coordinate. We", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 721, 329, 733 ], "spans": [ { "bbox": [ 105, 721, 329, 733 ], "score": 1.0, "content": "refer the readers to Appendix D for further elaboration.", "type": "text" } ], "index": 46 } ], "index": 40.5 } ], "page_idx": 7, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 750, 309, 761 ], "spans": [ { "bbox": [ 302, 750, 309, 761 ], "score": 1.0, "content": "8", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 115, 115, 492, 251 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 81, 502, 111 ], "group_id": 0, "lines": [ { "bbox": [ 106, 80, 504, 92 ], "spans": [ { "bbox": [ 106, 80, 239, 92 ], "score": 1.0, "content": "Table 3: Performance comparison", "type": "text" }, { "bbox": [ 239, 81, 279, 91 ], "score": 0.82, "content": "( \\operatorname { A c c } @ 0 . 5 )", "type": "inline_equation" }, { "bbox": [ 279, 80, 504, 92 ], "score": 1.0, "content": "on the referring expression comprehension (RefCOCO, Ref-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 91, 504, 102 ], "spans": [ { "bbox": [ 106, 91, 137, 101 ], "score": 0.43, "content": "\\mathrm { C O C O + }", "type": "inline_equation" }, { "bbox": [ 138, 91, 366, 102 ], "score": 1.0, "content": ", RefCOCOg) and phrase grounding (Flickr30k Entities) tasks.", "type": "text" }, { "bbox": [ 367, 92, 373, 100 ], "score": 0.7, "content": "^ *", "type": "inline_equation" }, { "bbox": [ 374, 91, 504, 102 ], "score": 1.0, "content": "indicates that the method is specifi-", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 100, 237, 113 ], "spans": [ { "bbox": [ 105, 100, 237, 113 ], "score": 1.0, "content": "cally fine-tuned in the second stage.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "table_body", "bbox": [ 115, 115, 492, 251 ], "group_id": 0, "lines": [ { "bbox": [ 115, 115, 492, 251 ], "spans": [ { "bbox": [ 115, 115, 492, 251 ], "score": 0.984, "html": "
Modelsval RefCOCOtestBvaRefCOCOtestBRefCOCOgFlickr30k Enties
MAttNet (Yu et al., 2018)76.40 80.43 69.28|64.93 70.26 56.00|66.67 67.011
OFA-L (Wang et al., 2022b)79.9683.67 76.3968.2976.00 61.7567.57 67.581
TransVG (Deng et al., 2021)81.0282.7278.3564.8270.70 56.9468.6767.7379.10
UNITER (Chen et al., 2020)81.4187.0474.1775.90 81.45 66.7074.02 68.6711
VILLA (Gan et al., 2020)82.3987.48 74.8476.17 81.54 66.8476.18 76.71/1
UniTAB (Yang et al., 2022)86.3288.8480.6178.70 83.2269.4879.96 79.9778.7679.58
MDETR (Kamath et al.,2021)86.7589.58 81.4179.52 84.09 70.6281.64 80.8982.3*83.8*
Shikra-7B (Chen et al.,2023b)87.01 90.6180.24|81.60 87.36 72.1282.27 82.1975.8476.54
Ferret-7B87.49 91.35 82.4580.78 87.38 73.1483.93 84.7680.3982.21
Shikra-13B(Chen et al.,2023b) 87.83 91.11 81.81|82.89 87.79 74.41|82.64 83.16[77.4178.44
Ferret-13B89.4892.41 84.3682.81 88.14 75.1785.83 86.3481.1384.76
", "type": "table", "image_path": "ab9a70b159c976205ce90d7e462cd4e3241a103b88c2c862e4ab4f63c696961e.jpg" } ] } ], "index": 4, "virtual_lines": [ { "bbox": [ 115, 115, 492, 160.33333333333334 ], "spans": [], "index": 3 }, { "bbox": [ 115, 160.33333333333334, 492, 205.66666666666669 ], "spans": [], "index": 4 }, { "bbox": [ 115, 205.66666666666669, 492, 251.00000000000003 ], "spans": [], "index": 5 } ] } ], "index": 2.5 }, { "type": "text", "bbox": [ 107, 263, 505, 351 ], "lines": [], "index": 9.5, "bbox_fs": [ 105, 263, 506, 352 ], "lines_deleted": true }, { "type": "text", "bbox": [ 106, 356, 505, 445 ], "lines": [ { "bbox": [ 106, 357, 505, 369 ], "spans": [ { "bbox": [ 106, 357, 505, 369 ], "score": 1.0, "content": "Grounded captioning. The grounded captioning task requires the model to generate a caption and", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 368, 506, 380 ], "spans": [ { "bbox": [ 105, 368, 506, 380 ], "score": 1.0, "content": "ground all generated noun phrases to image regions. The final predictions generally consist of three", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 379, 505, 391 ], "spans": [ { "bbox": [ 105, 379, 505, 391 ], "score": 1.0, "content": "parts, i.e., the text caption, visual regions as boxes, and the grounding alignments between words", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 390, 505, 402 ], "spans": [ { "bbox": [ 106, 390, 505, 402 ], "score": 1.0, "content": "and boxes. Following the established benchmarks on the Flickr30k Entities dataset, we evaluate", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 401, 505, 414 ], "spans": [ { "bbox": [ 106, 401, 505, 414 ], "score": 1.0, "content": "captioning and grounding separately with the captioning metrics and grounding F1 scores, respec-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 411, 506, 425 ], "spans": [ { "bbox": [ 105, 411, 133, 425 ], "score": 1.0, "content": "tively.", "type": "text" }, { "bbox": [ 134, 412, 158, 423 ], "score": 0.9, "content": "F 1 _ { a l l }", "type": "inline_equation" }, { "bbox": [ 159, 411, 461, 425 ], "score": 1.0, "content": "evaluates grounding as a multi-label classification problem. We also report", "type": "text" }, { "bbox": [ 462, 412, 486, 423 ], "score": 0.91, "content": "F 1 _ { l o c }", "type": "inline_equation" }, { "bbox": [ 487, 411, 506, 425 ], "score": 1.0, "content": "that", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 422, 506, 435 ], "spans": [ { "bbox": [ 105, 422, 506, 435 ], "score": 1.0, "content": "only computes the grounding score on correctly predicted object words. Results are summarized in", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 433, 287, 445 ], "spans": [ { "bbox": [ 106, 433, 287, 445 ], "score": 1.0, "content": "Table 2, and Ferret achieves state-of-the-art.", "type": "text" } ], "index": 21 } ], "index": 17.5, "bbox_fs": [ 105, 357, 506, 445 ] }, { "type": "title", "bbox": [ 110, 447, 469, 459 ], "lines": [ { "bbox": [ 108, 447, 470, 459 ], "spans": [ { "bbox": [ 108, 447, 470, 459 ], "score": 1.0, "content": "4.3 FERRET-BENCH: MULTIMODAL CHATTING WITH REFERRING AND GROUNDING", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 106, 462, 505, 594 ], "lines": [ { "bbox": [ 105, 461, 506, 475 ], "spans": [ { "bbox": [ 105, 461, 506, 475 ], "score": 1.0, "content": "Multimodal chatting has been an emergent ability of MLLMs. Previous benchmarks (Liu et al.,", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 473, 506, 487 ], "spans": [ { "bbox": [ 105, 473, 506, 487 ], "score": 1.0, "content": "2023b) mainly evaluate conversation, detailed description, and complex reasoning via GPT-4 as", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 485, 505, 497 ], "spans": [ { "bbox": [ 105, 485, 505, 497 ], "score": 1.0, "content": "a judge. Yet, a gap exists as no dataset currently evaluates multimodal chatting that necessitates", "type": "text" } ], "index": 25 }, { "bbox": [ 104, 495, 505, 508 ], "spans": [ { "bbox": [ 104, 495, 505, 508 ], "score": 1.0, "content": "referring or grounding actions, e.g., instances where individuals reference an unfamiliar object and", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 506, 505, 519 ], "spans": [ { "bbox": [ 106, 506, 505, 519 ], "score": 1.0, "content": "inquire about its purpose. To benchmark this intriguing and practical capability, we introduce Ferret-", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 517, 505, 530 ], "spans": [ { "bbox": [ 105, 517, 505, 530 ], "score": 1.0, "content": "Bench that covers three kinds of region-based questions evaluating referring and grounding capa-", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 528, 506, 541 ], "spans": [ { "bbox": [ 106, 528, 506, 541 ], "score": 1.0, "content": "bility: (i) Referring Description: models are asked to describe a referred region based on its in-", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 539, 506, 552 ], "spans": [ { "bbox": [ 105, 539, 506, 552 ], "score": 1.0, "content": "teraction with surrounding objects. (ii) Referring Reasoning: models need to reason on top of", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 550, 506, 563 ], "spans": [ { "bbox": [ 105, 550, 506, 563 ], "score": 1.0, "content": "one or more referred regions correctly. (iii) Grounding in Conversation: models are required to", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 561, 506, 575 ], "spans": [ { "bbox": [ 105, 561, 506, 575 ], "score": 1.0, "content": "reason correctly and accurately ground/localize the objects/regions necessary for the reasoning. For", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 572, 506, 586 ], "spans": [ { "bbox": [ 105, 572, 506, 586 ], "score": 1.0, "content": "the ease of benchmarking other methods, we represent the regions with boxes instead of points or", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 582, 178, 596 ], "spans": [ { "bbox": [ 105, 582, 178, 596 ], "score": 1.0, "content": "free-form shapes.", "type": "text" } ], "index": 34 } ], "index": 28.5, "bbox_fs": [ 104, 461, 506, 596 ] }, { "type": "text", "bbox": [ 106, 600, 505, 732 ], "lines": [ { "bbox": [ 106, 600, 505, 613 ], "spans": [ { "bbox": [ 106, 600, 505, 613 ], "score": 1.0, "content": "Specifically, we randomly sample 40 images from the COCO validation set for each type of question,", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 611, 505, 623 ], "spans": [ { "bbox": [ 106, 611, 505, 623 ], "score": 1.0, "content": "and generate the questions and GPT-4’s answers following the instruction generation pipeline in", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 622, 505, 634 ], "spans": [ { "bbox": [ 106, 622, 505, 634 ], "score": 1.0, "content": "Sec. 3.2. Following Liu et al. (2023b), we feed the question and image into MLLMs to obtain", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 633, 505, 645 ], "spans": [ { "bbox": [ 106, 633, 505, 645 ], "score": 1.0, "content": "the predicted answer, and then prompt GPT-4 to rate the predicted answer and pseudo answer from", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 643, 505, 657 ], "spans": [ { "bbox": [ 105, 643, 505, 657 ], "score": 1.0, "content": "GPT-4 based on the ground-truth textual scene description (object, relationship, region caption,", "type": "text" } ], "index": 39 }, { "bbox": [ 104, 654, 505, 668 ], "spans": [ { "bbox": [ 104, 654, 505, 668 ], "score": 1.0, "content": "global caption). GPT-4 evaluates both the precision of referring understanding, object grounding,", "type": "text" } ], "index": 40 }, { "bbox": [ 104, 664, 505, 679 ], "spans": [ { "bbox": [ 104, 664, 505, 679 ], "score": 1.0, "content": "and correctness of semantics. The rating score ranges from 1 to 10, in which higher means better.", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "We calculate the ratio of the predicted answer’s score and the GPT-4 answer’s score, which is then", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 687, 506, 701 ], "spans": [ { "bbox": [ 105, 687, 506, 701 ], "score": 1.0, "content": "presented as a percentage to measure the performance of MLLMs. We also asked GPT-4 to give a", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "comprehensive review for the rating and found that GPT-4 is good at measuring the degree of spatial", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "precision, such as how much the predicted bounding box diverges from the GT box coordinate. We", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 721, 329, 733 ], "spans": [ { "bbox": [ 105, 721, 329, 733 ], "score": 1.0, "content": "refer the readers to Appendix D for further elaboration.", "type": "text" } ], "index": 46 } ], "index": 40.5, "bbox_fs": [ 104, 600, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 126, 95, 482, 181 ], "blocks": [ { "type": "table_caption", "bbox": [ 124, 81, 485, 92 ], "group_id": 0, "lines": [ { "bbox": [ 123, 79, 487, 93 ], "spans": [ { "bbox": [ 123, 79, 487, 93 ], "score": 1.0, "content": "Table 4: Results on LLaVA-Bench and the proposed Ferret-Bench via GPT4-as-a-Judge evaluation.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "table_body", "bbox": [ 126, 95, 482, 181 ], "group_id": 0, "lines": [ { "bbox": [ 126, 95, 482, 181 ], "spans": [ { "bbox": [ 126, 95, 482, 181 ], "score": 0.98, "html": "
LLaVA-BenchFerret-Bench
ConversatioDon xAvg.Deferig Reeig ginAvg
LLaVA885.468.392.181.941.431.728.834.0
Kosmos-271.763.474.970.051.833.748.444.6
Shikra-7B80.670.788.179.946.041.650.145.9
Ferret-7B84.479.496.386.768.767.357.564.5
Ferret-13B85.280.996.487.570.668.759.766.3
", "type": "table", "image_path": "9ee8d7e1478b5b9ad88ca6e7397f4fa08177f5929a687da7756108066422122f.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 126, 95, 482, 123.66666666666667 ], "spans": [], "index": 1 }, { "bbox": [ 126, 123.66666666666667, 482, 152.33333333333334 ], "spans": [], "index": 2 }, { "bbox": [ 126, 152.33333333333334, 482, 181.0 ], "spans": [], "index": 3 } ] } ], "index": 1.0 }, { "type": "table", "bbox": [ 108, 218, 289, 279 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 187, 290, 217 ], "group_id": 1, "lines": [ { "bbox": [ 106, 186, 292, 198 ], "spans": [ { "bbox": [ 106, 186, 292, 198 ], "score": 1.0, "content": "Table 5: Ablation study on the mutual benefit of", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 197, 291, 208 ], "spans": [ { "bbox": [ 105, 197, 291, 208 ], "score": 1.0, "content": "grounding data and referring data. We evaluate Ac-", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 205, 291, 219 ], "spans": [ { "bbox": [ 105, 205, 215, 219 ], "score": 1.0, "content": "curacy for LVIS referring and", "type": "text" }, { "bbox": [ 215, 207, 235, 216 ], "score": 0.83, "content": "\\mathbf { R } \\ @ 1", "type": "inline_equation" }, { "bbox": [ 236, 205, 291, 219 ], "score": 1.0, "content": "for grounding.", "type": "text" } ], "index": 6 } ], "index": 5 }, { "type": "table_body", "bbox": [ 108, 218, 289, 279 ], "group_id": 1, "lines": [ { "bbox": [ 108, 218, 289, 279 ], "spans": [ { "bbox": [ 108, 218, 289, 279 ], "score": 0.976, "html": "
ModelReferring (LVIS)| Grounding
PointBoxFlickr30k
Ferret67.979.480.4
w/o Grounding data65.475.6×
w/o Referring data××79.8
", "type": "table", "image_path": "eda53cc861de10d6df5488df971bb14386ff3a031ba148f77f1130cebc277f84.jpg" } ] } ], "index": 8.5, "virtual_lines": [ { "bbox": [ 108, 218, 289, 233.25 ], "spans": [], "index": 7 }, { "bbox": [ 108, 233.25, 289, 248.5 ], "spans": [], "index": 8 }, { "bbox": [ 108, 248.5, 289, 263.75 ], "spans": [], "index": 9 }, { "bbox": [ 108, 263.75, 289, 279.0 ], "spans": [], "index": 10 } ] } ], "index": 6.75 }, { "type": "table", "bbox": [ 305, 222, 502, 275 ], "blocks": [ { "type": "table_caption", "bbox": [ 304, 191, 504, 221 ], "group_id": 2, "lines": [ { "bbox": [ 304, 190, 505, 203 ], "spans": [ { "bbox": [ 304, 190, 505, 203 ], "score": 1.0, "content": "Table 6: Ablation study on the effectiveness of the pro-", "type": "text" } ], "index": 11 }, { "bbox": [ 303, 201, 505, 213 ], "spans": [ { "bbox": [ 303, 201, 505, 213 ], "score": 1.0, "content": "posed spatial-aware visual sampler. Accuracy is used to", "type": "text" } ], "index": 12 }, { "bbox": [ 304, 210, 394, 223 ], "spans": [ { "bbox": [ 304, 210, 394, 223 ], "score": 1.0, "content": "evaluate LVIS referring.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "table_body", "bbox": [ 305, 222, 502, 275 ], "group_id": 2, "lines": [ { "bbox": [ 305, 222, 502, 275 ], "spans": [ { "bbox": [ 305, 222, 502, 275 ], "score": 0.97, "html": "
ModuleReferring (LVIS)
Point Box Free-form
Spatial-aware Visual Sampler 67.979.469.8
Visual Sampler in SEEM67.177.268.9
", "type": "table", "image_path": "6324eaa714f03f6680c7df041b628550b308cd62adf7bf8869ea0141035bb9e3.jpg" } ] } ], "index": 15.5, "virtual_lines": [ { "bbox": [ 305, 222, 502, 235.25 ], "spans": [], "index": 14 }, { "bbox": [ 305, 235.25, 502, 248.5 ], "spans": [], "index": 15 }, { "bbox": [ 305, 248.5, 502, 261.75 ], "spans": [], "index": 16 }, { "bbox": [ 305, 261.75, 502, 275.0 ], "spans": [], "index": 17 } ] } ], "index": 13.75 }, { "type": "text", "bbox": [ 107, 285, 505, 340 ], "lines": [ { "bbox": [ 106, 286, 504, 298 ], "spans": [ { "bbox": [ 106, 286, 504, 298 ], "score": 1.0, "content": "We use LLaVA-Bench (Liu et al., 2023b) and the proposed Ferret-Bench to compare Ferret with", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 296, 505, 309 ], "spans": [ { "bbox": [ 105, 296, 505, 309 ], "score": 1.0, "content": "previous models, including LLaVA (Liu et al., 2023b), Shikra (Chen et al., 2023b), and Kosmos-", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 307, 504, 319 ], "spans": [ { "bbox": [ 106, 307, 504, 319 ], "score": 1.0, "content": "2 (Peng et al., 2023). Results are summarized in Table 4. Ferret achieves superior performance in", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 318, 506, 331 ], "spans": [ { "bbox": [ 105, 318, 506, 331 ], "score": 1.0, "content": "all types of tasks, boosting the score for the detailed description category from 68.3 to 80.9, and", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 329, 445, 342 ], "spans": [ { "bbox": [ 105, 329, 445, 342 ], "score": 1.0, "content": "especially excels at the three new tasks demanding referring and grounding abilities.", "type": "text" } ], "index": 22 } ], "index": 20 }, { "type": "title", "bbox": [ 107, 350, 177, 360 ], "lines": [ { "bbox": [ 106, 348, 178, 362 ], "spans": [ { "bbox": [ 106, 348, 178, 362 ], "score": 1.0, "content": "4.4 ABLATION", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 107, 365, 503, 387 ], "lines": [ { "bbox": [ 105, 363, 505, 378 ], "spans": [ { "bbox": [ 105, 363, 505, 378 ], "score": 1.0, "content": "In the ablation studies below, in default, we ablate Ferret-7B and mainly evaluate in referring object", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 376, 389, 388 ], "spans": [ { "bbox": [ 105, 376, 389, 388 ], "score": 1.0, "content": "classification and grounding tasks on Flickr30k Entities validation set.", "type": "text" } ], "index": 25 } ], "index": 24.5 }, { "type": "text", "bbox": [ 107, 393, 505, 426 ], "lines": [ { "bbox": [ 105, 391, 505, 406 ], "spans": [ { "bbox": [ 105, 391, 505, 406 ], "score": 1.0, "content": "Mutual benefits of grounding and referring. As shown in Table 5, grounding and referring, as", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 404, 505, 416 ], "spans": [ { "bbox": [ 105, 404, 505, 416 ], "score": 1.0, "content": "two main capabilities emphasized in this paper, can actually benefit each other. Particularly, when", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 414, 481, 428 ], "spans": [ { "bbox": [ 105, 414, 481, 428 ], "score": 1.0, "content": "adding grounding data into training, the referring performance gets improved, and vice versa.", "type": "text" } ], "index": 28 } ], "index": 27 }, { "type": "text", "bbox": [ 107, 432, 505, 476 ], "lines": [ { "bbox": [ 105, 431, 505, 444 ], "spans": [ { "bbox": [ 105, 431, 505, 444 ], "score": 1.0, "content": "Spatial-aware Visual Sampler. We ablate the effectiveness of the spatial-aware visual sampler", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 442, 506, 455 ], "spans": [ { "bbox": [ 105, 442, 506, 455 ], "score": 1.0, "content": "by replacing it with the visual sampler in SEEM (Zou et al., 2023), where they average the features", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 454, 505, 467 ], "spans": [ { "bbox": [ 105, 454, 505, 467 ], "score": 1.0, "content": "of all the sampled points as the region feature. As we can see in Table 6, ours can outperform the", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 465, 311, 477 ], "spans": [ { "bbox": [ 105, 465, 311, 477 ], "score": 1.0, "content": "previous visual sampler in all three referring tasks.", "type": "text" } ], "index": 32 } ], "index": 30.5 }, { "type": "text", "bbox": [ 107, 481, 503, 504 ], "lines": [ { "bbox": [ 105, 479, 505, 495 ], "spans": [ { "bbox": [ 105, 479, 505, 495 ], "score": 1.0, "content": "LLM model size. We study how much LLM model size influences the performance of referring", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 492, 454, 505 ], "spans": [ { "bbox": [ 106, 492, 454, 505 ], "score": 1.0, "content": "and grounding. As seen in Table 1-4, having a larger LM backbone can generally help.", "type": "text" } ], "index": 34 } ], "index": 33.5 }, { "type": "title", "bbox": [ 107, 513, 240, 524 ], "lines": [ { "bbox": [ 105, 512, 241, 525 ], "spans": [ { "bbox": [ 105, 512, 241, 525 ], "score": 1.0, "content": "4.5 OBJECT HALLUCINATION", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 107, 528, 505, 572 ], "lines": [ { "bbox": [ 105, 528, 506, 541 ], "spans": [ { "bbox": [ 105, 528, 506, 541 ], "score": 1.0, "content": "Attribute to the incorporation of fine-grained spatial knowledge and negative mining, Ferret also", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 540, 505, 551 ], "spans": [ { "bbox": [ 105, 540, 505, 551 ], "score": 1.0, "content": "exhibits strong power against the hallucination problem. We evaluate object hallucinations on the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 550, 505, 562 ], "spans": [ { "bbox": [ 105, 550, 505, 562 ], "score": 1.0, "content": "POPE benchmark (Li et al., 2023e). Results are summarized in Table 7. Ferret has exhibited perfor-", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 561, 479, 573 ], "spans": [ { "bbox": [ 105, 561, 479, 573 ], "score": 1.0, "content": "mance comparable to Shikra (Chen et al., 2023b), and far surpasses recent popular MLLMs.9", "type": "text" } ], "index": 39 } ], "index": 37.5 }, { "type": "title", "bbox": [ 108, 584, 195, 596 ], "lines": [ { "bbox": [ 104, 582, 197, 600 ], "spans": [ { "bbox": [ 104, 582, 197, 600 ], "score": 1.0, "content": "5 CONCLUSION", "type": "text" } ], "index": 40 } ], "index": 40 }, { "type": "text", "bbox": [ 107, 604, 505, 670 ], "lines": [ { "bbox": [ 106, 604, 505, 617 ], "spans": [ { "bbox": [ 106, 604, 505, 617 ], "score": 1.0, "content": "We present Ferret, a new multimodal large language model adept at referring and grounding. Fer-", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 615, 505, 627 ], "spans": [ { "bbox": [ 105, 615, 505, 627 ], "score": 1.0, "content": "ret can refer image regions in any free-form shape, and automatically establish grounding for text", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 627, 505, 638 ], "spans": [ { "bbox": [ 106, 627, 505, 638 ], "score": 1.0, "content": "deemed groundable by the model. We have curated the GRIT dataset for model training, and the", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 637, 505, 649 ], "spans": [ { "bbox": [ 105, 637, 505, 649 ], "score": 1.0, "content": "Ferret-Bench dataset for evaluation. Ferret, like most MLLMs, may produce harmful and counter-", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 649, 505, 660 ], "spans": [ { "bbox": [ 105, 649, 505, 660 ], "score": 1.0, "content": "factual responses. For future work, inspired by LISA (Lai et al., 2023), we plan to enhance Ferret to", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 659, 383, 672 ], "spans": [ { "bbox": [ 105, 659, 383, 672 ], "score": 1.0, "content": "be able to output segmentation masks in addition to bounding boxes.", "type": "text" } ], "index": 46 } ], "index": 43.5 } ], "page_idx": 8, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 751, 309, 759 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "9", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 106, 680, 505, 731 ], "lines": [ { "bbox": [ 118, 678, 505, 694 ], "spans": [ { "bbox": [ 118, 678, 505, 694 ], "score": 1.0, "content": "8The result on LLaVA-Bench is obtained by evaluating LLaVA released checkpoint. The slight discrepancy", "type": "text" } ] }, { "bbox": [ 106, 690, 505, 702 ], "spans": [ { "bbox": [ 106, 690, 505, 702 ], "score": 1.0, "content": "might be due to evolving GPT4 APIs. For Ferret-Bench, we employ the same conversation template as Ferret,", "type": "text" } ] }, { "bbox": [ 105, 700, 506, 713 ], "spans": [ { "bbox": [ 105, 700, 506, 713 ], "score": 1.0, "content": "providing LLaVA with a predefined input size, resizing all coordinates accordingly, and generating a response.", "type": "text" } ] }, { "bbox": [ 118, 708, 506, 725 ], "spans": [ { "bbox": [ 118, 708, 506, 725 ], "score": 1.0, "content": "9Unlike other methods, Ferret refrains from relying on VQA. This decision stems from our observation that", "type": "text" } ] }, { "bbox": [ 106, 721, 475, 734 ], "spans": [ { "bbox": [ 106, 721, 475, 734 ], "score": 1.0, "content": "VQA answers tend to be concise, and this brevity can restrict the conversational capabilities of LLMs.", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 126, 95, 482, 181 ], "blocks": [ { "type": "table_caption", "bbox": [ 124, 81, 485, 92 ], "group_id": 0, "lines": [ { "bbox": [ 123, 79, 487, 93 ], "spans": [ { "bbox": [ 123, 79, 487, 93 ], "score": 1.0, "content": "Table 4: Results on LLaVA-Bench and the proposed Ferret-Bench via GPT4-as-a-Judge evaluation.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "table_body", "bbox": [ 126, 95, 482, 181 ], "group_id": 0, "lines": [ { "bbox": [ 126, 95, 482, 181 ], "spans": [ { "bbox": [ 126, 95, 482, 181 ], "score": 0.98, "html": "
LLaVA-BenchFerret-Bench
ConversatioDon xAvg.Deferig Reeig ginAvg
LLaVA885.468.392.181.941.431.728.834.0
Kosmos-271.763.474.970.051.833.748.444.6
Shikra-7B80.670.788.179.946.041.650.145.9
Ferret-7B84.479.496.386.768.767.357.564.5
Ferret-13B85.280.996.487.570.668.759.766.3
", "type": "table", "image_path": "9ee8d7e1478b5b9ad88ca6e7397f4fa08177f5929a687da7756108066422122f.jpg" } ] } ], "index": 2, "virtual_lines": [ { "bbox": [ 126, 95, 482, 123.66666666666667 ], "spans": [], "index": 1 }, { "bbox": [ 126, 123.66666666666667, 482, 152.33333333333334 ], "spans": [], "index": 2 }, { "bbox": [ 126, 152.33333333333334, 482, 181.0 ], "spans": [], "index": 3 } ] } ], "index": 1.0 }, { "type": "table", "bbox": [ 108, 218, 289, 279 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 187, 290, 217 ], "group_id": 1, "lines": [ { "bbox": [ 106, 186, 292, 198 ], "spans": [ { "bbox": [ 106, 186, 292, 198 ], "score": 1.0, "content": "Table 5: Ablation study on the mutual benefit of", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 197, 291, 208 ], "spans": [ { "bbox": [ 105, 197, 291, 208 ], "score": 1.0, "content": "grounding data and referring data. We evaluate Ac-", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 205, 291, 219 ], "spans": [ { "bbox": [ 105, 205, 215, 219 ], "score": 1.0, "content": "curacy for LVIS referring and", "type": "text" }, { "bbox": [ 215, 207, 235, 216 ], "score": 0.83, "content": "\\mathbf { R } \\ @ 1", "type": "inline_equation" }, { "bbox": [ 236, 205, 291, 219 ], "score": 1.0, "content": "for grounding.", "type": "text" } ], "index": 6 } ], "index": 5 }, { "type": "table_body", "bbox": [ 108, 218, 289, 279 ], "group_id": 1, "lines": [ { "bbox": [ 108, 218, 289, 279 ], "spans": [ { "bbox": [ 108, 218, 289, 279 ], "score": 0.976, "html": "
ModelReferring (LVIS)| Grounding
PointBoxFlickr30k
Ferret67.979.480.4
w/o Grounding data65.475.6×
w/o Referring data××79.8
", "type": "table", "image_path": "eda53cc861de10d6df5488df971bb14386ff3a031ba148f77f1130cebc277f84.jpg" } ] } ], "index": 8.5, "virtual_lines": [ { "bbox": [ 108, 218, 289, 233.25 ], "spans": [], "index": 7 }, { "bbox": [ 108, 233.25, 289, 248.5 ], "spans": [], "index": 8 }, { "bbox": [ 108, 248.5, 289, 263.75 ], "spans": [], "index": 9 }, { "bbox": [ 108, 263.75, 289, 279.0 ], "spans": [], "index": 10 } ] } ], "index": 6.75 }, { "type": "table", "bbox": [ 305, 222, 502, 275 ], "blocks": [ { "type": "table_caption", "bbox": [ 304, 191, 504, 221 ], "group_id": 2, "lines": [ { "bbox": [ 304, 190, 505, 203 ], "spans": [ { "bbox": [ 304, 190, 505, 203 ], "score": 1.0, "content": "Table 6: Ablation study on the effectiveness of the pro-", "type": "text" } ], "index": 11 }, { "bbox": [ 303, 201, 505, 213 ], "spans": [ { "bbox": [ 303, 201, 505, 213 ], "score": 1.0, "content": "posed spatial-aware visual sampler. Accuracy is used to", "type": "text" } ], "index": 12 }, { "bbox": [ 304, 210, 394, 223 ], "spans": [ { "bbox": [ 304, 210, 394, 223 ], "score": 1.0, "content": "evaluate LVIS referring.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "table_body", "bbox": [ 305, 222, 502, 275 ], "group_id": 2, "lines": [ { "bbox": [ 305, 222, 502, 275 ], "spans": [ { "bbox": [ 305, 222, 502, 275 ], "score": 0.97, "html": "
ModuleReferring (LVIS)
Point Box Free-form
Spatial-aware Visual Sampler 67.979.469.8
Visual Sampler in SEEM67.177.268.9
", "type": "table", "image_path": "6324eaa714f03f6680c7df041b628550b308cd62adf7bf8869ea0141035bb9e3.jpg" } ] } ], "index": 15.5, "virtual_lines": [ { "bbox": [ 305, 222, 502, 235.25 ], "spans": [], "index": 14 }, { "bbox": [ 305, 235.25, 502, 248.5 ], "spans": [], "index": 15 }, { "bbox": [ 305, 248.5, 502, 261.75 ], "spans": [], "index": 16 }, { "bbox": [ 305, 261.75, 502, 275.0 ], "spans": [], "index": 17 } ] } ], "index": 13.75 }, { "type": "text", "bbox": [ 107, 285, 505, 340 ], "lines": [ { "bbox": [ 106, 286, 504, 298 ], "spans": [ { "bbox": [ 106, 286, 504, 298 ], "score": 1.0, "content": "We use LLaVA-Bench (Liu et al., 2023b) and the proposed Ferret-Bench to compare Ferret with", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 296, 505, 309 ], "spans": [ { "bbox": [ 105, 296, 505, 309 ], "score": 1.0, "content": "previous models, including LLaVA (Liu et al., 2023b), Shikra (Chen et al., 2023b), and Kosmos-", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 307, 504, 319 ], "spans": [ { "bbox": [ 106, 307, 504, 319 ], "score": 1.0, "content": "2 (Peng et al., 2023). Results are summarized in Table 4. Ferret achieves superior performance in", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 318, 506, 331 ], "spans": [ { "bbox": [ 105, 318, 506, 331 ], "score": 1.0, "content": "all types of tasks, boosting the score for the detailed description category from 68.3 to 80.9, and", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 329, 445, 342 ], "spans": [ { "bbox": [ 105, 329, 445, 342 ], "score": 1.0, "content": "especially excels at the three new tasks demanding referring and grounding abilities.", "type": "text" } ], "index": 22 } ], "index": 20, "bbox_fs": [ 105, 286, 506, 342 ] }, { "type": "title", "bbox": [ 107, 350, 177, 360 ], "lines": [ { "bbox": [ 106, 348, 178, 362 ], "spans": [ { "bbox": [ 106, 348, 178, 362 ], "score": 1.0, "content": "4.4 ABLATION", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 107, 365, 503, 387 ], "lines": [ { "bbox": [ 105, 363, 505, 378 ], "spans": [ { "bbox": [ 105, 363, 505, 378 ], "score": 1.0, "content": "In the ablation studies below, in default, we ablate Ferret-7B and mainly evaluate in referring object", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 376, 389, 388 ], "spans": [ { "bbox": [ 105, 376, 389, 388 ], "score": 1.0, "content": "classification and grounding tasks on Flickr30k Entities validation set.", "type": "text" } ], "index": 25 } ], "index": 24.5, "bbox_fs": [ 105, 363, 505, 388 ] }, { "type": "text", "bbox": [ 107, 393, 505, 426 ], "lines": [ { "bbox": [ 105, 391, 505, 406 ], "spans": [ { "bbox": [ 105, 391, 505, 406 ], "score": 1.0, "content": "Mutual benefits of grounding and referring. As shown in Table 5, grounding and referring, as", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 404, 505, 416 ], "spans": [ { "bbox": [ 105, 404, 505, 416 ], "score": 1.0, "content": "two main capabilities emphasized in this paper, can actually benefit each other. Particularly, when", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 414, 481, 428 ], "spans": [ { "bbox": [ 105, 414, 481, 428 ], "score": 1.0, "content": "adding grounding data into training, the referring performance gets improved, and vice versa.", "type": "text" } ], "index": 28 } ], "index": 27, "bbox_fs": [ 105, 391, 505, 428 ] }, { "type": "text", "bbox": [ 107, 432, 505, 476 ], "lines": [ { "bbox": [ 105, 431, 505, 444 ], "spans": [ { "bbox": [ 105, 431, 505, 444 ], "score": 1.0, "content": "Spatial-aware Visual Sampler. We ablate the effectiveness of the spatial-aware visual sampler", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 442, 506, 455 ], "spans": [ { "bbox": [ 105, 442, 506, 455 ], "score": 1.0, "content": "by replacing it with the visual sampler in SEEM (Zou et al., 2023), where they average the features", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 454, 505, 467 ], "spans": [ { "bbox": [ 105, 454, 505, 467 ], "score": 1.0, "content": "of all the sampled points as the region feature. As we can see in Table 6, ours can outperform the", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 465, 311, 477 ], "spans": [ { "bbox": [ 105, 465, 311, 477 ], "score": 1.0, "content": "previous visual sampler in all three referring tasks.", "type": "text" } ], "index": 32 } ], "index": 30.5, "bbox_fs": [ 105, 431, 506, 477 ] }, { "type": "text", "bbox": [ 107, 481, 503, 504 ], "lines": [ { "bbox": [ 105, 479, 505, 495 ], "spans": [ { "bbox": [ 105, 479, 505, 495 ], "score": 1.0, "content": "LLM model size. We study how much LLM model size influences the performance of referring", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 492, 454, 505 ], "spans": [ { "bbox": [ 106, 492, 454, 505 ], "score": 1.0, "content": "and grounding. As seen in Table 1-4, having a larger LM backbone can generally help.", "type": "text" } ], "index": 34 } ], "index": 33.5, "bbox_fs": [ 105, 479, 505, 505 ] }, { "type": "title", "bbox": [ 107, 513, 240, 524 ], "lines": [ { "bbox": [ 105, 512, 241, 525 ], "spans": [ { "bbox": [ 105, 512, 241, 525 ], "score": 1.0, "content": "4.5 OBJECT HALLUCINATION", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 107, 528, 505, 572 ], "lines": [ { "bbox": [ 105, 528, 506, 541 ], "spans": [ { "bbox": [ 105, 528, 506, 541 ], "score": 1.0, "content": "Attribute to the incorporation of fine-grained spatial knowledge and negative mining, Ferret also", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 540, 505, 551 ], "spans": [ { "bbox": [ 105, 540, 505, 551 ], "score": 1.0, "content": "exhibits strong power against the hallucination problem. We evaluate object hallucinations on the", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 550, 505, 562 ], "spans": [ { "bbox": [ 105, 550, 505, 562 ], "score": 1.0, "content": "POPE benchmark (Li et al., 2023e). Results are summarized in Table 7. Ferret has exhibited perfor-", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 561, 479, 573 ], "spans": [ { "bbox": [ 105, 561, 479, 573 ], "score": 1.0, "content": "mance comparable to Shikra (Chen et al., 2023b), and far surpasses recent popular MLLMs.9", "type": "text" } ], "index": 39 } ], "index": 37.5, "bbox_fs": [ 105, 528, 506, 573 ] }, { "type": "title", "bbox": [ 108, 584, 195, 596 ], "lines": [ { "bbox": [ 104, 582, 197, 600 ], "spans": [ { "bbox": [ 104, 582, 197, 600 ], "score": 1.0, "content": "5 CONCLUSION", "type": "text" } ], "index": 40 } ], "index": 40 }, { "type": "text", "bbox": [ 107, 604, 505, 670 ], "lines": [ { "bbox": [ 106, 604, 505, 617 ], "spans": [ { "bbox": [ 106, 604, 505, 617 ], "score": 1.0, "content": "We present Ferret, a new multimodal large language model adept at referring and grounding. Fer-", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 615, 505, 627 ], "spans": [ { "bbox": [ 105, 615, 505, 627 ], "score": 1.0, "content": "ret can refer image regions in any free-form shape, and automatically establish grounding for text", "type": "text" } ], "index": 42 }, { "bbox": [ 106, 627, 505, 638 ], "spans": [ { "bbox": [ 106, 627, 505, 638 ], "score": 1.0, "content": "deemed groundable by the model. We have curated the GRIT dataset for model training, and the", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 637, 505, 649 ], "spans": [ { "bbox": [ 105, 637, 505, 649 ], "score": 1.0, "content": "Ferret-Bench dataset for evaluation. Ferret, like most MLLMs, may produce harmful and counter-", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 649, 505, 660 ], "spans": [ { "bbox": [ 105, 649, 505, 660 ], "score": 1.0, "content": "factual responses. For future work, inspired by LISA (Lai et al., 2023), we plan to enhance Ferret to", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 659, 383, 672 ], "spans": [ { "bbox": [ 105, 659, 383, 672 ], "score": 1.0, "content": "be able to output segmentation masks in addition to bounding boxes.", "type": "text" } ], "index": 46 } ], "index": 43.5, "bbox_fs": [ 105, 604, 505, 672 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 81, 176, 93 ], "lines": [ { "bbox": [ 106, 82, 176, 94 ], "spans": [ { "bbox": [ 106, 82, 176, 94 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 100, 92, 507, 732 ], "lines": [ { "bbox": [ 105, 100, 506, 115 ], "spans": [ { "bbox": [ 105, 100, 506, 115 ], "score": 1.0, "content": "Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal,", "type": "text" } ], "index": 1 }, { "bbox": [ 115, 112, 505, 125 ], "spans": [ { "bbox": [ 115, 112, 505, 125 ], "score": 1.0, "content": "Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multi-", "type": "text" } ], "index": 2 }, { "bbox": [ 115, 122, 396, 138 ], "spans": [ { "bbox": [ 115, 122, 396, 138 ], "score": 1.0, "content": "modal model of the internet. arXiv preprint arXiv:2201.07520, 2022.", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 143, 505, 156 ], "spans": [ { "bbox": [ 105, 143, 505, 156 ], "score": 1.0, "content": "Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel", "type": "text" } ], "index": 4 }, { "bbox": [ 114, 152, 506, 169 ], "spans": [ { "bbox": [ 114, 152, 506, 169 ], "score": 1.0, "content": "Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language", "type": "text" } ], "index": 5 }, { "bbox": [ 114, 164, 396, 179 ], "spans": [ { "bbox": [ 114, 164, 396, 179 ], "score": 1.0, "content": "model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022.", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 185, 505, 198 ], "spans": [ { "bbox": [ 106, 185, 505, 198 ], "score": 1.0, "content": "Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani", "type": "text" } ], "index": 7 }, { "bbox": [ 114, 195, 506, 209 ], "spans": [ { "bbox": [ 114, 195, 506, 209 ], "score": 1.0, "content": "Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel", "type": "text" } ], "index": 8 }, { "bbox": [ 114, 205, 505, 222 ], "spans": [ { "bbox": [ 114, 205, 505, 222 ], "score": 1.0, "content": "Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo, March 2023. URL https:", "type": "text" } ], "index": 9 }, { "bbox": [ 114, 217, 314, 233 ], "spans": [ { "bbox": [ 114, 217, 314, 233 ], "score": 1.0, "content": "//doi.org/10.5281/zenodo.7733589.", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 237, 505, 253 ], "spans": [ { "bbox": [ 105, 237, 505, 253 ], "score": 1.0, "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal,", "type": "text" } ], "index": 11 }, { "bbox": [ 114, 248, 507, 264 ], "spans": [ { "bbox": [ 114, 248, 507, 264 ], "score": 1.0, "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are", "type": "text" } ], "index": 12 }, { "bbox": [ 115, 260, 488, 273 ], "spans": [ { "bbox": [ 115, 260, 488, 273 ], "score": 1.0, "content": "few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020.", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 280, 505, 293 ], "spans": [ { "bbox": [ 106, 280, 505, 293 ], "score": 1.0, "content": "Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-", "type": "text" } ], "index": 14 }, { "bbox": [ 114, 289, 506, 306 ], "spans": [ { "bbox": [ 114, 289, 506, 306 ], "score": 1.0, "content": "enhanced visual instruction tuning for multimodal large language models. arXiv preprint", "type": "text" } ], "index": 15 }, { "bbox": [ 115, 302, 225, 316 ], "spans": [ { "bbox": [ 115, 302, 225, 316 ], "score": 1.0, "content": "arXiv:2308.13437, 2023a.", "type": "text" } ], "index": 16 }, { "bbox": [ 104, 319, 507, 338 ], "spans": [ { "bbox": [ 104, 319, 507, 338 ], "score": 1.0, "content": "Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 334, 465, 347 ], "spans": [ { "bbox": [ 115, 334, 465, 347 ], "score": 1.0, "content": "multimodal llm’s referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023b.", "type": "text" } ], "index": 18 }, { "bbox": [ 104, 350, 507, 369 ], "spans": [ { "bbox": [ 104, 350, 507, 369 ], "score": 1.0, "content": "Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language", "type": "text" } ], "index": 19 }, { "bbox": [ 115, 365, 449, 378 ], "spans": [ { "bbox": [ 115, 365, 449, 378 ], "score": 1.0, "content": "modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021.", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 383, 505, 399 ], "spans": [ { "bbox": [ 105, 383, 505, 399 ], "score": 1.0, "content": "Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey Hinton. A unified", "type": "text" } ], "index": 21 }, { "bbox": [ 114, 395, 427, 410 ], "spans": [ { "bbox": [ 114, 395, 427, 410 ], "score": 1.0, "content": "sequence interface for vision tasks. arXiv preprint arXiv:2206.07669, 2022a.", "type": "text" } ], "index": 22 }, { "bbox": [ 104, 414, 505, 430 ], "spans": [ { "bbox": [ 104, 414, 505, 430 ], "score": 1.0, "content": "Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian", "type": "text" } ], "index": 23 }, { "bbox": [ 114, 425, 505, 441 ], "spans": [ { "bbox": [ 114, 425, 505, 441 ], "score": 1.0, "content": "Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 437, 381, 452 ], "spans": [ { "bbox": [ 115, 437, 381, 452 ], "score": 1.0, "content": "language-image model. arXiv preprint arXiv:2209.06794, 2022b.", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 457, 505, 471 ], "spans": [ { "bbox": [ 105, 457, 505, 471 ], "score": 1.0, "content": "Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Car-", "type": "text" } ], "index": 26 }, { "bbox": [ 113, 466, 508, 485 ], "spans": [ { "bbox": [ 113, 466, 508, 485 ], "score": 1.0, "content": "los Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a", "type": "text" } ], "index": 27 }, { "bbox": [ 114, 478, 447, 495 ], "spans": [ { "bbox": [ 114, 478, 447, 495 ], "score": 1.0, "content": "multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023c.", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 499, 506, 514 ], "spans": [ { "bbox": [ 105, 499, 506, 514 ], "score": 1.0, "content": "Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and", "type": "text" } ], "index": 29 }, { "bbox": [ 113, 510, 452, 525 ], "spans": [ { "bbox": [ 113, 510, 452, 525 ], "score": 1.0, "content": "Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 528, 507, 547 ], "spans": [ { "bbox": [ 105, 528, 507, 547 ], "score": 1.0, "content": "Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng,", "type": "text" } ], "index": 31 }, { "bbox": [ 114, 541, 506, 555 ], "spans": [ { "bbox": [ 114, 541, 506, 555 ], "score": 1.0, "content": "Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot", "type": "text" } ], "index": 32 }, { "bbox": [ 113, 551, 507, 568 ], "spans": [ { "bbox": [ 113, 551, 208, 568 ], "score": 1.0, "content": "impressing gpt-4 with", "type": "text" }, { "bbox": [ 209, 552, 234, 563 ], "score": 0.83, "content": "9 0 \\% *", "type": "inline_equation" }, { "bbox": [ 234, 551, 507, 568 ], "score": 1.0, "content": "chatgpt quality. See https://vicuna. lmsys. org (accessed 14 April", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 564, 171, 577 ], "spans": [ { "bbox": [ 115, 564, 171, 577 ], "score": 1.0, "content": "2023), 2023.", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 583, 506, 598 ], "spans": [ { "bbox": [ 105, 583, 506, 598 ], "score": 1.0, "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam", "type": "text" } ], "index": 35 }, { "bbox": [ 115, 594, 506, 609 ], "spans": [ { "bbox": [ 115, 594, 506, 609 ], "score": 1.0, "content": "Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm:", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 605, 453, 621 ], "spans": [ { "bbox": [ 115, 605, 453, 621 ], "score": 1.0, "content": "Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022.", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 623, 505, 640 ], "spans": [ { "bbox": [ 105, 623, 505, 640 ], "score": 1.0, "content": "Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang,", "type": "text" } ], "index": 38 }, { "bbox": [ 113, 635, 507, 653 ], "spans": [ { "bbox": [ 113, 635, 507, 653 ], "score": 1.0, "content": "Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general-purpose vision-language", "type": "text" } ], "index": 39 }, { "bbox": [ 115, 647, 407, 662 ], "spans": [ { "bbox": [ 115, 647, 407, 662 ], "score": 1.0, "content": "models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023.", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 668, 504, 681 ], "spans": [ { "bbox": [ 106, 668, 504, 681 ], "score": 1.0, "content": "Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-", "type": "text" } ], "index": 41 }, { "bbox": [ 115, 679, 343, 693 ], "spans": [ { "bbox": [ 115, 679, 343, 693 ], "score": 1.0, "content": "end visual grounding with transformers. In ICCV, 2021.", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 698, 506, 713 ], "spans": [ { "bbox": [ 105, 698, 506, 713 ], "score": 1.0, "content": "Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter,", "type": "text" } ], "index": 43 }, { "bbox": [ 116, 710, 505, 723 ], "spans": [ { "bbox": [ 116, 710, 505, 723 ], "score": 1.0, "content": "Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. PaLM-E: An embodied multi-", "type": "text" } ], "index": 44 }, { "bbox": [ 114, 720, 377, 735 ], "spans": [ { "bbox": [ 114, 720, 377, 735 ], "score": 1.0, "content": "modal language model. arXiv preprint arXiv:2303.03378, 2023.", "type": "text" } ], "index": 45 } ], "index": 23 } ], "page_idx": 9, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "10", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 81, 176, 93 ], "lines": [ { "bbox": [ 106, 82, 176, 94 ], "spans": [ { "bbox": [ 106, 82, 176, 94 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "list", "bbox": [ 100, 92, 507, 732 ], "lines": [ { "bbox": [ 105, 100, 506, 115 ], "spans": [ { "bbox": [ 105, 100, 506, 115 ], "score": 1.0, "content": "Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal,", "type": "text" } ], "index": 1, "is_list_start_line": true }, { "bbox": [ 115, 112, 505, 125 ], "spans": [ { "bbox": [ 115, 112, 505, 125 ], "score": 1.0, "content": "Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multi-", "type": "text" } ], "index": 2 }, { "bbox": [ 115, 122, 396, 138 ], "spans": [ { "bbox": [ 115, 122, 396, 138 ], "score": 1.0, "content": "modal model of the internet. arXiv preprint arXiv:2201.07520, 2022.", "type": "text" } ], "index": 3, "is_list_end_line": true }, { "bbox": [ 105, 143, 505, 156 ], "spans": [ { "bbox": [ 105, 143, 505, 156 ], "score": 1.0, "content": "Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel", "type": "text" } ], "index": 4, "is_list_start_line": true }, { "bbox": [ 114, 152, 506, 169 ], "spans": [ { "bbox": [ 114, 152, 506, 169 ], "score": 1.0, "content": "Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language", "type": "text" } ], "index": 5 }, { "bbox": [ 114, 164, 396, 179 ], "spans": [ { "bbox": [ 114, 164, 396, 179 ], "score": 1.0, "content": "model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022.", "type": "text" } ], "index": 6, "is_list_end_line": true }, { "bbox": [ 106, 185, 505, 198 ], "spans": [ { "bbox": [ 106, 185, 505, 198 ], "score": 1.0, "content": "Anas Awadalla, Irena Gao, Joshua Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani", "type": "text" } ], "index": 7, "is_list_start_line": true }, { "bbox": [ 114, 195, 506, 209 ], "spans": [ { "bbox": [ 114, 195, 506, 209 ], "score": 1.0, "content": "Marathe, Yonatan Bitton, Samir Gadre, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel", "type": "text" } ], "index": 8 }, { "bbox": [ 114, 205, 505, 222 ], "spans": [ { "bbox": [ 114, 205, 505, 222 ], "score": 1.0, "content": "Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo, March 2023. URL https:", "type": "text" } ], "index": 9 }, { "bbox": [ 114, 217, 314, 233 ], "spans": [ { "bbox": [ 114, 217, 314, 233 ], "score": 1.0, "content": "//doi.org/10.5281/zenodo.7733589.", "type": "text" } ], "index": 10, "is_list_end_line": true }, { "bbox": [ 105, 237, 505, 253 ], "spans": [ { "bbox": [ 105, 237, 505, 253 ], "score": 1.0, "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal,", "type": "text" } ], "index": 11, "is_list_start_line": true }, { "bbox": [ 114, 248, 507, 264 ], "spans": [ { "bbox": [ 114, 248, 507, 264 ], "score": 1.0, "content": "Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are", "type": "text" } ], "index": 12 }, { "bbox": [ 115, 260, 488, 273 ], "spans": [ { "bbox": [ 115, 260, 488, 273 ], "score": 1.0, "content": "few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020.", "type": "text" } ], "index": 13, "is_list_end_line": true }, { "bbox": [ 106, 280, 505, 293 ], "spans": [ { "bbox": [ 106, 280, 505, 293 ], "score": 1.0, "content": "Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-", "type": "text" } ], "index": 14, "is_list_start_line": true }, { "bbox": [ 114, 289, 506, 306 ], "spans": [ { "bbox": [ 114, 289, 506, 306 ], "score": 1.0, "content": "enhanced visual instruction tuning for multimodal large language models. arXiv preprint", "type": "text" } ], "index": 15 }, { "bbox": [ 115, 302, 225, 316 ], "spans": [ { "bbox": [ 115, 302, 225, 316 ], "score": 1.0, "content": "arXiv:2308.13437, 2023a.", "type": "text" } ], "index": 16, "is_list_end_line": true }, { "bbox": [ 104, 319, 507, 338 ], "spans": [ { "bbox": [ 104, 319, 507, 338 ], "score": 1.0, "content": "Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing", "type": "text" } ], "index": 17, "is_list_start_line": true }, { "bbox": [ 115, 334, 465, 347 ], "spans": [ { "bbox": [ 115, 334, 465, 347 ], "score": 1.0, "content": "multimodal llm’s referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023b.", "type": "text" } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 104, 350, 507, 369 ], "spans": [ { "bbox": [ 104, 350, 507, 369 ], "score": 1.0, "content": "Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language", "type": "text" } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 115, 365, 449, 378 ], "spans": [ { "bbox": [ 115, 365, 449, 378 ], "score": 1.0, "content": "modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021.", "type": "text" } ], "index": 20, "is_list_end_line": true }, { "bbox": [ 105, 383, 505, 399 ], "spans": [ { "bbox": [ 105, 383, 505, 399 ], "score": 1.0, "content": "Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey Hinton. A unified", "type": "text" } ], "index": 21, "is_list_start_line": true }, { "bbox": [ 114, 395, 427, 410 ], "spans": [ { "bbox": [ 114, 395, 427, 410 ], "score": 1.0, "content": "sequence interface for vision tasks. arXiv preprint arXiv:2206.07669, 2022a.", "type": "text" } ], "index": 22, "is_list_end_line": true }, { "bbox": [ 104, 414, 505, 430 ], "spans": [ { "bbox": [ 104, 414, 505, 430 ], "score": 1.0, "content": "Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian", "type": "text" } ], "index": 23, "is_list_start_line": true }, { "bbox": [ 114, 425, 505, 441 ], "spans": [ { "bbox": [ 114, 425, 505, 441 ], "score": 1.0, "content": "Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 437, 381, 452 ], "spans": [ { "bbox": [ 115, 437, 381, 452 ], "score": 1.0, "content": "language-image model. arXiv preprint arXiv:2209.06794, 2022b.", "type": "text" } ], "index": 25, "is_list_end_line": true }, { "bbox": [ 105, 457, 505, 471 ], "spans": [ { "bbox": [ 105, 457, 505, 471 ], "score": 1.0, "content": "Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Car-", "type": "text" } ], "index": 26, "is_list_start_line": true }, { "bbox": [ 113, 466, 508, 485 ], "spans": [ { "bbox": [ 113, 466, 508, 485 ], "score": 1.0, "content": "los Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a", "type": "text" } ], "index": 27 }, { "bbox": [ 114, 478, 447, 495 ], "spans": [ { "bbox": [ 114, 478, 447, 495 ], "score": 1.0, "content": "multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023c.", "type": "text" } ], "index": 28, "is_list_end_line": true }, { "bbox": [ 105, 499, 506, 514 ], "spans": [ { "bbox": [ 105, 499, 506, 514 ], "score": 1.0, "content": "Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and", "type": "text" } ], "index": 29, "is_list_start_line": true }, { "bbox": [ 113, 510, 452, 525 ], "spans": [ { "bbox": [ 113, 510, 452, 525 ], "score": 1.0, "content": "Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020.", "type": "text" } ], "index": 30, "is_list_end_line": true }, { "bbox": [ 105, 528, 507, 547 ], "spans": [ { "bbox": [ 105, 528, 507, 547 ], "score": 1.0, "content": "Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng,", "type": "text" } ], "index": 31, "is_list_start_line": true }, { "bbox": [ 114, 541, 506, 555 ], "spans": [ { "bbox": [ 114, 541, 506, 555 ], "score": 1.0, "content": "Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot", "type": "text" } ], "index": 32 }, { "bbox": [ 113, 551, 507, 568 ], "spans": [ { "bbox": [ 113, 551, 208, 568 ], "score": 1.0, "content": "impressing gpt-4 with", "type": "text" }, { "bbox": [ 209, 552, 234, 563 ], "score": 0.83, "content": "9 0 \\% *", "type": "inline_equation" }, { "bbox": [ 234, 551, 507, 568 ], "score": 1.0, "content": "chatgpt quality. See https://vicuna. lmsys. org (accessed 14 April", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 564, 171, 577 ], "spans": [ { "bbox": [ 115, 564, 171, 577 ], "score": 1.0, "content": "2023), 2023.", "type": "text" } ], "index": 34, "is_list_end_line": true }, { "bbox": [ 105, 583, 506, 598 ], "spans": [ { "bbox": [ 105, 583, 506, 598 ], "score": 1.0, "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam", "type": "text" } ], "index": 35, "is_list_start_line": true }, { "bbox": [ 115, 594, 506, 609 ], "spans": [ { "bbox": [ 115, 594, 506, 609 ], "score": 1.0, "content": "Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm:", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 605, 453, 621 ], "spans": [ { "bbox": [ 115, 605, 453, 621 ], "score": 1.0, "content": "Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022.", "type": "text" } ], "index": 37, "is_list_end_line": true }, { "bbox": [ 105, 623, 505, 640 ], "spans": [ { "bbox": [ 105, 623, 505, 640 ], "score": 1.0, "content": "Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang,", "type": "text" } ], "index": 38, "is_list_start_line": true }, { "bbox": [ 113, 635, 507, 653 ], "spans": [ { "bbox": [ 113, 635, 507, 653 ], "score": 1.0, "content": "Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general-purpose vision-language", "type": "text" } ], "index": 39 }, { "bbox": [ 115, 647, 407, 662 ], "spans": [ { "bbox": [ 115, 647, 407, 662 ], "score": 1.0, "content": "models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023.", "type": "text" } ], "index": 40, "is_list_end_line": true }, { "bbox": [ 106, 668, 504, 681 ], "spans": [ { "bbox": [ 106, 668, 504, 681 ], "score": 1.0, "content": "Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-", "type": "text" } ], "index": 41, "is_list_start_line": true }, { "bbox": [ 115, 679, 343, 693 ], "spans": [ { "bbox": [ 115, 679, 343, 693 ], "score": 1.0, "content": "end visual grounding with transformers. In ICCV, 2021.", "type": "text" } ], "index": 42, "is_list_end_line": true }, { "bbox": [ 105, 698, 506, 713 ], "spans": [ { "bbox": [ 105, 698, 506, 713 ], "score": 1.0, "content": "Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter,", "type": "text" } ], "index": 43, "is_list_start_line": true }, { "bbox": [ 116, 710, 505, 723 ], "spans": [ { "bbox": [ 116, 710, 505, 723 ], "score": 1.0, "content": "Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. PaLM-E: An embodied multi-", "type": "text" } ], "index": 44 }, { "bbox": [ 114, 720, 377, 735 ], "spans": [ { "bbox": [ 114, 720, 377, 735 ], "score": 1.0, "content": "modal language model. arXiv preprint arXiv:2303.03378, 2023.", "type": "text" } ], "index": 45, "is_list_end_line": true }, { "bbox": [ 105, 82, 506, 96 ], "spans": [ { "bbox": [ 105, 82, 506, 96 ], "score": 1.0, "content": "Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial", "type": "text", "cross_page": true } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 115, 95, 410, 106 ], "spans": [ { "bbox": [ 115, 95, 410, 106 ], "score": 1.0, "content": "training for vision-and-language representation learning. NeurIPS, 2020.", "type": "text", "cross_page": true } ], "index": 1, "is_list_end_line": true }, { "bbox": [ 105, 112, 505, 127 ], "spans": [ { "bbox": [ 105, 112, 505, 127 ], "score": 1.0, "content": "Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmen-", "type": "text", "cross_page": true } ], "index": 2, "is_list_start_line": true }, { "bbox": [ 114, 124, 506, 138 ], "spans": [ { "bbox": [ 114, 124, 506, 138 ], "score": 1.0, "content": "tation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition,", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 114, 135, 208, 147 ], "spans": [ { "bbox": [ 114, 135, 208, 147 ], "score": 1.0, "content": "pp. 5356–5364, 2019.", "type": "text", "cross_page": true } ], "index": 4, "is_list_end_line": true }, { "bbox": [ 105, 153, 505, 168 ], "spans": [ { "bbox": [ 105, 153, 505, 168 ], "score": 1.0, "content": "Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In ´ Proceedings of the", "type": "text", "cross_page": true } ], "index": 5, "is_list_start_line": true }, { "bbox": [ 114, 165, 412, 179 ], "spans": [ { "bbox": [ 114, 165, 412, 179 ], "score": 1.0, "content": "IEEE international conference on computer vision, pp. 2961–2969, 2017.", "type": "text", "cross_page": true } ], "index": 6, "is_list_end_line": true }, { "bbox": [ 106, 185, 504, 197 ], "spans": [ { "bbox": [ 106, 185, 504, 197 ], "score": 1.0, "content": "Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Car-", "type": "text", "cross_page": true } ], "index": 7, "is_list_start_line": true }, { "bbox": [ 114, 195, 506, 210 ], "spans": [ { "bbox": [ 114, 195, 506, 210 ], "score": 1.0, "content": "ion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 114, 206, 440, 219 ], "spans": [ { "bbox": [ 114, 206, 440, 219 ], "score": 1.0, "content": "IEEE/CVF International Conference on Computer Vision, pp. 1780–1790, 2021.", "type": "text", "cross_page": true } ], "index": 9, "is_list_end_line": true }, { "bbox": [ 104, 224, 506, 239 ], "spans": [ { "bbox": [ 104, 224, 506, 239 ], "score": 1.0, "content": "Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to", "type": "text", "cross_page": true } ], "index": 10, "is_list_start_line": true }, { "bbox": [ 116, 237, 505, 249 ], "spans": [ { "bbox": [ 116, 237, 505, 249 ], "score": 1.0, "content": "objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 114, 247, 403, 261 ], "spans": [ { "bbox": [ 114, 247, 403, 261 ], "score": 1.0, "content": "methods in natural language processing (EMNLP), pp. 787–798, 2014.", "type": "text", "cross_page": true } ], "index": 12, "is_list_end_line": true }, { "bbox": [ 106, 267, 505, 279 ], "spans": [ { "bbox": [ 106, 267, 505, 279 ], "score": 1.0, "content": "Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete", "type": "text", "cross_page": true } ], "index": 13, "is_list_start_line": true }, { "bbox": [ 115, 277, 506, 292 ], "spans": [ { "bbox": [ 115, 277, 506, 292 ], "score": 1.0, "content": "Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv", "type": "text", "cross_page": true } ], "index": 14 }, { "bbox": [ 114, 289, 254, 302 ], "spans": [ { "bbox": [ 114, 289, 254, 302 ], "score": 1.0, "content": "preprint arXiv:2304.02643, 2023.", "type": "text", "cross_page": true } ], "index": 15, "is_list_end_line": true }, { "bbox": [ 104, 307, 506, 322 ], "spans": [ { "bbox": [ 104, 307, 506, 322 ], "score": 1.0, "content": "Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language", "type": "text", "cross_page": true } ], "index": 16, "is_list_start_line": true }, { "bbox": [ 114, 318, 317, 333 ], "spans": [ { "bbox": [ 114, 318, 317, 333 ], "score": 1.0, "content": "models. arXiv preprint arXiv:2305.17216, 2023a.", "type": "text", "cross_page": true } ], "index": 17, "is_list_end_line": true }, { "bbox": [ 105, 338, 506, 352 ], "spans": [ { "bbox": [ 105, 338, 506, 352 ], "score": 1.0, "content": "Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for", "type": "text", "cross_page": true } ], "index": 18, "is_list_start_line": true }, { "bbox": [ 114, 349, 380, 363 ], "spans": [ { "bbox": [ 114, 349, 380, 363 ], "score": 1.0, "content": "multimodal generation. arXiv preprint arXiv:2301.13823, 2023b.", "type": "text", "cross_page": true } ], "index": 19, "is_list_end_line": true }, { "bbox": [ 104, 367, 506, 384 ], "spans": [ { "bbox": [ 104, 367, 506, 384 ], "score": 1.0, "content": "Emiel Krahmer and Kees Van Deemter. Computational generation of referring expressions: A", "type": "text", "cross_page": true } ], "index": 20, "is_list_start_line": true }, { "bbox": [ 114, 379, 348, 393 ], "spans": [ { "bbox": [ 114, 379, 348, 393 ], "score": 1.0, "content": "survey. Computational Linguistics, 38(1):173–218, 2012.", "type": "text", "cross_page": true } ], "index": 21, "is_list_end_line": true }, { "bbox": [ 106, 399, 505, 411 ], "spans": [ { "bbox": [ 106, 399, 505, 411 ], "score": 1.0, "content": "Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie", "type": "text", "cross_page": true } ], "index": 22, "is_list_start_line": true }, { "bbox": [ 114, 408, 506, 425 ], "spans": [ { "bbox": [ 114, 408, 506, 425 ], "score": 1.0, "content": "Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting lan-", "type": "text", "cross_page": true } ], "index": 23 }, { "bbox": [ 114, 421, 505, 435 ], "spans": [ { "bbox": [ 114, 421, 505, 435 ], "score": 1.0, "content": "guage and vision using crowdsourced dense image annotations. International journal of computer", "type": "text", "cross_page": true } ], "index": 24 }, { "bbox": [ 116, 432, 219, 444 ], "spans": [ { "bbox": [ 116, 432, 219, 444 ], "score": 1.0, "content": "vision, 123:32–73, 2017.", "type": "text", "cross_page": true } ], "index": 25, "is_list_end_line": true }, { "bbox": [ 106, 452, 504, 464 ], "spans": [ { "bbox": [ 106, 452, 504, 464 ], "score": 1.0, "content": "Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Rea-", "type": "text", "cross_page": true } ], "index": 26, "is_list_start_line": true }, { "bbox": [ 115, 462, 469, 477 ], "spans": [ { "bbox": [ 115, 462, 469, 477 ], "score": 1.0, "content": "soning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023.", "type": "text", "cross_page": true } ], "index": 27, "is_list_end_line": true }, { "bbox": [ 105, 481, 506, 495 ], "spans": [ { "bbox": [ 105, 481, 506, 495 ], "score": 1.0, "content": "Hugo Laurenc¸on, Lucile Saulnier, Leo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, ´", "type": "text", "cross_page": true } ], "index": 28, "is_list_start_line": true }, { "bbox": [ 115, 492, 506, 506 ], "spans": [ { "bbox": [ 115, 492, 506, 506 ], "score": 1.0, "content": "Thomas Wang, Siddharth Karamcheti, Alexander M Rush, Douwe Kiela, et al. Obelisc: An open", "type": "text", "cross_page": true } ], "index": 29 }, { "bbox": [ 115, 502, 505, 517 ], "spans": [ { "bbox": [ 115, 502, 505, 517 ], "score": 1.0, "content": "web-scale filtered dataset of interleaved image-text documents. arXiv preprint arXiv:2306.16527,", "type": "text", "cross_page": true } ], "index": 30 }, { "bbox": [ 114, 514, 144, 529 ], "spans": [ { "bbox": [ 114, 514, 144, 529 ], "score": 1.0, "content": "2023.", "type": "text", "cross_page": true } ], "index": 31, "is_list_end_line": true }, { "bbox": [ 106, 534, 505, 546 ], "spans": [ { "bbox": [ 106, 534, 505, 546 ], "score": 1.0, "content": "Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A", "type": "text", "cross_page": true } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 115, 544, 501, 558 ], "spans": [ { "bbox": [ 115, 544, 501, 558 ], "score": 1.0, "content": "multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023a.", "type": "text", "cross_page": true } ], "index": 33 }, { "bbox": [ 105, 562, 505, 578 ], "spans": [ { "bbox": [ 105, 562, 505, 578 ], "score": 1.0, "content": "Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao.", "type": "text", "cross_page": true } ], "index": 34, "is_list_start_line": true }, { "bbox": [ 114, 573, 506, 590 ], "spans": [ { "bbox": [ 114, 573, 506, 590 ], "score": 1.0, "content": "Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint", "type": "text", "cross_page": true } ], "index": 35 }, { "bbox": [ 115, 586, 225, 598 ], "spans": [ { "bbox": [ 115, 586, 225, 598 ], "score": 1.0, "content": "arXiv:2309.10020, 2023b.", "type": "text", "cross_page": true } ], "index": 36, "is_list_end_line": true }, { "bbox": [ 104, 604, 505, 619 ], "spans": [ { "bbox": [ 104, 604, 505, 619 ], "score": 1.0, "content": "Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-", "type": "text", "cross_page": true } ], "index": 37, "is_list_start_line": true }, { "bbox": [ 114, 616, 506, 630 ], "spans": [ { "bbox": [ 114, 616, 506, 630 ], "score": 1.0, "content": "image pre-training with frozen image encoders and large language models. arXiv preprint", "type": "text", "cross_page": true } ], "index": 38 }, { "bbox": [ 114, 627, 225, 640 ], "spans": [ { "bbox": [ 114, 627, 225, 640 ], "score": 1.0, "content": "arXiv:2301.12597, 2023c.", "type": "text", "cross_page": true } ], "index": 39, "is_list_end_line": true }, { "bbox": [ 104, 645, 505, 660 ], "spans": [ { "bbox": [ 104, 645, 505, 660 ], "score": 1.0, "content": "Lei Li, Yuwei Yin, Shicheng Li, Liang Chen, Peiyi Wang, Shuhuai Ren, Mukai Li, Yazheng Yang,", "type": "text", "cross_page": true } ], "index": 40, "is_list_start_line": true }, { "bbox": [ 113, 656, 505, 672 ], "spans": [ { "bbox": [ 113, 656, 225, 672 ], "score": 1.0, "content": "Jingjing Xu, Xu Sun, et al.", "type": "text", "cross_page": true }, { "bbox": [ 225, 657, 246, 668 ], "score": 0.27, "content": "\\mathbf { M } ^ { 3 } \\mathrm { i t }", "type": "inline_equation", "cross_page": true }, { "bbox": [ 247, 656, 505, 672 ], "score": 1.0, "content": ": A large-scale dataset towards multi-modal multilingual instruc-", "type": "text", "cross_page": true } ], "index": 41 }, { "bbox": [ 116, 669, 333, 681 ], "spans": [ { "bbox": [ 116, 669, 333, 681 ], "score": 1.0, "content": "tion tuning. arXiv preprint arXiv:2306.04387, 2023d.", "type": "text", "cross_page": true } ], "index": 42, "is_list_end_line": true }, { "bbox": [ 105, 687, 505, 702 ], "spans": [ { "bbox": [ 105, 687, 505, 702 ], "score": 1.0, "content": "Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Li-", "type": "text", "cross_page": true } ], "index": 43, "is_list_start_line": true }, { "bbox": [ 114, 698, 505, 713 ], "spans": [ { "bbox": [ 114, 698, 505, 713 ], "score": 1.0, "content": "juan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training.", "type": "text", "cross_page": true } ], "index": 44 }, { "bbox": [ 114, 708, 505, 725 ], "spans": [ { "bbox": [ 114, 708, 505, 725 ], "score": 1.0, "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.", "type": "text", "cross_page": true } ], "index": 45 }, { "bbox": [ 117, 721, 201, 733 ], "spans": [ { "bbox": [ 117, 721, 201, 733 ], "score": 1.0, "content": "10965–10975, 2022.", "type": "text", "cross_page": true } ], "index": 46, "is_list_end_line": true } ], "index": 23, "bbox_fs": [ 104, 100, 508, 735 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 99, 65, 507, 746 ], "lines": [ { "bbox": [ 105, 82, 506, 96 ], "spans": [ { "bbox": [ 105, 82, 506, 96 ], "score": 1.0, "content": "Zhe Gan, Yen-Chun Chen, Linjie Li, Chen Zhu, Yu Cheng, and Jingjing Liu. Large-scale adversarial", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 95, 410, 106 ], "spans": [ { "bbox": [ 115, 95, 410, 106 ], "score": 1.0, "content": "training for vision-and-language representation learning. NeurIPS, 2020.", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 112, 505, 127 ], "spans": [ { "bbox": [ 105, 112, 505, 127 ], "score": 1.0, "content": "Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmen-", "type": "text" } ], "index": 2 }, { "bbox": [ 114, 124, 506, 138 ], "spans": [ { "bbox": [ 114, 124, 506, 138 ], "score": 1.0, "content": "tation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition,", "type": "text" } ], "index": 3 }, { "bbox": [ 114, 135, 208, 147 ], "spans": [ { "bbox": [ 114, 135, 208, 147 ], "score": 1.0, "content": "pp. 5356–5364, 2019.", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 153, 505, 168 ], "spans": [ { "bbox": [ 105, 153, 505, 168 ], "score": 1.0, "content": "Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In ´ Proceedings of the", "type": "text" } ], "index": 5 }, { "bbox": [ 114, 165, 412, 179 ], "spans": [ { "bbox": [ 114, 165, 412, 179 ], "score": 1.0, "content": "IEEE international conference on computer vision, pp. 2961–2969, 2017.", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 185, 504, 197 ], "spans": [ { "bbox": [ 106, 185, 504, 197 ], "score": 1.0, "content": "Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Car-", "type": "text" } ], "index": 7 }, { "bbox": [ 114, 195, 506, 210 ], "spans": [ { "bbox": [ 114, 195, 506, 210 ], "score": 1.0, "content": "ion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the", "type": "text" } ], "index": 8 }, { "bbox": [ 114, 206, 440, 219 ], "spans": [ { "bbox": [ 114, 206, 440, 219 ], "score": 1.0, "content": "IEEE/CVF International Conference on Computer Vision, pp. 1780–1790, 2021.", "type": "text" } ], "index": 9 }, { "bbox": [ 104, 224, 506, 239 ], "spans": [ { "bbox": [ 104, 224, 506, 239 ], "score": 1.0, "content": "Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 237, 505, 249 ], "spans": [ { "bbox": [ 116, 237, 505, 249 ], "score": 1.0, "content": "objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical", "type": "text" } ], "index": 11 }, { "bbox": [ 114, 247, 403, 261 ], "spans": [ { "bbox": [ 114, 247, 403, 261 ], "score": 1.0, "content": "methods in natural language processing (EMNLP), pp. 787–798, 2014.", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 267, 505, 279 ], "spans": [ { "bbox": [ 106, 267, 505, 279 ], "score": 1.0, "content": "Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 277, 506, 292 ], "spans": [ { "bbox": [ 115, 277, 506, 292 ], "score": 1.0, "content": "Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv", "type": "text" } ], "index": 14 }, { "bbox": [ 114, 289, 254, 302 ], "spans": [ { "bbox": [ 114, 289, 254, 302 ], "score": 1.0, "content": "preprint arXiv:2304.02643, 2023.", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 307, 506, 322 ], "spans": [ { "bbox": [ 104, 307, 506, 322 ], "score": 1.0, "content": "Jing Yu Koh, Daniel Fried, and Ruslan Salakhutdinov. Generating images with multimodal language", "type": "text" } ], "index": 16 }, { "bbox": [ 114, 318, 317, 333 ], "spans": [ { "bbox": [ 114, 318, 317, 333 ], "score": 1.0, "content": "models. arXiv preprint arXiv:2305.17216, 2023a.", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 338, 506, 352 ], "spans": [ { "bbox": [ 105, 338, 506, 352 ], "score": 1.0, "content": "Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for", "type": "text" } ], "index": 18 }, { "bbox": [ 114, 349, 380, 363 ], "spans": [ { "bbox": [ 114, 349, 380, 363 ], "score": 1.0, "content": "multimodal generation. arXiv preprint arXiv:2301.13823, 2023b.", "type": "text" } ], "index": 19 }, { "bbox": [ 104, 367, 506, 384 ], "spans": [ { "bbox": [ 104, 367, 506, 384 ], "score": 1.0, "content": "Emiel Krahmer and Kees Van Deemter. Computational generation of referring expressions: A", "type": "text" } ], "index": 20 }, { "bbox": [ 114, 379, 348, 393 ], "spans": [ { "bbox": [ 114, 379, 348, 393 ], "score": 1.0, "content": "survey. Computational Linguistics, 38(1):173–218, 2012.", "type": "text" } ], "index": 21 }, { "bbox": [ 106, 399, 505, 411 ], "spans": [ { "bbox": [ 106, 399, 505, 411 ], "score": 1.0, "content": "Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie", "type": "text" } ], "index": 22 }, { "bbox": [ 114, 408, 506, 425 ], "spans": [ { "bbox": [ 114, 408, 506, 425 ], "score": 1.0, "content": "Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting lan-", "type": "text" } ], "index": 23 }, { "bbox": [ 114, 421, 505, 435 ], "spans": [ { "bbox": [ 114, 421, 505, 435 ], "score": 1.0, "content": "guage and vision using crowdsourced dense image annotations. International journal of computer", "type": "text" } ], "index": 24 }, { "bbox": [ 116, 432, 219, 444 ], "spans": [ { "bbox": [ 116, 432, 219, 444 ], "score": 1.0, "content": "vision, 123:32–73, 2017.", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 452, 504, 464 ], "spans": [ { "bbox": [ 106, 452, 504, 464 ], "score": 1.0, "content": "Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Rea-", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 462, 469, 477 ], "spans": [ { "bbox": [ 115, 462, 469, 477 ], "score": 1.0, "content": "soning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023.", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 481, 506, 495 ], "spans": [ { "bbox": [ 105, 481, 506, 495 ], "score": 1.0, "content": "Hugo Laurenc¸on, Lucile Saulnier, Leo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, ´", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 492, 506, 506 ], "spans": [ { "bbox": [ 115, 492, 506, 506 ], "score": 1.0, "content": "Thomas Wang, Siddharth Karamcheti, Alexander M Rush, Douwe Kiela, et al. Obelisc: An open", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 502, 505, 517 ], "spans": [ { "bbox": [ 115, 502, 505, 517 ], "score": 1.0, "content": "web-scale filtered dataset of interleaved image-text documents. arXiv preprint arXiv:2306.16527,", "type": "text" } ], "index": 30 }, { "bbox": [ 114, 514, 144, 529 ], "spans": [ { "bbox": [ 114, 514, 144, 529 ], "score": 1.0, "content": "2023.", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 534, 505, 546 ], "spans": [ { "bbox": [ 106, 534, 505, 546 ], "score": 1.0, "content": "Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 544, 501, 558 ], "spans": [ { "bbox": [ 115, 544, 501, 558 ], "score": 1.0, "content": "multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023a.", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 562, 505, 578 ], "spans": [ { "bbox": [ 105, 562, 505, 578 ], "score": 1.0, "content": "Chunyuan Li, Zhe Gan, Zhengyuan Yang, Jianwei Yang, Linjie Li, Lijuan Wang, and Jianfeng Gao.", "type": "text" } ], "index": 34 }, { "bbox": [ 114, 573, 506, 590 ], "spans": [ { "bbox": [ 114, 573, 506, 590 ], "score": 1.0, "content": "Multimodal foundation models: From specialists to general-purpose assistants. arXiv preprint", "type": "text" } ], "index": 35 }, { "bbox": [ 115, 586, 225, 598 ], "spans": [ { "bbox": [ 115, 586, 225, 598 ], "score": 1.0, "content": "arXiv:2309.10020, 2023b.", "type": "text" } ], "index": 36 }, { "bbox": [ 104, 604, 505, 619 ], "spans": [ { "bbox": [ 104, 604, 505, 619 ], "score": 1.0, "content": "Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-", "type": "text" } ], "index": 37 }, { "bbox": [ 114, 616, 506, 630 ], "spans": [ { "bbox": [ 114, 616, 506, 630 ], "score": 1.0, "content": "image pre-training with frozen image encoders and large language models. arXiv preprint", "type": "text" } ], "index": 38 }, { "bbox": [ 114, 627, 225, 640 ], "spans": [ { "bbox": [ 114, 627, 225, 640 ], "score": 1.0, "content": "arXiv:2301.12597, 2023c.", "type": "text" } ], "index": 39 }, { "bbox": [ 104, 645, 505, 660 ], "spans": [ { "bbox": [ 104, 645, 505, 660 ], "score": 1.0, "content": "Lei Li, Yuwei Yin, Shicheng Li, Liang Chen, Peiyi Wang, Shuhuai Ren, Mukai Li, Yazheng Yang,", "type": "text" } ], "index": 40 }, { "bbox": [ 113, 656, 505, 672 ], "spans": [ { "bbox": [ 113, 656, 225, 672 ], "score": 1.0, "content": "Jingjing Xu, Xu Sun, et al.", "type": "text" }, { "bbox": [ 225, 657, 246, 668 ], "score": 0.27, "content": "\\mathbf { M } ^ { 3 } \\mathrm { i t }", "type": "inline_equation" }, { "bbox": [ 247, 656, 505, 672 ], "score": 1.0, "content": ": A large-scale dataset towards multi-modal multilingual instruc-", "type": "text" } ], "index": 41 }, { "bbox": [ 116, 669, 333, 681 ], "spans": [ { "bbox": [ 116, 669, 333, 681 ], "score": 1.0, "content": "tion tuning. arXiv preprint arXiv:2306.04387, 2023d.", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 687, 505, 702 ], "spans": [ { "bbox": [ 105, 687, 505, 702 ], "score": 1.0, "content": "Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Li-", "type": "text" } ], "index": 43 }, { "bbox": [ 114, 698, 505, 713 ], "spans": [ { "bbox": [ 114, 698, 505, 713 ], "score": 1.0, "content": "juan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training.", "type": "text" } ], "index": 44 }, { "bbox": [ 114, 708, 505, 725 ], "spans": [ { "bbox": [ 114, 708, 505, 725 ], "score": 1.0, "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.", "type": "text" } ], "index": 45 }, { "bbox": [ 117, 721, 201, 733 ], "spans": [ { "bbox": [ 117, 721, 201, 733 ], "score": 1.0, "content": "10965–10975, 2022.", "type": "text" } ], "index": 46 } ], "index": 23 } ], "page_idx": 10, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 310, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 765 ], "spans": [ { "bbox": [ 299, 750, 312, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 99, 65, 507, 746 ], "lines": [], "index": 23, "bbox_fs": [ 104, 82, 506, 733 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 82, 504, 106 ], "lines": [ { "bbox": [ 105, 80, 506, 96 ], "spans": [ { "bbox": [ 105, 80, 506, 96 ], "score": 1.0, "content": "Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating", "type": "text" } ], "index": 0 }, { "bbox": [ 116, 93, 498, 106 ], "spans": [ { "bbox": [ 116, 93, 498, 106 ], "score": 1.0, "content": "object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023e.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 101, 95, 507, 753 ], "lines": [ { "bbox": [ 106, 115, 505, 127 ], "spans": [ { "bbox": [ 106, 115, 505, 127 ], "score": 1.0, "content": "Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr", "type": "text" } ], "index": 2 }, { "bbox": [ 114, 124, 505, 139 ], "spans": [ { "bbox": [ 114, 124, 505, 139 ], "score": 1.0, "content": "Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ´ Computer", "type": "text" } ], "index": 3 }, { "bbox": [ 117, 136, 505, 150 ], "spans": [ { "bbox": [ 117, 136, 505, 150 ], "score": 1.0, "content": "Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014,", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 147, 333, 160 ], "spans": [ { "bbox": [ 116, 147, 333, 160 ], "score": 1.0, "content": "Proceedings, Part V 13, pp. 740–755. Springer, 2014.", "type": "text" } ], "index": 5 }, { "bbox": [ 104, 165, 506, 183 ], "spans": [ { "bbox": [ 104, 165, 506, 183 ], "score": 1.0, "content": "Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large", "type": "text" } ], "index": 6 }, { "bbox": [ 115, 178, 486, 192 ], "spans": [ { "bbox": [ 115, 178, 486, 192 ], "score": 1.0, "content": "multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023a.", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 200, 505, 213 ], "spans": [ { "bbox": [ 106, 200, 505, 213 ], "score": 1.0, "content": "Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv", "type": "text" } ], "index": 8 }, { "bbox": [ 114, 210, 261, 224 ], "spans": [ { "bbox": [ 114, 210, 261, 224 ], "score": 1.0, "content": "preprint arXiv:2304.08485, 2023b.", "type": "text" } ], "index": 9 }, { "bbox": [ 104, 229, 506, 246 ], "spans": [ { "bbox": [ 104, 229, 506, 246 ], "score": 1.0, "content": "Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 242, 221, 256 ], "spans": [ { "bbox": [ 115, 242, 221, 256 ], "score": 1.0, "content": "arXiv:1711.05101, 2017.", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 262, 505, 276 ], "spans": [ { "bbox": [ 106, 262, 505, 276 ], "score": 1.0, "content": "Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi.", "type": "text" } ], "index": 12 }, { "bbox": [ 114, 272, 506, 289 ], "spans": [ { "bbox": [ 114, 272, 506, 289 ], "score": 1.0, "content": "Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 285, 218, 299 ], "spans": [ { "bbox": [ 115, 285, 218, 299 ], "score": 1.0, "content": "arXiv:2206.08916, 2022.", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 305, 505, 319 ], "spans": [ { "bbox": [ 106, 305, 505, 319 ], "score": 1.0, "content": "Ruotian Luo and Gregory Shakhnarovich. Comprehension-guided referring expressions. In Pro-", "type": "text" } ], "index": 15 }, { "bbox": [ 115, 316, 505, 330 ], "spans": [ { "bbox": [ 115, 316, 505, 330 ], "score": 1.0, "content": "ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7102–7111,", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 327, 144, 341 ], "spans": [ { "bbox": [ 115, 327, 144, 341 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 347, 505, 361 ], "spans": [ { "bbox": [ 106, 347, 505, 361 ], "score": 1.0, "content": "Chih-Yao Ma, Yannis Kalantidis, Ghassan AlRegib, Peter Vajda, Marcus Rohrbach, and Zsolt Kira.", "type": "text" } ], "index": 18 }, { "bbox": [ 115, 360, 504, 373 ], "spans": [ { "bbox": [ 115, 360, 504, 373 ], "score": 1.0, "content": "Learning to generate grounded visual captions without localization supervision. In ECCV, 2020.", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 379, 505, 393 ], "spans": [ { "bbox": [ 105, 379, 505, 393 ], "score": 1.0, "content": "Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local", "type": "text" } ], "index": 20 }, { "bbox": [ 114, 390, 505, 406 ], "spans": [ { "bbox": [ 114, 390, 505, 406 ], "score": 1.0, "content": "geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123,", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 401, 143, 415 ], "spans": [ { "bbox": [ 115, 401, 143, 415 ], "score": 1.0, "content": "2022.", "type": "text" } ], "index": 22 }, { "bbox": [ 104, 421, 505, 436 ], "spans": [ { "bbox": [ 104, 421, 505, 436 ], "score": 1.0, "content": "Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy.", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 433, 505, 447 ], "spans": [ { "bbox": [ 115, 433, 505, 447 ], "score": 1.0, "content": "Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE", "type": "text" } ], "index": 24 }, { "bbox": [ 114, 443, 410, 459 ], "spans": [ { "bbox": [ 114, 443, 410, 459 ], "score": 1.0, "content": "conference on computer vision and pattern recognition, pp. 11–20, 2016.", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 464, 506, 479 ], "spans": [ { "bbox": [ 106, 464, 506, 479 ], "score": 1.0, "content": "Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 476, 504, 489 ], "spans": [ { "bbox": [ 115, 476, 504, 489 ], "score": 1.0, "content": "referring expression understanding. In Computer Vision–ECCV 2016: 14th European Confer-", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 487, 505, 500 ], "spans": [ { "bbox": [ 115, 487, 505, 500 ], "score": 1.0, "content": "ence, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14, pp. 792–807.", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 497, 182, 511 ], "spans": [ { "bbox": [ 116, 497, 182, 511 ], "score": 1.0, "content": "Springer, 2016.", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 519, 468, 532 ], "spans": [ { "bbox": [ 106, 519, 468, 532 ], "score": 1.0, "content": "OpenAI. GPT-4 technical report. https://arxiv.org/abs/2303.08774, 2023a.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 539, 296, 552 ], "spans": [ { "bbox": [ 105, 539, 296, 552 ], "score": 1.0, "content": "OpenAI. Gpt-4 technical report. arXiv, 2023b.", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 560, 505, 573 ], "spans": [ { "bbox": [ 106, 560, 505, 573 ], "score": 1.0, "content": "Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 570, 506, 586 ], "spans": [ { "bbox": [ 115, 570, 506, 586 ], "score": 1.0, "content": "Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint", "type": "text" } ], "index": 33 }, { "bbox": [ 114, 582, 219, 595 ], "spans": [ { "bbox": [ 114, 582, 219, 595 ], "score": 1.0, "content": "arXiv:2306.14824, 2023.", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 601, 505, 617 ], "spans": [ { "bbox": [ 105, 601, 505, 617 ], "score": 1.0, "content": "Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svet-", "type": "text" } ], "index": 35 }, { "bbox": [ 114, 612, 506, 628 ], "spans": [ { "bbox": [ 114, 612, 506, 628 ], "score": 1.0, "content": "lana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-", "type": "text" } ], "index": 36 }, { "bbox": [ 114, 623, 506, 640 ], "spans": [ { "bbox": [ 114, 623, 506, 640 ], "score": 1.0, "content": "to-sentence models. In Proceedings of the IEEE international conference on computer vision, pp.", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 635, 192, 649 ], "spans": [ { "bbox": [ 115, 635, 192, 649 ], "score": 1.0, "content": "2641–2649, 2015.", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 655, 505, 670 ], "spans": [ { "bbox": [ 105, 655, 505, 670 ], "score": 1.0, "content": "Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets", "type": "text" } ], "index": 39 }, { "bbox": [ 115, 668, 505, 681 ], "spans": [ { "bbox": [ 115, 668, 505, 681 ], "score": 1.0, "content": "for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision", "type": "text" } ], "index": 40 }, { "bbox": [ 116, 678, 298, 692 ], "spans": [ { "bbox": [ 116, 678, 298, 692 ], "score": 1.0, "content": "and pattern recognition, pp. 652–660, 2017a.", "type": "text" } ], "index": 41 }, { "bbox": [ 106, 699, 505, 712 ], "spans": [ { "bbox": [ 106, 699, 505, 712 ], "score": 1.0, "content": "Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical fea-", "type": "text" } ], "index": 42 }, { "bbox": [ 114, 709, 505, 724 ], "spans": [ { "bbox": [ 114, 709, 505, 724 ], "score": 1.0, "content": "ture learning on point sets in a metric space. Advances in neural information processing systems,", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 720, 162, 734 ], "spans": [ { "bbox": [ 115, 720, 162, 734 ], "score": 1.0, "content": "30, 2017b.", "type": "text" } ], "index": 44 } ], "index": 23 } ], "page_idx": 11, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "12", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 82, 504, 106 ], "lines": [ { "bbox": [ 105, 80, 506, 96 ], "spans": [ { "bbox": [ 105, 80, 506, 96 ], "score": 1.0, "content": "Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating", "type": "text" } ], "index": 0 }, { "bbox": [ 116, 93, 498, 106 ], "spans": [ { "bbox": [ 116, 93, 498, 106 ], "score": 1.0, "content": "object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023e.", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 105, 80, 506, 106 ] }, { "type": "list", "bbox": [ 101, 95, 507, 753 ], "lines": [ { "bbox": [ 106, 115, 505, 127 ], "spans": [ { "bbox": [ 106, 115, 505, 127 ], "score": 1.0, "content": "Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr", "type": "text" } ], "index": 2, "is_list_start_line": true }, { "bbox": [ 114, 124, 505, 139 ], "spans": [ { "bbox": [ 114, 124, 505, 139 ], "score": 1.0, "content": "Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ´ Computer", "type": "text" } ], "index": 3 }, { "bbox": [ 117, 136, 505, 150 ], "spans": [ { "bbox": [ 117, 136, 505, 150 ], "score": 1.0, "content": "Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014,", "type": "text" } ], "index": 4 }, { "bbox": [ 116, 147, 333, 160 ], "spans": [ { "bbox": [ 116, 147, 333, 160 ], "score": 1.0, "content": "Proceedings, Part V 13, pp. 740–755. Springer, 2014.", "type": "text" } ], "index": 5, "is_list_end_line": true }, { "bbox": [ 104, 165, 506, 183 ], "spans": [ { "bbox": [ 104, 165, 506, 183 ], "score": 1.0, "content": "Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. Aligning large", "type": "text" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 115, 178, 486, 192 ], "spans": [ { "bbox": [ 115, 178, 486, 192 ], "score": 1.0, "content": "multi-modal model with robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023a.", "type": "text" } ], "index": 7, "is_list_end_line": true }, { "bbox": [ 106, 200, 505, 213 ], "spans": [ { "bbox": [ 106, 200, 505, 213 ], "score": 1.0, "content": "Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv", "type": "text" } ], "index": 8, "is_list_start_line": true }, { "bbox": [ 114, 210, 261, 224 ], "spans": [ { "bbox": [ 114, 210, 261, 224 ], "score": 1.0, "content": "preprint arXiv:2304.08485, 2023b.", "type": "text" } ], "index": 9, "is_list_end_line": true }, { "bbox": [ 104, 229, 506, 246 ], "spans": [ { "bbox": [ 104, 229, 506, 246 ], "score": 1.0, "content": "Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint", "type": "text" } ], "index": 10, "is_list_start_line": true }, { "bbox": [ 115, 242, 221, 256 ], "spans": [ { "bbox": [ 115, 242, 221, 256 ], "score": 1.0, "content": "arXiv:1711.05101, 2017.", "type": "text" } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 106, 262, 505, 276 ], "spans": [ { "bbox": [ 106, 262, 505, 276 ], "score": 1.0, "content": "Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi.", "type": "text" } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 114, 272, 506, 289 ], "spans": [ { "bbox": [ 114, 272, 506, 289 ], "score": 1.0, "content": "Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 285, 218, 299 ], "spans": [ { "bbox": [ 115, 285, 218, 299 ], "score": 1.0, "content": "arXiv:2206.08916, 2022.", "type": "text" } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 106, 305, 505, 319 ], "spans": [ { "bbox": [ 106, 305, 505, 319 ], "score": 1.0, "content": "Ruotian Luo and Gregory Shakhnarovich. Comprehension-guided referring expressions. In Pro-", "type": "text" } ], "index": 15, "is_list_start_line": true }, { "bbox": [ 115, 316, 505, 330 ], "spans": [ { "bbox": [ 115, 316, 505, 330 ], "score": 1.0, "content": "ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7102–7111,", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 327, 144, 341 ], "spans": [ { "bbox": [ 115, 327, 144, 341 ], "score": 1.0, "content": "2017.", "type": "text" } ], "index": 17, "is_list_end_line": true }, { "bbox": [ 106, 347, 505, 361 ], "spans": [ { "bbox": [ 106, 347, 505, 361 ], "score": 1.0, "content": "Chih-Yao Ma, Yannis Kalantidis, Ghassan AlRegib, Peter Vajda, Marcus Rohrbach, and Zsolt Kira.", "type": "text" } ], "index": 18, "is_list_start_line": true }, { "bbox": [ 115, 360, 504, 373 ], "spans": [ { "bbox": [ 115, 360, 504, 373 ], "score": 1.0, "content": "Learning to generate grounded visual captions without localization supervision. In ECCV, 2020.", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 379, 505, 393 ], "spans": [ { "bbox": [ 105, 379, 505, 393 ], "score": 1.0, "content": "Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local", "type": "text" } ], "index": 20, "is_list_start_line": true }, { "bbox": [ 114, 390, 505, 406 ], "spans": [ { "bbox": [ 114, 390, 505, 406 ], "score": 1.0, "content": "geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123,", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 401, 143, 415 ], "spans": [ { "bbox": [ 115, 401, 143, 415 ], "score": 1.0, "content": "2022.", "type": "text" } ], "index": 22, "is_list_end_line": true }, { "bbox": [ 104, 421, 505, 436 ], "spans": [ { "bbox": [ 104, 421, 505, 436 ], "score": 1.0, "content": "Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy.", "type": "text" } ], "index": 23, "is_list_start_line": true }, { "bbox": [ 115, 433, 505, 447 ], "spans": [ { "bbox": [ 115, 433, 505, 447 ], "score": 1.0, "content": "Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE", "type": "text" } ], "index": 24 }, { "bbox": [ 114, 443, 410, 459 ], "spans": [ { "bbox": [ 114, 443, 410, 459 ], "score": 1.0, "content": "conference on computer vision and pattern recognition, pp. 11–20, 2016.", "type": "text" } ], "index": 25, "is_list_end_line": true }, { "bbox": [ 106, 464, 506, 479 ], "spans": [ { "bbox": [ 106, 464, 506, 479 ], "score": 1.0, "content": "Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for", "type": "text" } ], "index": 26, "is_list_start_line": true }, { "bbox": [ 115, 476, 504, 489 ], "spans": [ { "bbox": [ 115, 476, 504, 489 ], "score": 1.0, "content": "referring expression understanding. In Computer Vision–ECCV 2016: 14th European Confer-", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 487, 505, 500 ], "spans": [ { "bbox": [ 115, 487, 505, 500 ], "score": 1.0, "content": "ence, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14, pp. 792–807.", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 497, 182, 511 ], "spans": [ { "bbox": [ 116, 497, 182, 511 ], "score": 1.0, "content": "Springer, 2016.", "type": "text" } ], "index": 29, "is_list_end_line": true }, { "bbox": [ 106, 519, 468, 532 ], "spans": [ { "bbox": [ 106, 519, 468, 532 ], "score": 1.0, "content": "OpenAI. GPT-4 technical report. https://arxiv.org/abs/2303.08774, 2023a.", "type": "text" } ], "index": 30, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 539, 296, 552 ], "spans": [ { "bbox": [ 105, 539, 296, 552 ], "score": 1.0, "content": "OpenAI. Gpt-4 technical report. arXiv, 2023b.", "type": "text" } ], "index": 31, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 106, 560, 505, 573 ], "spans": [ { "bbox": [ 106, 560, 505, 573 ], "score": 1.0, "content": "Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu", "type": "text" } ], "index": 32, "is_list_start_line": true }, { "bbox": [ 115, 570, 506, 586 ], "spans": [ { "bbox": [ 115, 570, 506, 586 ], "score": 1.0, "content": "Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint", "type": "text" } ], "index": 33 }, { "bbox": [ 114, 582, 219, 595 ], "spans": [ { "bbox": [ 114, 582, 219, 595 ], "score": 1.0, "content": "arXiv:2306.14824, 2023.", "type": "text" } ], "index": 34, "is_list_end_line": true }, { "bbox": [ 105, 601, 505, 617 ], "spans": [ { "bbox": [ 105, 601, 505, 617 ], "score": 1.0, "content": "Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svet-", "type": "text" } ], "index": 35, "is_list_start_line": true }, { "bbox": [ 114, 612, 506, 628 ], "spans": [ { "bbox": [ 114, 612, 506, 628 ], "score": 1.0, "content": "lana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-", "type": "text" } ], "index": 36 }, { "bbox": [ 114, 623, 506, 640 ], "spans": [ { "bbox": [ 114, 623, 506, 640 ], "score": 1.0, "content": "to-sentence models. In Proceedings of the IEEE international conference on computer vision, pp.", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 635, 192, 649 ], "spans": [ { "bbox": [ 115, 635, 192, 649 ], "score": 1.0, "content": "2641–2649, 2015.", "type": "text" } ], "index": 38, "is_list_end_line": true }, { "bbox": [ 105, 655, 505, 670 ], "spans": [ { "bbox": [ 105, 655, 505, 670 ], "score": 1.0, "content": "Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets", "type": "text" } ], "index": 39, "is_list_start_line": true }, { "bbox": [ 115, 668, 505, 681 ], "spans": [ { "bbox": [ 115, 668, 505, 681 ], "score": 1.0, "content": "for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision", "type": "text" } ], "index": 40 }, { "bbox": [ 116, 678, 298, 692 ], "spans": [ { "bbox": [ 116, 678, 298, 692 ], "score": 1.0, "content": "and pattern recognition, pp. 652–660, 2017a.", "type": "text" } ], "index": 41, "is_list_end_line": true }, { "bbox": [ 106, 699, 505, 712 ], "spans": [ { "bbox": [ 106, 699, 505, 712 ], "score": 1.0, "content": "Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical fea-", "type": "text" } ], "index": 42, "is_list_start_line": true }, { "bbox": [ 114, 709, 505, 724 ], "spans": [ { "bbox": [ 114, 709, 505, 724 ], "score": 1.0, "content": "ture learning on point sets in a metric space. Advances in neural information processing systems,", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 720, 162, 734 ], "spans": [ { "bbox": [ 115, 720, 162, 734 ], "score": 1.0, "content": "30, 2017b.", "type": "text" } ], "index": 44, "is_list_end_line": true }, { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,", "type": "text", "cross_page": true } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 115, 93, 506, 108 ], "spans": [ { "bbox": [ 115, 93, 506, 108 ], "score": 1.0, "content": "Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 113, 101, 506, 122 ], "spans": [ { "bbox": [ 113, 101, 506, 122 ], "score": 1.0, "content": "models from natural language supervision. In International conference on machine learning, pp.", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 115, 115, 226, 129 ], "spans": [ { "bbox": [ 115, 115, 226, 129 ], "score": 1.0, "content": "8748–8763. PMLR, 2021.", "type": "text", "cross_page": true } ], "index": 3, "is_list_end_line": true }, { "bbox": [ 105, 133, 506, 148 ], "spans": [ { "bbox": [ 105, 133, 506, 148 ], "score": 1.0, "content": "Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman ´", "type": "text", "cross_page": true } ], "index": 4, "is_list_start_line": true }, { "bbox": [ 115, 144, 505, 159 ], "spans": [ { "bbox": [ 115, 144, 505, 159 ], "score": 1.0, "content": "Castagne, Alexandra Sasha Luccioni, Franc¸ois Yvon, Matthias Gall ´ e, et al. Bloom: A 176b- ´", "type": "text", "cross_page": true } ], "index": 5 }, { "bbox": [ 114, 155, 492, 170 ], "spans": [ { "bbox": [ 114, 155, 492, 170 ], "score": 1.0, "content": "parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022.", "type": "text", "cross_page": true } ], "index": 6, "is_list_end_line": true }, { "bbox": [ 105, 176, 506, 190 ], "spans": [ { "bbox": [ 105, 176, 506, 190 ], "score": 1.0, "content": "Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian", "type": "text", "cross_page": true } ], "index": 7, "is_list_start_line": true }, { "bbox": [ 115, 187, 506, 201 ], "spans": [ { "bbox": [ 115, 187, 506, 201 ], "score": 1.0, "content": "Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 115, 197, 434, 213 ], "spans": [ { "bbox": [ 115, 197, 434, 213 ], "score": 1.0, "content": "IEEE/CVF international conference on computer vision, pp. 8430–8439, 2019.", "type": "text", "cross_page": true } ], "index": 9, "is_list_end_line": true }, { "bbox": [ 107, 218, 505, 230 ], "spans": [ { "bbox": [ 107, 218, 505, 230 ], "score": 1.0, "content": "Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao,", "type": "text", "cross_page": true } ], "index": 10, "is_list_start_line": true }, { "bbox": [ 113, 226, 507, 243 ], "spans": [ { "bbox": [ 113, 226, 507, 243 ], "score": 1.0, "content": "Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 114, 239, 254, 252 ], "spans": [ { "bbox": [ 114, 239, 254, 252 ], "score": 1.0, "content": "preprint arXiv:2307.05222, 2023.", "type": "text", "cross_page": true } ], "index": 12, "is_list_end_line": true }, { "bbox": [ 106, 258, 505, 270 ], "spans": [ { "bbox": [ 106, 258, 505, 270 ], "score": 1.0, "content": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee´", "type": "text", "cross_page": true } ], "index": 13, "is_list_start_line": true }, { "bbox": [ 115, 268, 506, 282 ], "spans": [ { "bbox": [ 115, 268, 506, 282 ], "score": 1.0, "content": "Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and `", "type": "text", "cross_page": true } ], "index": 14 }, { "bbox": [ 115, 280, 437, 294 ], "spans": [ { "bbox": [ 115, 280, 437, 294 ], "score": 1.0, "content": "efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023a.", "type": "text", "cross_page": true } ], "index": 15, "is_list_end_line": true }, { "bbox": [ 106, 300, 504, 311 ], "spans": [ { "bbox": [ 106, 300, 504, 311 ], "score": 1.0, "content": "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Niko-", "type": "text", "cross_page": true } ], "index": 16, "is_list_start_line": true }, { "bbox": [ 115, 309, 505, 323 ], "spans": [ { "bbox": [ 115, 309, 505, 323 ], "score": 1.0, "content": "lay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open founda-", "type": "text", "cross_page": true } ], "index": 17 }, { "bbox": [ 116, 322, 413, 334 ], "spans": [ { "bbox": [ 116, 322, 413, 334 ], "score": 1.0, "content": "tion and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023b.", "type": "text", "cross_page": true } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 105, 340, 506, 354 ], "spans": [ { "bbox": [ 105, 340, 506, 354 ], "score": 1.0, "content": "Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu,", "type": "text", "cross_page": true } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 115, 350, 505, 365 ], "spans": [ { "bbox": [ 115, 350, 505, 365 ], "score": 1.0, "content": "and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv", "type": "text", "cross_page": true } ], "index": 20 }, { "bbox": [ 115, 362, 260, 376 ], "spans": [ { "bbox": [ 115, 362, 260, 376 ], "score": 1.0, "content": "preprint arXiv:2205.14100, 2022a.", "type": "text", "cross_page": true } ], "index": 21, "is_list_end_line": true }, { "bbox": [ 105, 381, 506, 395 ], "spans": [ { "bbox": [ 105, 381, 506, 395 ], "score": 1.0, "content": "Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou,", "type": "text", "cross_page": true } ], "index": 22, "is_list_start_line": true }, { "bbox": [ 115, 392, 506, 406 ], "spans": [ { "bbox": [ 115, 392, 506, 406 ], "score": 1.0, "content": "Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a", "type": "text", "cross_page": true } ], "index": 23 }, { "bbox": [ 115, 403, 389, 417 ], "spans": [ { "bbox": [ 115, 403, 389, 417 ], "score": 1.0, "content": "simple sequence-to-sequence learning framework. In ICML, 2022b.", "type": "text", "cross_page": true } ], "index": 24, "is_list_end_line": true }, { "bbox": [ 105, 419, 506, 437 ], "spans": [ { "bbox": [ 105, 419, 506, 437 ], "score": 1.0, "content": "Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong", "type": "text", "cross_page": true } ], "index": 25, "is_list_start_line": true }, { "bbox": [ 113, 431, 506, 448 ], "spans": [ { "bbox": [ 113, 431, 506, 448 ], "score": 1.0, "content": "Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for", "type": "text", "cross_page": true } ], "index": 26 }, { "bbox": [ 116, 443, 362, 457 ], "spans": [ { "bbox": [ 116, 443, 362, 457 ], "score": 1.0, "content": "vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023.", "type": "text", "cross_page": true } ], "index": 27, "is_list_end_line": true }, { "bbox": [ 108, 464, 504, 475 ], "spans": [ { "bbox": [ 108, 464, 504, 475 ], "score": 1.0, "content": "Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon.", "type": "text", "cross_page": true } ], "index": 28, "is_list_start_line": true }, { "bbox": [ 115, 474, 506, 488 ], "spans": [ { "bbox": [ 115, 474, 506, 488 ], "score": 1.0, "content": "Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):", "type": "text", "cross_page": true } ], "index": 29 }, { "bbox": [ 115, 484, 168, 498 ], "spans": [ { "bbox": [ 115, 484, 168, 498 ], "score": 1.0, "content": "1–12, 2019.", "type": "text", "cross_page": true } ], "index": 30, "is_list_end_line": true }, { "bbox": [ 105, 504, 506, 518 ], "spans": [ { "bbox": [ 105, 504, 506, 518 ], "score": 1.0, "content": "Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple", "type": "text", "cross_page": true } ], "index": 31, "is_list_start_line": true }, { "bbox": [ 115, 515, 417, 529 ], "spans": [ { "bbox": [ 115, 515, 417, 529 ], "score": 1.0, "content": "visual language model pretraining with weak supervision. In ICLR, 2022c.", "type": "text", "cross_page": true } ], "index": 32, "is_list_end_line": true }, { "bbox": [ 105, 534, 506, 548 ], "spans": [ { "bbox": [ 105, 534, 506, 548 ], "score": 1.0, "content": "Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan", "type": "text", "cross_page": true } ], "index": 33, "is_list_start_line": true }, { "bbox": [ 115, 544, 506, 560 ], "spans": [ { "bbox": [ 115, 544, 506, 560 ], "score": 1.0, "content": "Wang. Grit: A generative region-to-text transformer for object understanding. arXiv preprint", "type": "text", "cross_page": true } ], "index": 34 }, { "bbox": [ 115, 555, 220, 569 ], "spans": [ { "bbox": [ 115, 555, 220, 569 ], "score": 1.0, "content": "arXiv:2212.00280, 2022.", "type": "text", "cross_page": true } ], "index": 35, "is_list_end_line": true }, { "bbox": [ 106, 576, 505, 588 ], "spans": [ { "bbox": [ 106, 576, 505, 588 ], "score": 1.0, "content": "Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu,", "type": "text", "cross_page": true } ], "index": 36, "is_list_start_line": true }, { "bbox": [ 113, 584, 505, 602 ], "spans": [ { "bbox": [ 113, 584, 505, 602 ], "score": 1.0, "content": "and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling.", "type": "text", "cross_page": true } ], "index": 37 }, { "bbox": [ 115, 597, 420, 612 ], "spans": [ { "bbox": [ 115, 597, 420, 612 ], "score": 1.0, "content": "In European Conference on Computer Vision, pp. 521–539. Springer, 2022.", "type": "text", "cross_page": true } ], "index": 38, "is_list_end_line": true }, { "bbox": [ 106, 616, 506, 630 ], "spans": [ { "bbox": [ 106, 616, 506, 630 ], "score": 1.0, "content": "Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen", "type": "text", "cross_page": true } ], "index": 39, "is_list_start_line": true }, { "bbox": [ 115, 628, 506, 642 ], "spans": [ { "bbox": [ 115, 628, 506, 642 ], "score": 1.0, "content": "Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models", "type": "text", "cross_page": true } ], "index": 40 }, { "bbox": [ 115, 638, 361, 652 ], "spans": [ { "bbox": [ 115, 638, 361, 652 ], "score": 1.0, "content": "with multimodality. arXiv preprint arXiv:2304.14178, 2023.", "type": "text", "cross_page": true } ], "index": 41, "is_list_end_line": true }, { "bbox": [ 105, 657, 506, 671 ], "spans": [ { "bbox": [ 105, 657, 506, 671 ], "score": 1.0, "content": "Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context", "type": "text", "cross_page": true } ], "index": 42, "is_list_start_line": true }, { "bbox": [ 115, 668, 505, 682 ], "spans": [ { "bbox": [ 115, 668, 505, 682 ], "score": 1.0, "content": "in referring expressions. In Computer Vision–ECCV 2016: 14th European Conference, Amster-", "type": "text", "cross_page": true } ], "index": 43 }, { "bbox": [ 116, 680, 504, 692 ], "spans": [ { "bbox": [ 116, 680, 504, 692 ], "score": 1.0, "content": "dam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pp. 69–85. Springer, 2016.", "type": "text", "cross_page": true } ], "index": 44 }, { "bbox": [ 105, 699, 506, 711 ], "spans": [ { "bbox": [ 105, 699, 506, 711 ], "score": 1.0, "content": "Licheng Yu, Hao Tan, Mohit Bansal, and Tamara L Berg. A joint speaker-listener-reinforcer model", "type": "text", "cross_page": true } ], "index": 45, "is_list_start_line": true }, { "bbox": [ 115, 710, 506, 724 ], "spans": [ { "bbox": [ 115, 710, 506, 724 ], "score": 1.0, "content": "for referring expressions. In Proceedings of the IEEE conference on computer vision and pattern", "type": "text", "cross_page": true } ], "index": 46 }, { "bbox": [ 115, 720, 257, 734 ], "spans": [ { "bbox": [ 115, 720, 257, 734 ], "score": 1.0, "content": "recognition, pp. 7282–7290, 2017.", "type": "text", "cross_page": true } ], "index": 47, "is_list_end_line": true }, { "bbox": [ 106, 83, 505, 95 ], "spans": [ { "bbox": [ 106, 83, 505, 95 ], "score": 1.0, "content": "Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mat-", "type": "text", "cross_page": true } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 115, 93, 476, 107 ], "spans": [ { "bbox": [ 115, 93, 476, 107 ], "score": 1.0, "content": "tnet: Modular attention network for referring expression comprehension. In CVPR, 2018.", "type": "text", "cross_page": true } ], "index": 1, "is_list_end_line": true }, { "bbox": [ 104, 111, 506, 126 ], "spans": [ { "bbox": [ 104, 111, 506, 126 ], "score": 1.0, "content": "Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun", "type": "text", "cross_page": true } ], "index": 2, "is_list_start_line": true }, { "bbox": [ 115, 123, 506, 137 ], "spans": [ { "bbox": [ 115, 123, 506, 137 ], "score": 1.0, "content": "Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multi-modal models:", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 115, 134, 419, 148 ], "spans": [ { "bbox": [ 115, 134, 419, 148 ], "score": 1.0, "content": "Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023.", "type": "text", "cross_page": true } ], "index": 4, "is_list_end_line": true }, { "bbox": [ 105, 153, 506, 167 ], "spans": [ { "bbox": [ 105, 153, 506, 167 ], "score": 1.0, "content": "Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection", "type": "text", "cross_page": true } ], "index": 5, "is_list_start_line": true }, { "bbox": [ 115, 164, 443, 177 ], "spans": [ { "bbox": [ 115, 164, 443, 177 ], "score": 1.0, "content": "with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023.", "type": "text", "cross_page": true } ], "index": 6, "is_list_end_line": true }, { "bbox": [ 106, 183, 505, 195 ], "spans": [ { "bbox": [ 106, 183, 505, 195 ], "score": 1.0, "content": "Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual", "type": "text", "cross_page": true } ], "index": 7, "is_list_start_line": true }, { "bbox": [ 114, 194, 506, 208 ], "spans": [ { "bbox": [ 114, 194, 506, 208 ], "score": 1.0, "content": "commonsense reasoning. In Proceedings of the IEEE/CVF conference on computer vision and", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 115, 206, 288, 218 ], "spans": [ { "bbox": [ 115, 206, 288, 218 ], "score": 1.0, "content": "pattern recognition, pp. 6720–6731, 2019.", "type": "text", "cross_page": true } ], "index": 9, "is_list_end_line": true }, { "bbox": [ 105, 223, 506, 237 ], "spans": [ { "bbox": [ 105, 223, 506, 237 ], "score": 1.0, "content": "Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan", "type": "text", "cross_page": true } ], "index": 10, "is_list_start_line": true }, { "bbox": [ 116, 235, 505, 248 ], "spans": [ { "bbox": [ 116, 235, 505, 248 ], "score": 1.0, "content": "Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 115, 246, 506, 259 ], "spans": [ { "bbox": [ 115, 246, 506, 259 ], "score": 1.0, "content": "language understanding. Advances in Neural Information Processing Systems, 35:36067–36080,", "type": "text", "cross_page": true } ], "index": 12 }, { "bbox": [ 116, 257, 141, 269 ], "spans": [ { "bbox": [ 116, 257, 141, 269 ], "score": 1.0, "content": "2022.", "type": "text", "cross_page": true } ], "index": 13, "is_list_end_line": true }, { "bbox": [ 105, 276, 506, 289 ], "spans": [ { "bbox": [ 105, 276, 506, 289 ], "score": 1.0, "content": "Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and", "type": "text", "cross_page": true } ], "index": 14, "is_list_start_line": true }, { "bbox": [ 115, 287, 506, 300 ], "spans": [ { "bbox": [ 115, 287, 506, 300 ], "score": 1.0, "content": "Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint", "type": "text", "cross_page": true } ], "index": 15 }, { "bbox": [ 115, 298, 218, 310 ], "spans": [ { "bbox": [ 115, 298, 218, 310 ], "score": 1.0, "content": "arXiv:2307.03601, 2023.", "type": "text", "cross_page": true } ], "index": 16, "is_list_end_line": true }, { "bbox": [ 106, 315, 505, 330 ], "spans": [ { "bbox": [ 106, 315, 505, 330 ], "score": 1.0, "content": "Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt:", "type": "text", "cross_page": true } ], "index": 17, "is_list_start_line": true }, { "bbox": [ 115, 327, 472, 341 ], "spans": [ { "bbox": [ 115, 327, 472, 341 ], "score": 1.0, "content": "Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023.", "type": "text", "cross_page": true } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 105, 347, 505, 359 ], "spans": [ { "bbox": [ 105, 347, 505, 359 ], "score": 1.0, "content": "Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. Grounded", "type": "text", "cross_page": true } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 116, 358, 257, 370 ], "spans": [ { "bbox": [ 116, 358, 257, 370 ], "score": 1.0, "content": "video description. In CVPR, 2019.", "type": "text", "cross_page": true } ], "index": 20, "is_list_end_line": true }, { "bbox": [ 105, 374, 506, 391 ], "spans": [ { "bbox": [ 105, 374, 506, 391 ], "score": 1.0, "content": "Yuanen Zhou, Meng Wang, Daqing Liu, Zhenzhen Hu, and Hanwang Zhang. More grounded image", "type": "text", "cross_page": true } ], "index": 21, "is_list_start_line": true }, { "bbox": [ 115, 389, 392, 400 ], "spans": [ { "bbox": [ 115, 389, 392, 400 ], "score": 1.0, "content": "captioning by distilling image-text matching model. In CVPR, 2020.", "type": "text", "cross_page": true } ], "index": 22, "is_list_end_line": true }, { "bbox": [ 105, 406, 505, 420 ], "spans": [ { "bbox": [ 105, 406, 505, 420 ], "score": 1.0, "content": "Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: En-", "type": "text", "cross_page": true } ], "index": 23, "is_list_start_line": true }, { "bbox": [ 115, 417, 506, 431 ], "spans": [ { "bbox": [ 115, 417, 506, 431 ], "score": 1.0, "content": "hancing vision-language understanding with advanced large language models. arXiv preprint", "type": "text", "cross_page": true } ], "index": 24 }, { "bbox": [ 115, 428, 223, 440 ], "spans": [ { "bbox": [ 115, 428, 223, 440 ], "score": 1.0, "content": "arXiv:2304.10592, 2023a.", "type": "text", "cross_page": true } ], "index": 25, "is_list_end_line": true }, { "bbox": [ 105, 445, 506, 462 ], "spans": [ { "bbox": [ 105, 445, 506, 462 ], "score": 1.0, "content": "Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Young-", "type": "text", "cross_page": true } ], "index": 26, "is_list_start_line": true }, { "bbox": [ 115, 458, 505, 472 ], "spans": [ { "bbox": [ 115, 458, 505, 472 ], "score": 1.0, "content": "jae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-", "type": "text", "cross_page": true } ], "index": 27 }, { "bbox": [ 115, 469, 464, 483 ], "spans": [ { "bbox": [ 115, 469, 464, 483 ], "score": 1.0, "content": "scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023b.", "type": "text", "cross_page": true } ], "index": 28, "is_list_end_line": true }, { "bbox": [ 105, 486, 505, 503 ], "spans": [ { "bbox": [ 105, 486, 505, 503 ], "score": 1.0, "content": "Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Seg-", "type": "text", "cross_page": true } ], "index": 29, "is_list_start_line": true }, { "bbox": [ 115, 499, 441, 512 ], "spans": [ { "bbox": [ 115, 499, 441, 512 ], "score": 1.0, "content": "ment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023.", "type": "text", "cross_page": true } ], "index": 30, "is_list_end_line": true } ], "index": 23, "bbox_fs": [ 104, 115, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 101, 58, 506, 732 ], "lines": [ { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 506, 108 ], "spans": [ { "bbox": [ 115, 93, 506, 108 ], "score": 1.0, "content": "Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual", "type": "text" } ], "index": 1 }, { "bbox": [ 113, 101, 506, 122 ], "spans": [ { "bbox": [ 113, 101, 506, 122 ], "score": 1.0, "content": "models from natural language supervision. In International conference on machine learning, pp.", "type": "text" } ], "index": 2 }, { "bbox": [ 115, 115, 226, 129 ], "spans": [ { "bbox": [ 115, 115, 226, 129 ], "score": 1.0, "content": "8748–8763. PMLR, 2021.", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 133, 506, 148 ], "spans": [ { "bbox": [ 105, 133, 506, 148 ], "score": 1.0, "content": "Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman ´", "type": "text" } ], "index": 4 }, { "bbox": [ 115, 144, 505, 159 ], "spans": [ { "bbox": [ 115, 144, 505, 159 ], "score": 1.0, "content": "Castagne, Alexandra Sasha Luccioni, Franc¸ois Yvon, Matthias Gall ´ e, et al. Bloom: A 176b- ´", "type": "text" } ], "index": 5 }, { "bbox": [ 114, 155, 492, 170 ], "spans": [ { "bbox": [ 114, 155, 492, 170 ], "score": 1.0, "content": "parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022.", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 176, 506, 190 ], "spans": [ { "bbox": [ 105, 176, 506, 190 ], "score": 1.0, "content": "Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 187, 506, 201 ], "spans": [ { "bbox": [ 115, 187, 506, 201 ], "score": 1.0, "content": "Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 197, 434, 213 ], "spans": [ { "bbox": [ 115, 197, 434, 213 ], "score": 1.0, "content": "IEEE/CVF international conference on computer vision, pp. 8430–8439, 2019.", "type": "text" } ], "index": 9 }, { "bbox": [ 107, 218, 505, 230 ], "spans": [ { "bbox": [ 107, 218, 505, 230 ], "score": 1.0, "content": "Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao,", "type": "text" } ], "index": 10 }, { "bbox": [ 113, 226, 507, 243 ], "spans": [ { "bbox": [ 113, 226, 507, 243 ], "score": 1.0, "content": "Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv", "type": "text" } ], "index": 11 }, { "bbox": [ 114, 239, 254, 252 ], "spans": [ { "bbox": [ 114, 239, 254, 252 ], "score": 1.0, "content": "preprint arXiv:2307.05222, 2023.", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 258, 505, 270 ], "spans": [ { "bbox": [ 106, 258, 505, 270 ], "score": 1.0, "content": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee´", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 268, 506, 282 ], "spans": [ { "bbox": [ 115, 268, 506, 282 ], "score": 1.0, "content": "Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and `", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 280, 437, 294 ], "spans": [ { "bbox": [ 115, 280, 437, 294 ], "score": 1.0, "content": "efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023a.", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 300, 504, 311 ], "spans": [ { "bbox": [ 106, 300, 504, 311 ], "score": 1.0, "content": "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Niko-", "type": "text" } ], "index": 16 }, { "bbox": [ 115, 309, 505, 323 ], "spans": [ { "bbox": [ 115, 309, 505, 323 ], "score": 1.0, "content": "lay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open founda-", "type": "text" } ], "index": 17 }, { "bbox": [ 116, 322, 413, 334 ], "spans": [ { "bbox": [ 116, 322, 413, 334 ], "score": 1.0, "content": "tion and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023b.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 340, 506, 354 ], "spans": [ { "bbox": [ 105, 340, 506, 354 ], "score": 1.0, "content": "Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu,", "type": "text" } ], "index": 19 }, { "bbox": [ 115, 350, 505, 365 ], "spans": [ { "bbox": [ 115, 350, 505, 365 ], "score": 1.0, "content": "and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv", "type": "text" } ], "index": 20 }, { "bbox": [ 115, 362, 260, 376 ], "spans": [ { "bbox": [ 115, 362, 260, 376 ], "score": 1.0, "content": "preprint arXiv:2205.14100, 2022a.", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 381, 506, 395 ], "spans": [ { "bbox": [ 105, 381, 506, 395 ], "score": 1.0, "content": "Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou,", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 392, 506, 406 ], "spans": [ { "bbox": [ 115, 392, 506, 406 ], "score": 1.0, "content": "Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 403, 389, 417 ], "spans": [ { "bbox": [ 115, 403, 389, 417 ], "score": 1.0, "content": "simple sequence-to-sequence learning framework. In ICML, 2022b.", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 419, 506, 437 ], "spans": [ { "bbox": [ 105, 419, 506, 437 ], "score": 1.0, "content": "Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong", "type": "text" } ], "index": 25 }, { "bbox": [ 113, 431, 506, 448 ], "spans": [ { "bbox": [ 113, 431, 506, 448 ], "score": 1.0, "content": "Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for", "type": "text" } ], "index": 26 }, { "bbox": [ 116, 443, 362, 457 ], "spans": [ { "bbox": [ 116, 443, 362, 457 ], "score": 1.0, "content": "vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023.", "type": "text" } ], "index": 27 }, { "bbox": [ 108, 464, 504, 475 ], "spans": [ { "bbox": [ 108, 464, 504, 475 ], "score": 1.0, "content": "Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon.", "type": "text" } ], "index": 28 }, { "bbox": [ 115, 474, 506, 488 ], "spans": [ { "bbox": [ 115, 474, 506, 488 ], "score": 1.0, "content": "Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 484, 168, 498 ], "spans": [ { "bbox": [ 115, 484, 168, 498 ], "score": 1.0, "content": "1–12, 2019.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 504, 506, 518 ], "spans": [ { "bbox": [ 105, 504, 506, 518 ], "score": 1.0, "content": "Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 515, 417, 529 ], "spans": [ { "bbox": [ 115, 515, 417, 529 ], "score": 1.0, "content": "visual language model pretraining with weak supervision. In ICLR, 2022c.", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 534, 506, 548 ], "spans": [ { "bbox": [ 105, 534, 506, 548 ], "score": 1.0, "content": "Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 544, 506, 560 ], "spans": [ { "bbox": [ 115, 544, 506, 560 ], "score": 1.0, "content": "Wang. Grit: A generative region-to-text transformer for object understanding. arXiv preprint", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 555, 220, 569 ], "spans": [ { "bbox": [ 115, 555, 220, 569 ], "score": 1.0, "content": "arXiv:2212.00280, 2022.", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 576, 505, 588 ], "spans": [ { "bbox": [ 106, 576, 505, 588 ], "score": 1.0, "content": "Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu,", "type": "text" } ], "index": 36 }, { "bbox": [ 113, 584, 505, 602 ], "spans": [ { "bbox": [ 113, 584, 505, 602 ], "score": 1.0, "content": "and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling.", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 597, 420, 612 ], "spans": [ { "bbox": [ 115, 597, 420, 612 ], "score": 1.0, "content": "In European Conference on Computer Vision, pp. 521–539. Springer, 2022.", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 616, 506, 630 ], "spans": [ { "bbox": [ 106, 616, 506, 630 ], "score": 1.0, "content": "Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen", "type": "text" } ], "index": 39 }, { "bbox": [ 115, 628, 506, 642 ], "spans": [ { "bbox": [ 115, 628, 506, 642 ], "score": 1.0, "content": "Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models", "type": "text" } ], "index": 40 }, { "bbox": [ 115, 638, 361, 652 ], "spans": [ { "bbox": [ 115, 638, 361, 652 ], "score": 1.0, "content": "with multimodality. arXiv preprint arXiv:2304.14178, 2023.", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 657, 506, 671 ], "spans": [ { "bbox": [ 105, 657, 506, 671 ], "score": 1.0, "content": "Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context", "type": "text" } ], "index": 42 }, { "bbox": [ 115, 668, 505, 682 ], "spans": [ { "bbox": [ 115, 668, 505, 682 ], "score": 1.0, "content": "in referring expressions. In Computer Vision–ECCV 2016: 14th European Conference, Amster-", "type": "text" } ], "index": 43 }, { "bbox": [ 116, 680, 504, 692 ], "spans": [ { "bbox": [ 116, 680, 504, 692 ], "score": 1.0, "content": "dam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pp. 69–85. Springer, 2016.", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 699, 506, 711 ], "spans": [ { "bbox": [ 105, 699, 506, 711 ], "score": 1.0, "content": "Licheng Yu, Hao Tan, Mohit Bansal, and Tamara L Berg. A joint speaker-listener-reinforcer model", "type": "text" } ], "index": 45 }, { "bbox": [ 115, 710, 506, 724 ], "spans": [ { "bbox": [ 115, 710, 506, 724 ], "score": 1.0, "content": "for referring expressions. In Proceedings of the IEEE conference on computer vision and pattern", "type": "text" } ], "index": 46 }, { "bbox": [ 115, 720, 257, 734 ], "spans": [ { "bbox": [ 115, 720, 257, 734 ], "score": 1.0, "content": "recognition, pp. 7282–7290, 2017.", "type": "text" } ], "index": 47 } ], "index": 23.5 } ], "page_idx": 12, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "13", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 101, 58, 506, 732 ], "lines": [], "index": 23.5, "bbox_fs": [ 105, 82, 507, 734 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 103, 77, 506, 511 ], "lines": [ { "bbox": [ 106, 83, 505, 95 ], "spans": [ { "bbox": [ 106, 83, 505, 95 ], "score": 1.0, "content": "Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mat-", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 93, 476, 107 ], "spans": [ { "bbox": [ 115, 93, 476, 107 ], "score": 1.0, "content": "tnet: Modular attention network for referring expression comprehension. In CVPR, 2018.", "type": "text" } ], "index": 1 }, { "bbox": [ 104, 111, 506, 126 ], "spans": [ { "bbox": [ 104, 111, 506, 126 ], "score": 1.0, "content": "Lili Yu, Bowen Shi, Ramakanth Pasunuru, Benjamin Muller, Olga Golovneva, Tianlu Wang, Arun", "type": "text" } ], "index": 2 }, { "bbox": [ 115, 123, 506, 137 ], "spans": [ { "bbox": [ 115, 123, 506, 137 ], "score": 1.0, "content": "Babu, Binh Tang, Brian Karrer, Shelly Sheynin, et al. Scaling autoregressive multi-modal models:", "type": "text" } ], "index": 3 }, { "bbox": [ 115, 134, 419, 148 ], "spans": [ { "bbox": [ 115, 134, 419, 148 ], "score": 1.0, "content": "Pretraining and instruction tuning. arXiv preprint arXiv:2309.02591, 2023.", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 153, 506, 167 ], "spans": [ { "bbox": [ 105, 153, 506, 167 ], "score": 1.0, "content": "Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection", "type": "text" } ], "index": 5 }, { "bbox": [ 115, 164, 443, 177 ], "spans": [ { "bbox": [ 115, 164, 443, 177 ], "score": 1.0, "content": "with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023.", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 183, 505, 195 ], "spans": [ { "bbox": [ 106, 183, 505, 195 ], "score": 1.0, "content": "Rowan Zellers, Yonatan Bisk, Ali Farhadi, and Yejin Choi. From recognition to cognition: Visual", "type": "text" } ], "index": 7 }, { "bbox": [ 114, 194, 506, 208 ], "spans": [ { "bbox": [ 114, 194, 506, 208 ], "score": 1.0, "content": "commonsense reasoning. In Proceedings of the IEEE/CVF conference on computer vision and", "type": "text" } ], "index": 8 }, { "bbox": [ 115, 206, 288, 218 ], "spans": [ { "bbox": [ 115, 206, 288, 218 ], "score": 1.0, "content": "pattern recognition, pp. 6720–6731, 2019.", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 223, 506, 237 ], "spans": [ { "bbox": [ 105, 223, 506, 237 ], "score": 1.0, "content": "Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan", "type": "text" } ], "index": 10 }, { "bbox": [ 116, 235, 505, 248 ], "spans": [ { "bbox": [ 116, 235, 505, 248 ], "score": 1.0, "content": "Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-", "type": "text" } ], "index": 11 }, { "bbox": [ 115, 246, 506, 259 ], "spans": [ { "bbox": [ 115, 246, 506, 259 ], "score": 1.0, "content": "language understanding. Advances in Neural Information Processing Systems, 35:36067–36080,", "type": "text" } ], "index": 12 }, { "bbox": [ 116, 257, 141, 269 ], "spans": [ { "bbox": [ 116, 257, 141, 269 ], "score": 1.0, "content": "2022.", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 276, 506, 289 ], "spans": [ { "bbox": [ 105, 276, 506, 289 ], "score": 1.0, "content": "Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and", "type": "text" } ], "index": 14 }, { "bbox": [ 115, 287, 506, 300 ], "spans": [ { "bbox": [ 115, 287, 506, 300 ], "score": 1.0, "content": "Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint", "type": "text" } ], "index": 15 }, { "bbox": [ 115, 298, 218, 310 ], "spans": [ { "bbox": [ 115, 298, 218, 310 ], "score": 1.0, "content": "arXiv:2307.03601, 2023.", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 315, 505, 330 ], "spans": [ { "bbox": [ 106, 315, 505, 330 ], "score": 1.0, "content": "Yang Zhao, Zhijie Lin, Daquan Zhou, Zilong Huang, Jiashi Feng, and Bingyi Kang. Bubogpt:", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 327, 472, 341 ], "spans": [ { "bbox": [ 115, 327, 472, 341 ], "score": 1.0, "content": "Enabling visual grounding in multi-modal llms. arXiv preprint arXiv:2307.08581, 2023.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 347, 505, 359 ], "spans": [ { "bbox": [ 105, 347, 505, 359 ], "score": 1.0, "content": "Luowei Zhou, Yannis Kalantidis, Xinlei Chen, Jason J Corso, and Marcus Rohrbach. Grounded", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 358, 257, 370 ], "spans": [ { "bbox": [ 116, 358, 257, 370 ], "score": 1.0, "content": "video description. In CVPR, 2019.", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 374, 506, 391 ], "spans": [ { "bbox": [ 105, 374, 506, 391 ], "score": 1.0, "content": "Yuanen Zhou, Meng Wang, Daqing Liu, Zhenzhen Hu, and Hanwang Zhang. More grounded image", "type": "text" } ], "index": 21 }, { "bbox": [ 115, 389, 392, 400 ], "spans": [ { "bbox": [ 115, 389, 392, 400 ], "score": 1.0, "content": "captioning by distilling image-text matching model. In CVPR, 2020.", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 406, 505, 420 ], "spans": [ { "bbox": [ 105, 406, 505, 420 ], "score": 1.0, "content": "Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: En-", "type": "text" } ], "index": 23 }, { "bbox": [ 115, 417, 506, 431 ], "spans": [ { "bbox": [ 115, 417, 506, 431 ], "score": 1.0, "content": "hancing vision-language understanding with advanced large language models. arXiv preprint", "type": "text" } ], "index": 24 }, { "bbox": [ 115, 428, 223, 440 ], "spans": [ { "bbox": [ 115, 428, 223, 440 ], "score": 1.0, "content": "arXiv:2304.10592, 2023a.", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 445, 506, 462 ], "spans": [ { "bbox": [ 105, 445, 506, 462 ], "score": 1.0, "content": "Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Young-", "type": "text" } ], "index": 26 }, { "bbox": [ 115, 458, 505, 472 ], "spans": [ { "bbox": [ 115, 458, 505, 472 ], "score": 1.0, "content": "jae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 469, 464, 483 ], "spans": [ { "bbox": [ 115, 469, 464, 483 ], "score": 1.0, "content": "scale corpus of images interleaved with text. arXiv preprint arXiv:2304.06939, 2023b.", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 486, 505, 503 ], "spans": [ { "bbox": [ 105, 486, 505, 503 ], "score": 1.0, "content": "Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Seg-", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 499, 441, 512 ], "spans": [ { "bbox": [ 115, 499, 441, 512 ], "score": 1.0, "content": "ment everything everywhere all at once. arXiv preprint arXiv:2304.06718, 2023.", "type": "text" } ], "index": 30 } ], "index": 15 } ], "page_idx": 13, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 763 ], "spans": [ { "bbox": [ 299, 750, 313, 763 ], "score": 1.0, "content": "14", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 103, 77, 506, 511 ], "lines": [], "index": 15, "bbox_fs": [ 104, 83, 506, 512 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 122, 106, 486, 272 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 80, 502, 102 ], "group_id": 0, "lines": [ { "bbox": [ 105, 79, 504, 93 ], "spans": [ { "bbox": [ 105, 79, 504, 93 ], "score": 1.0, "content": "Table 7: Results on the object hallucination benchmark using the POPE evaluation pipeline (Li et al.,", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 90, 140, 104 ], "spans": [ { "bbox": [ 105, 90, 140, 104 ], "score": 1.0, "content": "2023e).", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 122, 106, 486, 272 ], "group_id": 0, "lines": [ { "bbox": [ 122, 106, 486, 272 ], "spans": [ { "bbox": [ 122, 106, 486, 272 ], "score": 0.985, "html": "
DatasetsMetricsFerret Shikra InstructBLIPMiniGPT4 LLaVAMM-GPTmPLUG-Owl
RandomAccuracy (↑)90.2486.9088.5779.6750.3750.1053.97
Precision (↑)97.7294.4084.0978.2450.1950.0552.07
Recall(↑)83.0079.2695.1382.2099.13100.0099.60
F1 Score (↑)89.7686.1989.2780.1766.6466.7168.39
Yes43.7843.2656.5752.5398.7799.9095.63
PopularAccuracy (↑)84.9083.9782.7769.7349.8750.0050.90
Precision (↑)88.2487.5576.2765.8649.9350.0050.46
Recall(↑)80.5379.2095.1381.9399.27100.0099.40
F1 Score (↑)84.2183.1684.6673.0266.4466.6766.94
Yes45.6345.2362.3762.2099.40100.0098.57
AdversarialAccuracy (↑)82.3683.1072.1065.1749.7050.0050.67
Precision (↑)83.6085.6065.1361.1949.8550.0050.34
Recall(↑)80.5379.6095.1382.9399.07100.0099.33
F1 Score (↑)82.0082.4977.3270.4266.3266.6766.82
Yes48.1846.5073.0367.7799.37100.0098.67
", "type": "table", "image_path": "97fd21bd217498b31db4b97bab1f0f9d496f7754f5164ba95cf83e0f75274862.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 122, 106, 486, 161.33333333333334 ], "spans": [], "index": 2 }, { "bbox": [ 122, 161.33333333333334, 486, 216.66666666666669 ], "spans": [], "index": 3 }, { "bbox": [ 122, 216.66666666666669, 486, 272.0 ], "spans": [], "index": 4 } ] } ], "index": 1.75 }, { "type": "table", "bbox": [ 106, 361, 506, 462 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 296, 507, 352 ], "group_id": 1, "lines": [ { "bbox": [ 106, 297, 507, 309 ], "spans": [ { "bbox": [ 106, 297, 507, 309 ], "score": 1.0, "content": "Table 8: Comparison of Ferret v.s. recent MLLMs integrating spatial awareness. ‘Convention’ refers", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 307, 507, 321 ], "spans": [ { "bbox": [ 105, 307, 507, 321 ], "score": 1.0, "content": "to a comprehensive collection of publicly available data that has been transformed using templates,", "type": "text" } ], "index": 6 }, { "bbox": [ 107, 319, 506, 331 ], "spans": [ { "bbox": [ 107, 319, 506, 331 ], "score": 1.0, "content": "‘GPT-Generate’ signifies the generated refer/ground datasets employing GPT, and ‘Robustness’ de-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 329, 507, 343 ], "spans": [ { "bbox": [ 105, 329, 507, 343 ], "score": 1.0, "content": "notes datasets aimed at mitigating hallucination and improving robustness. Section 3 explains more", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 340, 182, 352 ], "spans": [ { "bbox": [ 106, 340, 182, 352 ], "score": 1.0, "content": "details about each.", "type": "text" } ], "index": 9 } ], "index": 7 }, { "type": "table_body", "bbox": [ 106, 361, 506, 462 ], "group_id": 1, "lines": [ { "bbox": [ 106, 361, 506, 462 ], "spans": [ { "bbox": [ 106, 361, 506, 462 ], "score": 0.979, "html": "
ModelInput TypesGrOutpungData ConstructionQu antaetivr uval. w. Chat
Point Box Free-formConvention GPT-Generate Robustness
BuboGPTxxx√:√×
Vision-LLM
Kosmos-2
Shikraxxνν√xxx√x
GPT4-ROIννννxxν×xxxxxxνxxxxν
PVITXxxxxxxν√
Ferret<
", "type": "table", "image_path": "050a897eac948cd863f25a4797cc3683864b14a7b0a04d8e6edbf91c1b9d6bdf.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 106, 361, 506, 394.6666666666667 ], "spans": [], "index": 10 }, { "bbox": [ 106, 394.6666666666667, 506, 428.33333333333337 ], "spans": [], "index": 11 }, { "bbox": [ 106, 428.33333333333337, 506, 462.00000000000006 ], "spans": [], "index": 12 } ] } ], "index": 9.0 }, { "type": "title", "bbox": [ 108, 489, 214, 502 ], "lines": [ { "bbox": [ 105, 488, 216, 504 ], "spans": [ { "bbox": [ 105, 488, 216, 504 ], "score": 1.0, "content": "A RELATED WORK", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 106, 517, 505, 660 ], "lines": [ { "bbox": [ 105, 516, 506, 531 ], "spans": [ { "bbox": [ 105, 516, 506, 531 ], "score": 1.0, "content": "Multimodal large language models (MLLMs). Large Language Models (LLMs), including", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 527, 505, 541 ], "spans": [ { "bbox": [ 105, 527, 505, 541 ], "score": 1.0, "content": "GPTs (Brown et al., 2020; OpenAI, 2023a), PaLM (Chowdhery et al., 2022), BLOOM (Scao et al.,", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 536, 506, 554 ], "spans": [ { "bbox": [ 104, 536, 506, 554 ], "score": 1.0, "content": "2022), and LLaMA (Touvron et al., 2023a;b), have revolutionized research in NLP, spurring sig-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 549, 505, 563 ], "spans": [ { "bbox": [ 105, 549, 505, 563 ], "score": 1.0, "content": "nificant advances in multimodal language models as well. Early models primarily focused on", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 561, 505, 574 ], "spans": [ { "bbox": [ 105, 561, 505, 574 ], "score": 1.0, "content": "large-scale image-text pre-training. Notable examples include SimVLM (Wang et al., 2022c),", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 572, 506, 585 ], "spans": [ { "bbox": [ 106, 572, 506, 585 ], "score": 1.0, "content": "GIT (Wang et al., 2022a), PaLI (Chen et al., 2022b), PaLI-X (Chen et al., 2023c), BLIP-2 (Li", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 582, 505, 596 ], "spans": [ { "bbox": [ 105, 582, 505, 596 ], "score": 1.0, "content": "et al., 2023c), Flamingo (Alayrac et al., 2022), PaLM-E (Driess et al., 2023), CM3 (Aghajanyan", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 594, 506, 607 ], "spans": [ { "bbox": [ 105, 594, 506, 607 ], "score": 1.0, "content": "et al., 2022), and CM3Leon (Yu et al., 2023). Flamingo, in particular, pioneered the integration", "type": "text" } ], "index": 21 }, { "bbox": [ 104, 604, 506, 619 ], "spans": [ { "bbox": [ 104, 604, 506, 619 ], "score": 1.0, "content": "of a pre-trained CLIP image encoder with LLMs through gated cross-attention blocks, showcasing", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 615, 505, 629 ], "spans": [ { "bbox": [ 106, 615, 505, 629 ], "score": 1.0, "content": "emergent multimodal in-context few-shot learning capabilities. Its open-sourced variants, such as", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 625, 506, 641 ], "spans": [ { "bbox": [ 105, 625, 506, 641 ], "score": 1.0, "content": "OpenFlamingo (Awadalla et al., 2023) and IDEFICS (Laurenc¸on et al., 2023), have garnered sig-", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 637, 506, 650 ], "spans": [ { "bbox": [ 105, 637, 506, 650 ], "score": 1.0, "content": "nificant attention. Typically, these models undergo pre-training using millions or even billions of", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 648, 395, 661 ], "spans": [ { "bbox": [ 105, 648, 395, 661 ], "score": 1.0, "content": "image-text pairs and interleaved image-text datasets (Zhu et al., 2023b).", "type": "text" } ], "index": 26 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 666, 505, 732 ], "lines": [ { "bbox": [ 105, 665, 505, 678 ], "spans": [ { "bbox": [ 105, 665, 505, 678 ], "score": 1.0, "content": "On the other hand, recent research has increasingly focused on using pre-trained LLMs for visual", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 677, 505, 690 ], "spans": [ { "bbox": [ 105, 677, 505, 690 ], "score": 1.0, "content": "instruction tuning. Prominent examples include LLaVA (Liu et al., 2023b), MiniGPT-4 (Zhu et al.,", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 686, 505, 700 ], "spans": [ { "bbox": [ 105, 686, 505, 700 ], "score": 1.0, "content": "2023a), mPLUG-Owl (Ye et al., 2023), Otter (Li et al., 2023a), InstructBLIP (Dai et al., 2023),", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 699, 505, 712 ], "spans": [ { "bbox": [ 105, 699, 505, 712 ], "score": 1.0, "content": "to name a few. In addition to text generation, recent models like FROMAGe (Koh et al., 2023b),", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 709, 506, 722 ], "spans": [ { "bbox": [ 105, 709, 506, 722 ], "score": 1.0, "content": "GILL (Koh et al., 2023a), Emu (Sun et al., 2023), have also enabled MLLMs for image retrieval and", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 721, 446, 733 ], "spans": [ { "bbox": [ 105, 721, 446, 733 ], "score": 1.0, "content": "image generation. Please refer to Chapter 5 of Li et al. (2023b) for a detailed review.", "type": "text" } ], "index": 32 } ], "index": 29.5 } ], "page_idx": 14, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "15", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 122, 106, 486, 272 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 80, 502, 102 ], "group_id": 0, "lines": [ { "bbox": [ 105, 79, 504, 93 ], "spans": [ { "bbox": [ 105, 79, 504, 93 ], "score": 1.0, "content": "Table 7: Results on the object hallucination benchmark using the POPE evaluation pipeline (Li et al.,", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 90, 140, 104 ], "spans": [ { "bbox": [ 105, 90, 140, 104 ], "score": 1.0, "content": "2023e).", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 122, 106, 486, 272 ], "group_id": 0, "lines": [ { "bbox": [ 122, 106, 486, 272 ], "spans": [ { "bbox": [ 122, 106, 486, 272 ], "score": 0.985, "html": "
DatasetsMetricsFerret Shikra InstructBLIPMiniGPT4 LLaVAMM-GPTmPLUG-Owl
RandomAccuracy (↑)90.2486.9088.5779.6750.3750.1053.97
Precision (↑)97.7294.4084.0978.2450.1950.0552.07
Recall(↑)83.0079.2695.1382.2099.13100.0099.60
F1 Score (↑)89.7686.1989.2780.1766.6466.7168.39
Yes43.7843.2656.5752.5398.7799.9095.63
PopularAccuracy (↑)84.9083.9782.7769.7349.8750.0050.90
Precision (↑)88.2487.5576.2765.8649.9350.0050.46
Recall(↑)80.5379.2095.1381.9399.27100.0099.40
F1 Score (↑)84.2183.1684.6673.0266.4466.6766.94
Yes45.6345.2362.3762.2099.40100.0098.57
AdversarialAccuracy (↑)82.3683.1072.1065.1749.7050.0050.67
Precision (↑)83.6085.6065.1361.1949.8550.0050.34
Recall(↑)80.5379.6095.1382.9399.07100.0099.33
F1 Score (↑)82.0082.4977.3270.4266.3266.6766.82
Yes48.1846.5073.0367.7799.37100.0098.67
", "type": "table", "image_path": "97fd21bd217498b31db4b97bab1f0f9d496f7754f5164ba95cf83e0f75274862.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 122, 106, 486, 161.33333333333334 ], "spans": [], "index": 2 }, { "bbox": [ 122, 161.33333333333334, 486, 216.66666666666669 ], "spans": [], "index": 3 }, { "bbox": [ 122, 216.66666666666669, 486, 272.0 ], "spans": [], "index": 4 } ] } ], "index": 1.75 }, { "type": "table", "bbox": [ 106, 361, 506, 462 ], "blocks": [ { "type": "table_caption", "bbox": [ 106, 296, 507, 352 ], "group_id": 1, "lines": [ { "bbox": [ 106, 297, 507, 309 ], "spans": [ { "bbox": [ 106, 297, 507, 309 ], "score": 1.0, "content": "Table 8: Comparison of Ferret v.s. recent MLLMs integrating spatial awareness. ‘Convention’ refers", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 307, 507, 321 ], "spans": [ { "bbox": [ 105, 307, 507, 321 ], "score": 1.0, "content": "to a comprehensive collection of publicly available data that has been transformed using templates,", "type": "text" } ], "index": 6 }, { "bbox": [ 107, 319, 506, 331 ], "spans": [ { "bbox": [ 107, 319, 506, 331 ], "score": 1.0, "content": "‘GPT-Generate’ signifies the generated refer/ground datasets employing GPT, and ‘Robustness’ de-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 329, 507, 343 ], "spans": [ { "bbox": [ 105, 329, 507, 343 ], "score": 1.0, "content": "notes datasets aimed at mitigating hallucination and improving robustness. Section 3 explains more", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 340, 182, 352 ], "spans": [ { "bbox": [ 106, 340, 182, 352 ], "score": 1.0, "content": "details about each.", "type": "text" } ], "index": 9 } ], "index": 7 }, { "type": "table_body", "bbox": [ 106, 361, 506, 462 ], "group_id": 1, "lines": [ { "bbox": [ 106, 361, 506, 462 ], "spans": [ { "bbox": [ 106, 361, 506, 462 ], "score": 0.979, "html": "
ModelInput TypesGrOutpungData ConstructionQu antaetivr uval. w. Chat
Point Box Free-formConvention GPT-Generate Robustness
BuboGPTxxx√:√×
Vision-LLM
Kosmos-2
Shikraxxνν√xxx√x
GPT4-ROIννννxxν×xxxxxxνxxxxν
PVITXxxxxxxν√
Ferret<
", "type": "table", "image_path": "050a897eac948cd863f25a4797cc3683864b14a7b0a04d8e6edbf91c1b9d6bdf.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 106, 361, 506, 394.6666666666667 ], "spans": [], "index": 10 }, { "bbox": [ 106, 394.6666666666667, 506, 428.33333333333337 ], "spans": [], "index": 11 }, { "bbox": [ 106, 428.33333333333337, 506, 462.00000000000006 ], "spans": [], "index": 12 } ] } ], "index": 9.0 }, { "type": "title", "bbox": [ 108, 489, 214, 502 ], "lines": [ { "bbox": [ 105, 488, 216, 504 ], "spans": [ { "bbox": [ 105, 488, 216, 504 ], "score": 1.0, "content": "A RELATED WORK", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 106, 517, 505, 660 ], "lines": [ { "bbox": [ 105, 516, 506, 531 ], "spans": [ { "bbox": [ 105, 516, 506, 531 ], "score": 1.0, "content": "Multimodal large language models (MLLMs). Large Language Models (LLMs), including", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 527, 505, 541 ], "spans": [ { "bbox": [ 105, 527, 505, 541 ], "score": 1.0, "content": "GPTs (Brown et al., 2020; OpenAI, 2023a), PaLM (Chowdhery et al., 2022), BLOOM (Scao et al.,", "type": "text" } ], "index": 15 }, { "bbox": [ 104, 536, 506, 554 ], "spans": [ { "bbox": [ 104, 536, 506, 554 ], "score": 1.0, "content": "2022), and LLaMA (Touvron et al., 2023a;b), have revolutionized research in NLP, spurring sig-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 549, 505, 563 ], "spans": [ { "bbox": [ 105, 549, 505, 563 ], "score": 1.0, "content": "nificant advances in multimodal language models as well. Early models primarily focused on", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 561, 505, 574 ], "spans": [ { "bbox": [ 105, 561, 505, 574 ], "score": 1.0, "content": "large-scale image-text pre-training. Notable examples include SimVLM (Wang et al., 2022c),", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 572, 506, 585 ], "spans": [ { "bbox": [ 106, 572, 506, 585 ], "score": 1.0, "content": "GIT (Wang et al., 2022a), PaLI (Chen et al., 2022b), PaLI-X (Chen et al., 2023c), BLIP-2 (Li", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 582, 505, 596 ], "spans": [ { "bbox": [ 105, 582, 505, 596 ], "score": 1.0, "content": "et al., 2023c), Flamingo (Alayrac et al., 2022), PaLM-E (Driess et al., 2023), CM3 (Aghajanyan", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 594, 506, 607 ], "spans": [ { "bbox": [ 105, 594, 506, 607 ], "score": 1.0, "content": "et al., 2022), and CM3Leon (Yu et al., 2023). Flamingo, in particular, pioneered the integration", "type": "text" } ], "index": 21 }, { "bbox": [ 104, 604, 506, 619 ], "spans": [ { "bbox": [ 104, 604, 506, 619 ], "score": 1.0, "content": "of a pre-trained CLIP image encoder with LLMs through gated cross-attention blocks, showcasing", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 615, 505, 629 ], "spans": [ { "bbox": [ 106, 615, 505, 629 ], "score": 1.0, "content": "emergent multimodal in-context few-shot learning capabilities. Its open-sourced variants, such as", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 625, 506, 641 ], "spans": [ { "bbox": [ 105, 625, 506, 641 ], "score": 1.0, "content": "OpenFlamingo (Awadalla et al., 2023) and IDEFICS (Laurenc¸on et al., 2023), have garnered sig-", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 637, 506, 650 ], "spans": [ { "bbox": [ 105, 637, 506, 650 ], "score": 1.0, "content": "nificant attention. Typically, these models undergo pre-training using millions or even billions of", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 648, 395, 661 ], "spans": [ { "bbox": [ 105, 648, 395, 661 ], "score": 1.0, "content": "image-text pairs and interleaved image-text datasets (Zhu et al., 2023b).", "type": "text" } ], "index": 26 } ], "index": 20, "bbox_fs": [ 104, 516, 506, 661 ] }, { "type": "text", "bbox": [ 107, 666, 505, 732 ], "lines": [ { "bbox": [ 105, 665, 505, 678 ], "spans": [ { "bbox": [ 105, 665, 505, 678 ], "score": 1.0, "content": "On the other hand, recent research has increasingly focused on using pre-trained LLMs for visual", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 677, 505, 690 ], "spans": [ { "bbox": [ 105, 677, 505, 690 ], "score": 1.0, "content": "instruction tuning. Prominent examples include LLaVA (Liu et al., 2023b), MiniGPT-4 (Zhu et al.,", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 686, 505, 700 ], "spans": [ { "bbox": [ 105, 686, 505, 700 ], "score": 1.0, "content": "2023a), mPLUG-Owl (Ye et al., 2023), Otter (Li et al., 2023a), InstructBLIP (Dai et al., 2023),", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 699, 505, 712 ], "spans": [ { "bbox": [ 105, 699, 505, 712 ], "score": 1.0, "content": "to name a few. In addition to text generation, recent models like FROMAGe (Koh et al., 2023b),", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 709, 506, 722 ], "spans": [ { "bbox": [ 105, 709, 506, 722 ], "score": 1.0, "content": "GILL (Koh et al., 2023a), Emu (Sun et al., 2023), have also enabled MLLMs for image retrieval and", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 721, 446, 733 ], "spans": [ { "bbox": [ 105, 721, 446, 733 ], "score": 1.0, "content": "image generation. Please refer to Chapter 5 of Li et al. (2023b) for a detailed review.", "type": "text" } ], "index": 32 } ], "index": 29.5, "bbox_fs": [ 105, 665, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 225 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "MLLMs for referring and grounding. In the realm of existing literature, works such as Kosmos-", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 94, 505, 106 ], "spans": [ { "bbox": [ 105, 94, 505, 106 ], "score": 1.0, "content": "2 (Peng et al., 2023) and Shikra (Chen et al., 2023b), closely resemble ours as they also enable", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 505, 117 ], "spans": [ { "bbox": [ 105, 104, 505, 117 ], "score": 1.0, "content": "MLLMs for fine-grained image comprehension and open-world referring and grounding. Addi-", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 505, 128 ], "spans": [ { "bbox": [ 105, 115, 505, 128 ], "score": 1.0, "content": "tional works in this direction include GPT4ROI (Zhang et al., 2023), PVIT (Chen et al., 2023a),", "type": "text" } ], "index": 3 }, { "bbox": [ 104, 125, 506, 140 ], "spans": [ { "bbox": [ 104, 125, 506, 140 ], "score": 1.0, "content": "BuboGPT (Zhao et al., 2023), VisionLLM (Wang et al., 2023), and ContextDET (Zang et al., 2023).", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 136, 505, 151 ], "spans": [ { "bbox": [ 105, 136, 505, 151 ], "score": 1.0, "content": "Nevertheless, pivotal distinctions set our model apart. First, prior endeavors supported only bound-", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 147, 506, 163 ], "spans": [ { "bbox": [ 105, 147, 506, 163 ], "score": 1.0, "content": "ing boxes (and points in Shikra) as input. Conversely, due to Ferret’s innovative hybrid region repre-", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 158, 505, 173 ], "spans": [ { "bbox": [ 105, 158, 505, 173 ], "score": 1.0, "content": "sentation, we accommodate a broader range of free-form shapes for referring, encompassing points,", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 170, 506, 184 ], "spans": [ { "bbox": [ 105, 170, 506, 184 ], "score": 1.0, "content": "boxes, sketches, scribbles, polygons, and more. Second, we meticulously curate an extensive refer-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 180, 505, 195 ], "spans": [ { "bbox": [ 105, 180, 505, 195 ], "score": 1.0, "content": "and-ground instruction tuning dataset. Third, we introduce Ferret-Bench to facilitate forthcoming", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 192, 505, 205 ], "spans": [ { "bbox": [ 105, 192, 505, 205 ], "score": 1.0, "content": "research and enhance evaluation benchmarks in this direction. Lastly, our model exhibits superior", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 203, 505, 216 ], "spans": [ { "bbox": [ 105, 203, 505, 216 ], "score": 1.0, "content": "performance compared to previous works, notably mitigating object hallucination to a significant", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 214, 411, 227 ], "spans": [ { "bbox": [ 105, 214, 411, 227 ], "score": 1.0, "content": "extent. A more straightforward side-by-side comparison is shown in Tab. 8.", "type": "text" } ], "index": 12 } ], "index": 6 }, { "type": "text", "bbox": [ 107, 231, 505, 319 ], "lines": [ { "bbox": [ 105, 230, 505, 244 ], "spans": [ { "bbox": [ 105, 230, 505, 244 ], "score": 1.0, "content": "Unifying grounding and VL understanding. Our work is also related to previous work that aims", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 241, 505, 255 ], "spans": [ { "bbox": [ 105, 241, 505, 255 ], "score": 1.0, "content": "to unify text and bounding box output for vision-language (VL) models, such as UniTAB (Yang", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 252, 506, 266 ], "spans": [ { "bbox": [ 105, 252, 506, 266 ], "score": 1.0, "content": "et al., 2022), OFA (Wang et al., 2022b), and Unified-IO (Lu et al., 2022), which also represent", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 263, 505, 277 ], "spans": [ { "bbox": [ 105, 263, 505, 277 ], "score": 1.0, "content": "bounding boxes using a set of additional discrete tokens as proposed in Pix2Seq (Chen et al., 2021;", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 274, 505, 288 ], "spans": [ { "bbox": [ 105, 274, 505, 288 ], "score": 1.0, "content": "2022a). Ferret is unique in that (i) our model is built upon LLMs, marrying the power of LLMs and", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 286, 505, 298 ], "spans": [ { "bbox": [ 105, 286, 505, 298 ], "score": 1.0, "content": "grounding, thus unlocking new capabilities such as grounded instruction tuning, and (ii) we handle", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 297, 506, 310 ], "spans": [ { "bbox": [ 105, 297, 506, 310 ], "score": 1.0, "content": "bounding box coordinates as regular text tokens, avoiding the need for extra specialized tokens", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 308, 237, 320 ], "spans": [ { "bbox": [ 105, 308, 237, 320 ], "score": 1.0, "content": "dedicated to representing boxes.", "type": "text" } ], "index": 20 } ], "index": 16.5 }, { "type": "title", "bbox": [ 107, 339, 385, 352 ], "lines": [ { "bbox": [ 105, 339, 387, 353 ], "spans": [ { "bbox": [ 105, 339, 387, 353 ], "score": 1.0, "content": "B DISCUSSION ON LIMITATION AND FAILURE CASES", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 108, 367, 504, 389 ], "lines": [ { "bbox": [ 106, 367, 505, 380 ], "spans": [ { "bbox": [ 106, 367, 505, 380 ], "score": 1.0, "content": "We acknowledge certain specific failure scenarios and limitations for our models, which are detailed", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 378, 153, 390 ], "spans": [ { "bbox": [ 105, 378, 153, 390 ], "score": 1.0, "content": "as follows:", "type": "text" } ], "index": 23 } ], "index": 22.5 }, { "type": "text", "bbox": [ 107, 389, 504, 444 ], "lines": [ { "bbox": [ 105, 388, 505, 402 ], "spans": [ { "bbox": [ 105, 388, 505, 402 ], "score": 1.0, "content": "Failure Scenarios: (1). Referring to too many objects (more than 3) in one question might not be", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 400, 506, 413 ], "spans": [ { "bbox": [ 105, 400, 506, 413 ], "score": 1.0, "content": "as accurate as referring to each of them in separate conversations. This is likely due to a relative", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 410, 506, 425 ], "spans": [ { "bbox": [ 105, 410, 506, 425 ], "score": 1.0, "content": "scarcity of training data that mentions too many objects. (2). The referring and grounding of very", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 420, 506, 436 ], "spans": [ { "bbox": [ 105, 420, 506, 436 ], "score": 1.0, "content": "small objects is less accurate than large or medium objects. It’s a common challenge in object", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 432, 452, 447 ], "spans": [ { "bbox": [ 105, 432, 452, 447 ], "score": 1.0, "content": "detection. However, we think further improving input image resolution is able to help.", "type": "text" } ], "index": 28 } ], "index": 26 }, { "type": "text", "bbox": [ 107, 445, 505, 510 ], "lines": [ { "bbox": [ 105, 444, 504, 456 ], "spans": [ { "bbox": [ 105, 444, 504, 456 ], "score": 1.0, "content": "Limitations: (1). Not good at other languages because the training dataset is curated only in English.", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 455, 505, 468 ], "spans": [ { "bbox": [ 106, 455, 505, 468 ], "score": 1.0, "content": "Although Ferret shows some emergent referring and grounding capability in other languages, its", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 466, 505, 479 ], "spans": [ { "bbox": [ 105, 466, 505, 479 ], "score": 1.0, "content": "performance in other languages is still worse than in English. Future incorporation of multilingual", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 477, 504, 490 ], "spans": [ { "bbox": [ 105, 477, 504, 490 ], "score": 1.0, "content": "training data could potentially mitigate this. (2). Similar to many large language models, Ferret has", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 488, 505, 501 ], "spans": [ { "bbox": [ 105, 488, 505, 501 ], "score": 1.0, "content": "the potential to generate harmful or factually incorrect responses. (3). Ferret is not designed for", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 499, 281, 512 ], "spans": [ { "bbox": [ 105, 499, 281, 512 ], "score": 1.0, "content": "segmentation tasks requiring mask outputs.", "type": "text" } ], "index": 34 } ], "index": 31.5 }, { "type": "title", "bbox": [ 108, 531, 239, 543 ], "lines": [ { "bbox": [ 106, 530, 241, 546 ], "spans": [ { "bbox": [ 106, 530, 241, 546 ], "score": 1.0, "content": "C DETAILS OF DATASET", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "title", "bbox": [ 108, 558, 310, 570 ], "lines": [ { "bbox": [ 106, 558, 312, 571 ], "spans": [ { "bbox": [ 106, 558, 312, 571 ], "score": 1.0, "content": "C.1 TASK TEMPLATES FOR PUBLIC DATASETS", "type": "text" } ], "index": 36 } ], "index": 36 }, { "type": "text", "bbox": [ 107, 581, 505, 614 ], "lines": [ { "bbox": [ 105, 581, 505, 593 ], "spans": [ { "bbox": [ 105, 581, 505, 593 ], "score": 1.0, "content": "In Section 3.1, we mentioned using carefully designed task templates to convert public datasets such", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 593, 505, 604 ], "spans": [ { "bbox": [ 106, 593, 505, 604 ], "score": 1.0, "content": "as Visual Genome into instruction-following format. The task templates we used are provided in", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 603, 369, 615 ], "spans": [ { "bbox": [ 106, 603, 369, 615 ], "score": 1.0, "content": "Table 9. For simplicity, we only list three examples for each task.", "type": "text" } ], "index": 39 } ], "index": 38 }, { "type": "title", "bbox": [ 107, 632, 308, 644 ], "lines": [ { "bbox": [ 106, 631, 309, 645 ], "spans": [ { "bbox": [ 106, 631, 309, 645 ], "score": 1.0, "content": "C.2 DETAILS ON SPATIAL NEGATIVE MINING", "type": "text" } ], "index": 40 } ], "index": 40 }, { "type": "text", "bbox": [ 107, 654, 504, 732 ], "lines": [ { "bbox": [ 105, 654, 506, 668 ], "spans": [ { "bbox": [ 105, 654, 506, 668 ], "score": 1.0, "content": "In Section 3.3, we conducted negative sample mining for two aspects: (i) Image-conditioned Cate-", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 666, 505, 678 ], "spans": [ { "bbox": [ 105, 666, 505, 678 ], "score": 1.0, "content": "gory Localization, and (ii) Semantics-conditioned Category Localization. They use the same tem-", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 677, 505, 689 ], "spans": [ { "bbox": [ 105, 677, 505, 689 ], "score": 1.0, "content": "plate to convert the original data, which falls into the task of object hallucination in Table 9. Specif-", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 255, 700 ], "score": 1.0, "content": "ically, for the negative categories in", "type": "text" }, { "bbox": [ 256, 688, 270, 699 ], "score": 0.46, "content": "( i i )", "type": "inline_equation" }, { "bbox": [ 271, 687, 506, 700 ], "score": 1.0, "content": ", we prompt ChatGPT/GPT-4 to generate entities that are", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "most analogous to the original class, attribute, or quantity, e.g., ‘man’ vs. ‘woman’, ‘blue’ vs. ‘yel-", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 709, 506, 722 ], "spans": [ { "bbox": [ 105, 709, 506, 722 ], "score": 1.0, "content": "low’, ‘two’ vs. ‘three’. The prompt feed into ChatGPT/GPT-4 encompasses all the entities extracted", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 720, 503, 733 ], "spans": [ { "bbox": [ 105, 720, 503, 733 ], "score": 1.0, "content": "from 5 captions associated with one single image. We show the exact prompt template in Table 10.", "type": "text" } ], "index": 47 } ], "index": 44 } ], "page_idx": 15, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 505, 225 ], "lines": [ { "bbox": [ 105, 82, 505, 95 ], "spans": [ { "bbox": [ 105, 82, 505, 95 ], "score": 1.0, "content": "MLLMs for referring and grounding. In the realm of existing literature, works such as Kosmos-", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 94, 505, 106 ], "spans": [ { "bbox": [ 105, 94, 505, 106 ], "score": 1.0, "content": "2 (Peng et al., 2023) and Shikra (Chen et al., 2023b), closely resemble ours as they also enable", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 104, 505, 117 ], "spans": [ { "bbox": [ 105, 104, 505, 117 ], "score": 1.0, "content": "MLLMs for fine-grained image comprehension and open-world referring and grounding. Addi-", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 115, 505, 128 ], "spans": [ { "bbox": [ 105, 115, 505, 128 ], "score": 1.0, "content": "tional works in this direction include GPT4ROI (Zhang et al., 2023), PVIT (Chen et al., 2023a),", "type": "text" } ], "index": 3 }, { "bbox": [ 104, 125, 506, 140 ], "spans": [ { "bbox": [ 104, 125, 506, 140 ], "score": 1.0, "content": "BuboGPT (Zhao et al., 2023), VisionLLM (Wang et al., 2023), and ContextDET (Zang et al., 2023).", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 136, 505, 151 ], "spans": [ { "bbox": [ 105, 136, 505, 151 ], "score": 1.0, "content": "Nevertheless, pivotal distinctions set our model apart. First, prior endeavors supported only bound-", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 147, 506, 163 ], "spans": [ { "bbox": [ 105, 147, 506, 163 ], "score": 1.0, "content": "ing boxes (and points in Shikra) as input. Conversely, due to Ferret’s innovative hybrid region repre-", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 158, 505, 173 ], "spans": [ { "bbox": [ 105, 158, 505, 173 ], "score": 1.0, "content": "sentation, we accommodate a broader range of free-form shapes for referring, encompassing points,", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 170, 506, 184 ], "spans": [ { "bbox": [ 105, 170, 506, 184 ], "score": 1.0, "content": "boxes, sketches, scribbles, polygons, and more. Second, we meticulously curate an extensive refer-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 180, 505, 195 ], "spans": [ { "bbox": [ 105, 180, 505, 195 ], "score": 1.0, "content": "and-ground instruction tuning dataset. Third, we introduce Ferret-Bench to facilitate forthcoming", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 192, 505, 205 ], "spans": [ { "bbox": [ 105, 192, 505, 205 ], "score": 1.0, "content": "research and enhance evaluation benchmarks in this direction. Lastly, our model exhibits superior", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 203, 505, 216 ], "spans": [ { "bbox": [ 105, 203, 505, 216 ], "score": 1.0, "content": "performance compared to previous works, notably mitigating object hallucination to a significant", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 214, 411, 227 ], "spans": [ { "bbox": [ 105, 214, 411, 227 ], "score": 1.0, "content": "extent. A more straightforward side-by-side comparison is shown in Tab. 8.", "type": "text" } ], "index": 12 } ], "index": 6, "bbox_fs": [ 104, 82, 506, 227 ] }, { "type": "text", "bbox": [ 107, 231, 505, 319 ], "lines": [ { "bbox": [ 105, 230, 505, 244 ], "spans": [ { "bbox": [ 105, 230, 505, 244 ], "score": 1.0, "content": "Unifying grounding and VL understanding. Our work is also related to previous work that aims", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 241, 505, 255 ], "spans": [ { "bbox": [ 105, 241, 505, 255 ], "score": 1.0, "content": "to unify text and bounding box output for vision-language (VL) models, such as UniTAB (Yang", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 252, 506, 266 ], "spans": [ { "bbox": [ 105, 252, 506, 266 ], "score": 1.0, "content": "et al., 2022), OFA (Wang et al., 2022b), and Unified-IO (Lu et al., 2022), which also represent", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 263, 505, 277 ], "spans": [ { "bbox": [ 105, 263, 505, 277 ], "score": 1.0, "content": "bounding boxes using a set of additional discrete tokens as proposed in Pix2Seq (Chen et al., 2021;", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 274, 505, 288 ], "spans": [ { "bbox": [ 105, 274, 505, 288 ], "score": 1.0, "content": "2022a). Ferret is unique in that (i) our model is built upon LLMs, marrying the power of LLMs and", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 286, 505, 298 ], "spans": [ { "bbox": [ 105, 286, 505, 298 ], "score": 1.0, "content": "grounding, thus unlocking new capabilities such as grounded instruction tuning, and (ii) we handle", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 297, 506, 310 ], "spans": [ { "bbox": [ 105, 297, 506, 310 ], "score": 1.0, "content": "bounding box coordinates as regular text tokens, avoiding the need for extra specialized tokens", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 308, 237, 320 ], "spans": [ { "bbox": [ 105, 308, 237, 320 ], "score": 1.0, "content": "dedicated to representing boxes.", "type": "text" } ], "index": 20 } ], "index": 16.5, "bbox_fs": [ 105, 230, 506, 320 ] }, { "type": "title", "bbox": [ 107, 339, 385, 352 ], "lines": [ { "bbox": [ 105, 339, 387, 353 ], "spans": [ { "bbox": [ 105, 339, 387, 353 ], "score": 1.0, "content": "B DISCUSSION ON LIMITATION AND FAILURE CASES", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 108, 367, 504, 389 ], "lines": [ { "bbox": [ 106, 367, 505, 380 ], "spans": [ { "bbox": [ 106, 367, 505, 380 ], "score": 1.0, "content": "We acknowledge certain specific failure scenarios and limitations for our models, which are detailed", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 378, 153, 390 ], "spans": [ { "bbox": [ 105, 378, 153, 390 ], "score": 1.0, "content": "as follows:", "type": "text" } ], "index": 23 } ], "index": 22.5, "bbox_fs": [ 105, 367, 505, 390 ] }, { "type": "text", "bbox": [ 107, 389, 504, 444 ], "lines": [ { "bbox": [ 105, 388, 505, 402 ], "spans": [ { "bbox": [ 105, 388, 505, 402 ], "score": 1.0, "content": "Failure Scenarios: (1). Referring to too many objects (more than 3) in one question might not be", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 400, 506, 413 ], "spans": [ { "bbox": [ 105, 400, 506, 413 ], "score": 1.0, "content": "as accurate as referring to each of them in separate conversations. This is likely due to a relative", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 410, 506, 425 ], "spans": [ { "bbox": [ 105, 410, 506, 425 ], "score": 1.0, "content": "scarcity of training data that mentions too many objects. (2). The referring and grounding of very", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 420, 506, 436 ], "spans": [ { "bbox": [ 105, 420, 506, 436 ], "score": 1.0, "content": "small objects is less accurate than large or medium objects. It’s a common challenge in object", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 432, 452, 447 ], "spans": [ { "bbox": [ 105, 432, 452, 447 ], "score": 1.0, "content": "detection. However, we think further improving input image resolution is able to help.", "type": "text" } ], "index": 28 } ], "index": 26, "bbox_fs": [ 105, 388, 506, 447 ] }, { "type": "text", "bbox": [ 107, 445, 505, 510 ], "lines": [ { "bbox": [ 105, 444, 504, 456 ], "spans": [ { "bbox": [ 105, 444, 504, 456 ], "score": 1.0, "content": "Limitations: (1). Not good at other languages because the training dataset is curated only in English.", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 455, 505, 468 ], "spans": [ { "bbox": [ 106, 455, 505, 468 ], "score": 1.0, "content": "Although Ferret shows some emergent referring and grounding capability in other languages, its", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 466, 505, 479 ], "spans": [ { "bbox": [ 105, 466, 505, 479 ], "score": 1.0, "content": "performance in other languages is still worse than in English. Future incorporation of multilingual", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 477, 504, 490 ], "spans": [ { "bbox": [ 105, 477, 504, 490 ], "score": 1.0, "content": "training data could potentially mitigate this. (2). Similar to many large language models, Ferret has", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 488, 505, 501 ], "spans": [ { "bbox": [ 105, 488, 505, 501 ], "score": 1.0, "content": "the potential to generate harmful or factually incorrect responses. (3). Ferret is not designed for", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 499, 281, 512 ], "spans": [ { "bbox": [ 105, 499, 281, 512 ], "score": 1.0, "content": "segmentation tasks requiring mask outputs.", "type": "text" } ], "index": 34 } ], "index": 31.5, "bbox_fs": [ 105, 444, 505, 512 ] }, { "type": "title", "bbox": [ 108, 531, 239, 543 ], "lines": [ { "bbox": [ 106, 530, 241, 546 ], "spans": [ { "bbox": [ 106, 530, 241, 546 ], "score": 1.0, "content": "C DETAILS OF DATASET", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "title", "bbox": [ 108, 558, 310, 570 ], "lines": [ { "bbox": [ 106, 558, 312, 571 ], "spans": [ { "bbox": [ 106, 558, 312, 571 ], "score": 1.0, "content": "C.1 TASK TEMPLATES FOR PUBLIC DATASETS", "type": "text" } ], "index": 36 } ], "index": 36 }, { "type": "text", "bbox": [ 107, 581, 505, 614 ], "lines": [ { "bbox": [ 105, 581, 505, 593 ], "spans": [ { "bbox": [ 105, 581, 505, 593 ], "score": 1.0, "content": "In Section 3.1, we mentioned using carefully designed task templates to convert public datasets such", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 593, 505, 604 ], "spans": [ { "bbox": [ 106, 593, 505, 604 ], "score": 1.0, "content": "as Visual Genome into instruction-following format. The task templates we used are provided in", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 603, 369, 615 ], "spans": [ { "bbox": [ 106, 603, 369, 615 ], "score": 1.0, "content": "Table 9. For simplicity, we only list three examples for each task.", "type": "text" } ], "index": 39 } ], "index": 38, "bbox_fs": [ 105, 581, 505, 615 ] }, { "type": "title", "bbox": [ 107, 632, 308, 644 ], "lines": [ { "bbox": [ 106, 631, 309, 645 ], "spans": [ { "bbox": [ 106, 631, 309, 645 ], "score": 1.0, "content": "C.2 DETAILS ON SPATIAL NEGATIVE MINING", "type": "text" } ], "index": 40 } ], "index": 40 }, { "type": "text", "bbox": [ 107, 654, 504, 732 ], "lines": [ { "bbox": [ 105, 654, 506, 668 ], "spans": [ { "bbox": [ 105, 654, 506, 668 ], "score": 1.0, "content": "In Section 3.3, we conducted negative sample mining for two aspects: (i) Image-conditioned Cate-", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 666, 505, 678 ], "spans": [ { "bbox": [ 105, 666, 505, 678 ], "score": 1.0, "content": "gory Localization, and (ii) Semantics-conditioned Category Localization. They use the same tem-", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 677, 505, 689 ], "spans": [ { "bbox": [ 105, 677, 505, 689 ], "score": 1.0, "content": "plate to convert the original data, which falls into the task of object hallucination in Table 9. Specif-", "type": "text" } ], "index": 43 }, { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 255, 700 ], "score": 1.0, "content": "ically, for the negative categories in", "type": "text" }, { "bbox": [ 256, 688, 270, 699 ], "score": 0.46, "content": "( i i )", "type": "inline_equation" }, { "bbox": [ 271, 687, 506, 700 ], "score": 1.0, "content": ", we prompt ChatGPT/GPT-4 to generate entities that are", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "most analogous to the original class, attribute, or quantity, e.g., ‘man’ vs. ‘woman’, ‘blue’ vs. ‘yel-", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 709, 506, 722 ], "spans": [ { "bbox": [ 105, 709, 506, 722 ], "score": 1.0, "content": "low’, ‘two’ vs. ‘three’. The prompt feed into ChatGPT/GPT-4 encompasses all the entities extracted", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 720, 503, 733 ], "spans": [ { "bbox": [ 105, 720, 503, 733 ], "score": 1.0, "content": "from 5 captions associated with one single image. We show the exact prompt template in Table 10.", "type": "text" } ], "index": 47 } ], "index": 44, "bbox_fs": [ 105, 654, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 109, 112, 501, 347 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 80, 504, 102 ], "group_id": 0, "lines": [ { "bbox": [ 105, 79, 505, 93 ], "spans": [ { "bbox": [ 105, 79, 505, 93 ], "score": 1.0, "content": "Table 9: Examples of task templates Ferret used to transfer different public data types into the", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 92, 225, 103 ], "spans": [ { "bbox": [ 106, 92, 225, 103 ], "score": 1.0, "content": "instruction-following format.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 109, 112, 501, 347 ], "group_id": 0, "lines": [ { "bbox": [ 109, 112, 501, 347 ], "spans": [ { "bbox": [ 109, 112, 501, 347 ], "score": 0.984, "html": "
TaskThree randomly chosen examples from many.
Referring-ObjectWhat is the class of the object <location>within the image? Classify object <location> in the image. Identify the object <location> in the image.
Referring-RelationWhat does <object1><location1> do to <object2><location2> of the image? What is the physical relation between <object1><locationl>and <object2><location2>? Can you figure out the geometric relation of the <objectl><locationl>and <object2><location2>?
Referring-RegionDescribe the region <location> in a short phrase. What is in the region <location>? Describe in a phrase. Capture in a phrase:what's near region <location>in the picture?
REC.Where is<object>in the image? What are the coordinates for the given <object> in the image? Given the image,could you please tell me where is <object>
Phrase GroundingWhat are the locations of <objects>? Could you provide me with the exact locations of <objects>? Please indicate the positions of <objects> in the image?
Object Detection (O365)Detect all objects among <class> in the image. Perform object detection given the image within <class>. Given the image and set <class>,identify allthe objects that belong to the set.
Grounded CaptioningWhat is this photo about? Use concise language. Describe the overall picture in just a few words. What do you see happening in this image? Provide the answer in short.
Object HallucinationIs there a <object>in the image? Are there <object>in the image? Please tell me whether<object> exists in the image?
", "type": "table", "image_path": "5320f095ac9a8244a5928ae20f13d9a21d932e723e97adb30cf9e843722f60e4.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 109, 112, 501, 190.33333333333331 ], "spans": [], "index": 2 }, { "bbox": [ 109, 190.33333333333331, 501, 268.66666666666663 ], "spans": [], "index": 3 }, { "bbox": [ 109, 268.66666666666663, 501, 346.99999999999994 ], "spans": [], "index": 4 } ] } ], "index": 1.75 }, { "type": "text", "bbox": [ 132, 373, 477, 385 ], "lines": [ { "bbox": [ 132, 371, 479, 388 ], "spans": [ { "bbox": [ 132, 371, 479, 388 ], "score": 1.0, "content": "Table 10: In this example, we provide the prompt to generate the spatial negative sets.", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 119, 404, 478, 434 ], "lines": [ { "bbox": [ 116, 402, 480, 417 ], "spans": [ { "bbox": [ 116, 402, 163, 417 ], "score": 1.0, "content": "messages", "type": "text" }, { "bbox": [ 163, 405, 172, 413 ], "score": 0.53, "content": "=", "type": "inline_equation" }, { "bbox": [ 172, 402, 480, 417 ], "score": 1.0, "content": "[ {\"role\":\"system\", \"content\": f”’You are an AI visual assistant that can", "type": "text" } ], "index": 6 }, { "bbox": [ 117, 413, 479, 426 ], "spans": [ { "bbox": [ 117, 413, 479, 426 ], "score": 1.0, "content": "analyze a single image. You receive several entities given by a list, each describing the objects in", "type": "text" } ], "index": 7 }, { "bbox": [ 118, 423, 225, 436 ], "spans": [ { "bbox": [ 118, 423, 225, 436 ], "score": 1.0, "content": "the image you are observing.", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "text", "bbox": [ 119, 444, 478, 484 ], "lines": [ { "bbox": [ 117, 442, 480, 457 ], "spans": [ { "bbox": [ 117, 442, 480, 457 ], "score": 1.0, "content": "For each entity mentioned, change them with the most misleading entity name (may belong to", "type": "text" } ], "index": 9 }, { "bbox": [ 116, 452, 480, 466 ], "spans": [ { "bbox": [ 116, 452, 387, 466 ], "score": 1.0, "content": "the same category but are actually different) (nonexistent objects: man", "type": "text" }, { "bbox": [ 387, 455, 399, 463 ], "score": 0.81, "content": "", "type": "inline_equation" }, { "bbox": [ 400, 452, 480, 466 ], "score": 1.0, "content": "woman, nonexistent", "type": "text" } ], "index": 10 }, { "bbox": [ 117, 463, 479, 475 ], "spans": [ { "bbox": [ 117, 463, 186, 475 ], "score": 1.0, "content": "attributes: brown", "type": "text" }, { "bbox": [ 187, 465, 199, 473 ], "score": 0.81, "content": "", "type": "inline_equation" }, { "bbox": [ 199, 463, 333, 475 ], "score": 1.0, "content": "yellow, nonexistent quantities: two", "type": "text" }, { "bbox": [ 334, 465, 345, 473 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 346, 463, 479, 475 ], "score": 1.0, "content": "three, etc.). The instructions should", "type": "text" } ], "index": 11 }, { "bbox": [ 117, 474, 291, 486 ], "spans": [ { "bbox": [ 117, 474, 291, 486 ], "score": 1.0, "content": "contain interrogative and declarative sentences.", "type": "text" } ], "index": 12 } ], "index": 10.5 }, { "type": "text", "bbox": [ 118, 493, 477, 514 ], "lines": [ { "bbox": [ 118, 493, 478, 505 ], "spans": [ { "bbox": [ 118, 493, 478, 505 ], "score": 1.0, "content": "The output format needs to be a list only which contains the misleading entity names. Please follow", "type": "text" } ], "index": 13 }, { "bbox": [ 117, 502, 212, 515 ], "spans": [ { "bbox": [ 117, 502, 212, 515 ], "score": 1.0, "content": "the instructions carefully.", "type": "text" } ], "index": 14 } ], "index": 13.5 }, { "type": "text", "bbox": [ 118, 523, 380, 534 ], "lines": [ { "bbox": [ 117, 523, 381, 535 ], "spans": [ { "bbox": [ 117, 523, 381, 535 ], "score": 1.0, "content": "1. The length of the output list needs to be exactly equal to the input list.", "type": "text" } ], "index": 15 } ], "index": 15 }, { "type": "text", "bbox": [ 119, 543, 227, 554 ], "lines": [ { "bbox": [ 118, 543, 228, 554 ], "spans": [ { "bbox": [ 118, 543, 228, 554 ], "score": 1.0, "content": "2. Do not explain the reasons.", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "text", "bbox": [ 115, 563, 473, 574 ], "lines": [ { "bbox": [ 115, 563, 474, 575 ], "spans": [ { "bbox": [ 115, 563, 474, 575 ], "score": 1.0, "content": "3. Do not mention the input entities, at least the output name and input name needs to be different.", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 119, 583, 295, 594 ], "lines": [ { "bbox": [ 118, 582, 297, 594 ], "spans": [ { "bbox": [ 118, 582, 297, 594 ], "score": 1.0, "content": "4. Do not mention something abstract, like alien ¨ .", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 118, 603, 438, 614 ], "lines": [ { "bbox": [ 117, 601, 439, 615 ], "spans": [ { "bbox": [ 117, 601, 439, 615 ], "score": 1.0, "content": "5. When dealing with quantities, focus solely on increasing the numbers during revision.", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 118, 623, 478, 644 ], "lines": [ { "bbox": [ 116, 621, 480, 637 ], "spans": [ { "bbox": [ 116, 621, 480, 637 ], "score": 1.0, "content": "6. When dealing with words like ”a few”, ”a group”, ”several”, ”some”, etc., try changing the", "type": "text" } ], "index": 20 }, { "bbox": [ 117, 631, 261, 645 ], "spans": [ { "bbox": [ 117, 631, 190, 645 ], "score": 1.0, "content": "objects (A few men", "type": "text" }, { "bbox": [ 190, 633, 210, 642 ], "score": 0.84, "content": " \\mathbf { A }", "type": "inline_equation" }, { "bbox": [ 211, 631, 261, 645 ], "score": 1.0, "content": "few women).", "type": "text" } ], "index": 21 } ], "index": 20.5 }, { "type": "text", "bbox": [ 118, 653, 478, 684 ], "lines": [ { "bbox": [ 117, 652, 479, 665 ], "spans": [ { "bbox": [ 117, 652, 479, 665 ], "score": 1.0, "content": "7. Ensure that inclusive words are not substituted with their specific subsets. For example, if the word", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 661, 479, 676 ], "spans": [ { "bbox": [ 116, 661, 479, 676 ], "score": 1.0, "content": "is ”people,” avoid replacing it with genders like ”man” or ”woman.” Instead, consider modifying", "type": "text" } ], "index": 23 }, { "bbox": [ 117, 671, 352, 686 ], "spans": [ { "bbox": [ 117, 671, 283, 686 ], "score": 1.0, "content": "them to different categories, such as ”people”", "type": "text" }, { "bbox": [ 284, 674, 295, 682 ], "score": 0.8, "content": "", "type": "inline_equation" }, { "bbox": [ 296, 671, 352, 686 ], "score": 1.0, "content": "”animals.”.”’}]", "type": "text" } ], "index": 24 } ], "index": 23 } ], "page_idx": 16, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "17", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 109, 112, 501, 347 ], "blocks": [ { "type": "table_caption", "bbox": [ 107, 80, 504, 102 ], "group_id": 0, "lines": [ { "bbox": [ 105, 79, 505, 93 ], "spans": [ { "bbox": [ 105, 79, 505, 93 ], "score": 1.0, "content": "Table 9: Examples of task templates Ferret used to transfer different public data types into the", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 92, 225, 103 ], "spans": [ { "bbox": [ 106, 92, 225, 103 ], "score": 1.0, "content": "instruction-following format.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "table_body", "bbox": [ 109, 112, 501, 347 ], "group_id": 0, "lines": [ { "bbox": [ 109, 112, 501, 347 ], "spans": [ { "bbox": [ 109, 112, 501, 347 ], "score": 0.984, "html": "
TaskThree randomly chosen examples from many.
Referring-ObjectWhat is the class of the object <location>within the image? Classify object <location> in the image. Identify the object <location> in the image.
Referring-RelationWhat does <object1><location1> do to <object2><location2> of the image? What is the physical relation between <object1><locationl>and <object2><location2>? Can you figure out the geometric relation of the <objectl><locationl>and <object2><location2>?
Referring-RegionDescribe the region <location> in a short phrase. What is in the region <location>? Describe in a phrase. Capture in a phrase:what's near region <location>in the picture?
REC.Where is<object>in the image? What are the coordinates for the given <object> in the image? Given the image,could you please tell me where is <object>
Phrase GroundingWhat are the locations of <objects>? Could you provide me with the exact locations of <objects>? Please indicate the positions of <objects> in the image?
Object Detection (O365)Detect all objects among <class> in the image. Perform object detection given the image within <class>. Given the image and set <class>,identify allthe objects that belong to the set.
Grounded CaptioningWhat is this photo about? Use concise language. Describe the overall picture in just a few words. What do you see happening in this image? Provide the answer in short.
Object HallucinationIs there a <object>in the image? Are there <object>in the image? Please tell me whether<object> exists in the image?
", "type": "table", "image_path": "5320f095ac9a8244a5928ae20f13d9a21d932e723e97adb30cf9e843722f60e4.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 109, 112, 501, 190.33333333333331 ], "spans": [], "index": 2 }, { "bbox": [ 109, 190.33333333333331, 501, 268.66666666666663 ], "spans": [], "index": 3 }, { "bbox": [ 109, 268.66666666666663, 501, 346.99999999999994 ], "spans": [], "index": 4 } ] } ], "index": 1.75 }, { "type": "text", "bbox": [ 132, 373, 477, 385 ], "lines": [ { "bbox": [ 132, 371, 479, 388 ], "spans": [ { "bbox": [ 132, 371, 479, 388 ], "score": 1.0, "content": "Table 10: In this example, we provide the prompt to generate the spatial negative sets.", "type": "text" } ], "index": 5 } ], "index": 5, "bbox_fs": [ 132, 371, 479, 388 ] }, { "type": "text", "bbox": [ 119, 404, 478, 434 ], "lines": [ { "bbox": [ 116, 402, 480, 417 ], "spans": [ { "bbox": [ 116, 402, 163, 417 ], "score": 1.0, "content": "messages", "type": "text" }, { "bbox": [ 163, 405, 172, 413 ], "score": 0.53, "content": "=", "type": "inline_equation" }, { "bbox": [ 172, 402, 480, 417 ], "score": 1.0, "content": "[ {\"role\":\"system\", \"content\": f”’You are an AI visual assistant that can", "type": "text" } ], "index": 6 }, { "bbox": [ 117, 413, 479, 426 ], "spans": [ { "bbox": [ 117, 413, 479, 426 ], "score": 1.0, "content": "analyze a single image. You receive several entities given by a list, each describing the objects in", "type": "text" } ], "index": 7 }, { "bbox": [ 118, 423, 225, 436 ], "spans": [ { "bbox": [ 118, 423, 225, 436 ], "score": 1.0, "content": "the image you are observing.", "type": "text" } ], "index": 8 } ], "index": 7, "bbox_fs": [ 116, 402, 480, 436 ] }, { "type": "text", "bbox": [ 119, 444, 478, 484 ], "lines": [ { "bbox": [ 117, 442, 480, 457 ], "spans": [ { "bbox": [ 117, 442, 480, 457 ], "score": 1.0, "content": "For each entity mentioned, change them with the most misleading entity name (may belong to", "type": "text" } ], "index": 9 }, { "bbox": [ 116, 452, 480, 466 ], "spans": [ { "bbox": [ 116, 452, 387, 466 ], "score": 1.0, "content": "the same category but are actually different) (nonexistent objects: man", "type": "text" }, { "bbox": [ 387, 455, 399, 463 ], "score": 0.81, "content": "", "type": "inline_equation" }, { "bbox": [ 400, 452, 480, 466 ], "score": 1.0, "content": "woman, nonexistent", "type": "text" } ], "index": 10 }, { "bbox": [ 117, 463, 479, 475 ], "spans": [ { "bbox": [ 117, 463, 186, 475 ], "score": 1.0, "content": "attributes: brown", "type": "text" }, { "bbox": [ 187, 465, 199, 473 ], "score": 0.81, "content": "", "type": "inline_equation" }, { "bbox": [ 199, 463, 333, 475 ], "score": 1.0, "content": "yellow, nonexistent quantities: two", "type": "text" }, { "bbox": [ 334, 465, 345, 473 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 346, 463, 479, 475 ], "score": 1.0, "content": "three, etc.). The instructions should", "type": "text" } ], "index": 11 }, { "bbox": [ 117, 474, 291, 486 ], "spans": [ { "bbox": [ 117, 474, 291, 486 ], "score": 1.0, "content": "contain interrogative and declarative sentences.", "type": "text" } ], "index": 12 } ], "index": 10.5, "bbox_fs": [ 116, 442, 480, 486 ] }, { "type": "text", "bbox": [ 118, 493, 477, 514 ], "lines": [ { "bbox": [ 118, 493, 478, 505 ], "spans": [ { "bbox": [ 118, 493, 478, 505 ], "score": 1.0, "content": "The output format needs to be a list only which contains the misleading entity names. Please follow", "type": "text" } ], "index": 13 }, { "bbox": [ 117, 502, 212, 515 ], "spans": [ { "bbox": [ 117, 502, 212, 515 ], "score": 1.0, "content": "the instructions carefully.", "type": "text" } ], "index": 14 } ], "index": 13.5, "bbox_fs": [ 117, 493, 478, 515 ] }, { "type": "text", "bbox": [ 118, 523, 380, 534 ], "lines": [ { "bbox": [ 117, 523, 381, 535 ], "spans": [ { "bbox": [ 117, 523, 381, 535 ], "score": 1.0, "content": "1. The length of the output list needs to be exactly equal to the input list.", "type": "text" } ], "index": 15 } ], "index": 15, "bbox_fs": [ 117, 523, 381, 535 ] }, { "type": "text", "bbox": [ 119, 543, 227, 554 ], "lines": [ { "bbox": [ 118, 543, 228, 554 ], "spans": [ { "bbox": [ 118, 543, 228, 554 ], "score": 1.0, "content": "2. Do not explain the reasons.", "type": "text" } ], "index": 16 } ], "index": 16, "bbox_fs": [ 118, 543, 228, 554 ] }, { "type": "text", "bbox": [ 115, 563, 473, 574 ], "lines": [ { "bbox": [ 115, 563, 474, 575 ], "spans": [ { "bbox": [ 115, 563, 474, 575 ], "score": 1.0, "content": "3. Do not mention the input entities, at least the output name and input name needs to be different.", "type": "text" } ], "index": 17 } ], "index": 17, "bbox_fs": [ 115, 563, 474, 575 ] }, { "type": "text", "bbox": [ 119, 583, 295, 594 ], "lines": [ { "bbox": [ 118, 582, 297, 594 ], "spans": [ { "bbox": [ 118, 582, 297, 594 ], "score": 1.0, "content": "4. Do not mention something abstract, like alien ¨ .", "type": "text" } ], "index": 18 } ], "index": 18, "bbox_fs": [ 118, 582, 297, 594 ] }, { "type": "text", "bbox": [ 118, 603, 438, 614 ], "lines": [ { "bbox": [ 117, 601, 439, 615 ], "spans": [ { "bbox": [ 117, 601, 439, 615 ], "score": 1.0, "content": "5. When dealing with quantities, focus solely on increasing the numbers during revision.", "type": "text" } ], "index": 19 } ], "index": 19, "bbox_fs": [ 117, 601, 439, 615 ] }, { "type": "text", "bbox": [ 118, 623, 478, 644 ], "lines": [ { "bbox": [ 116, 621, 480, 637 ], "spans": [ { "bbox": [ 116, 621, 480, 637 ], "score": 1.0, "content": "6. When dealing with words like ”a few”, ”a group”, ”several”, ”some”, etc., try changing the", "type": "text" } ], "index": 20 }, { "bbox": [ 117, 631, 261, 645 ], "spans": [ { "bbox": [ 117, 631, 190, 645 ], "score": 1.0, "content": "objects (A few men", "type": "text" }, { "bbox": [ 190, 633, 210, 642 ], "score": 0.84, "content": " \\mathbf { A }", "type": "inline_equation" }, { "bbox": [ 211, 631, 261, 645 ], "score": 1.0, "content": "few women).", "type": "text" } ], "index": 21 } ], "index": 20.5, "bbox_fs": [ 116, 621, 480, 645 ] }, { "type": "text", "bbox": [ 118, 653, 478, 684 ], "lines": [ { "bbox": [ 117, 652, 479, 665 ], "spans": [ { "bbox": [ 117, 652, 479, 665 ], "score": 1.0, "content": "7. Ensure that inclusive words are not substituted with their specific subsets. For example, if the word", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 661, 479, 676 ], "spans": [ { "bbox": [ 116, 661, 479, 676 ], "score": 1.0, "content": "is ”people,” avoid replacing it with genders like ”man” or ”woman.” Instead, consider modifying", "type": "text" } ], "index": 23 }, { "bbox": [ 117, 671, 352, 686 ], "spans": [ { "bbox": [ 117, 671, 283, 686 ], "score": 1.0, "content": "them to different categories, such as ”people”", "type": "text" }, { "bbox": [ 284, 674, 295, 682 ], "score": 0.8, "content": "", "type": "inline_equation" }, { "bbox": [ 296, 671, 352, 686 ], "score": 1.0, "content": "”animals.”.”’}]", "type": "text" } ], "index": 24 } ], "index": 23, "bbox_fs": [ 116, 652, 479, 686 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 102, 505, 147 ], "lines": [ { "bbox": [ 105, 102, 505, 115 ], "spans": [ { "bbox": [ 105, 102, 505, 115 ], "score": 1.0, "content": "We provide some example prompts to generate refer-and-ground from ChatGPT/GPT-4. Prompt", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 114, 505, 126 ], "spans": [ { "bbox": [ 105, 114, 505, 126 ], "score": 1.0, "content": "and the in-context example of multiple-round visual conversation data are shown in Table 11 and", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 125, 505, 138 ], "spans": [ { "bbox": [ 105, 125, 505, 138 ], "score": 1.0, "content": "Table 12. Prompt and the in-context example of one-round reasoning data are shown in Table 13", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 136, 163, 147 ], "spans": [ { "bbox": [ 106, 136, 163, 147 ], "score": 1.0, "content": "and Table 14.", "type": "text" } ], "index": 3 } ], "index": 1.5 }, { "type": "text", "bbox": [ 106, 155, 504, 178 ], "lines": [ { "bbox": [ 105, 154, 505, 168 ], "spans": [ { "bbox": [ 105, 154, 505, 168 ], "score": 1.0, "content": "Table 11: In this example, we provide the prompt used to generate the conversation response for", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 167, 462, 179 ], "spans": [ { "bbox": [ 105, 167, 462, 179 ], "score": 1.0, "content": "refer-and-ground instruction tuning, following the practice of LLaVA (Liu et al., 2023b).", "type": "text" } ], "index": 5 } ], "index": 4.5 }, { "type": "text", "bbox": [ 119, 197, 478, 288 ], "lines": [ { "bbox": [ 116, 197, 479, 208 ], "spans": [ { "bbox": [ 116, 197, 163, 208 ], "score": 1.0, "content": "messages", "type": "text" }, { "bbox": [ 163, 198, 218, 208 ], "score": 0.38, "content": "\\mathbf { \\Sigma } = [ \\mathbf { \\Sigma } \\{ { \\ \" } \\Sigma \\circ \\mathrm { { 1 } } \\mathsf { e \" \\Sigma }", "type": "inline_equation" }, { "bbox": [ 218, 197, 479, 208 ], "score": 1.0, "content": ":\"system\", \"content\": f”’You are an AI visual assistant that", "type": "text" } ], "index": 6 }, { "bbox": [ 117, 206, 479, 219 ], "spans": [ { "bbox": [ 117, 206, 479, 219 ], "score": 1.0, "content": "can analyze a single image. You receive five global captions, each describing the same image you", "type": "text" } ], "index": 7 }, { "bbox": [ 117, 217, 479, 229 ], "spans": [ { "bbox": [ 117, 217, 479, 229 ], "score": 1.0, "content": "are observing. In addition, specific object locations within the image are given, along with detailed", "type": "text" } ], "index": 8 }, { "bbox": [ 117, 227, 478, 239 ], "spans": [ { "bbox": [ 117, 227, 478, 239 ], "score": 1.0, "content": "coordinates. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2)", "type": "text" } ], "index": 9 }, { "bbox": [ 117, 236, 479, 249 ], "spans": [ { "bbox": [ 117, 236, 479, 249 ], "score": 1.0, "content": "with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y,", "type": "text" } ], "index": 10 }, { "bbox": [ 117, 246, 479, 259 ], "spans": [ { "bbox": [ 117, 246, 165, 259 ], "score": 1.0, "content": "bottom right", "type": "text" }, { "bbox": [ 166, 249, 172, 256 ], "score": 0.28, "content": "\\mathbf { X }", "type": "inline_equation" }, { "bbox": [ 172, 246, 479, 259 ], "score": 1.0, "content": ", and bottom right y. Also, the relationships between pairs of objects are provided in", "type": "text" } ], "index": 11 }, { "bbox": [ 117, 257, 479, 269 ], "spans": [ { "bbox": [ 117, 257, 194, 269 ], "score": 1.0, "content": "the format of object", "type": "text" }, { "bbox": [ 194, 258, 206, 267 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 207, 257, 252, 269 ], "score": 1.0, "content": "relationship", "type": "text" }, { "bbox": [ 252, 258, 264, 267 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 264, 257, 479, 269 ], "score": 1.0, "content": "subject, where the object/subject are indexed by object id", "type": "text" } ], "index": 12 }, { "bbox": [ 117, 266, 479, 279 ], "spans": [ { "bbox": [ 117, 266, 479, 279 ], "score": 1.0, "content": "from previous object lists as well as the object names. Also, several region descriptions are given,", "type": "text" } ], "index": 13 }, { "bbox": [ 117, 277, 368, 289 ], "spans": [ { "bbox": [ 117, 277, 368, 289 ], "score": 1.0, "content": "each describing a box region of the image, with detailed coordinates.", "type": "text" } ], "index": 14 } ], "index": 10 }, { "type": "text", "bbox": [ 118, 297, 478, 327 ], "lines": [ { "bbox": [ 117, 295, 480, 309 ], "spans": [ { "bbox": [ 117, 295, 480, 309 ], "score": 1.0, "content": "Design a conversation between you and a person asking about this photo. Ask diverse questions and", "type": "text" } ], "index": 15 }, { "bbox": [ 117, 306, 479, 320 ], "spans": [ { "bbox": [ 117, 306, 479, 320 ], "score": 1.0, "content": "give corresponding answers. The answers should be in a tone that a visual AI assistant is seeing the", "type": "text" } ], "index": 16 }, { "bbox": [ 117, 316, 245, 329 ], "spans": [ { "bbox": [ 117, 316, 245, 329 ], "score": 1.0, "content": "image and answering the question.", "type": "text" } ], "index": 17 } ], "index": 16 }, { "type": "text", "bbox": [ 119, 336, 403, 347 ], "lines": [ { "bbox": [ 117, 336, 404, 348 ], "spans": [ { "bbox": [ 117, 336, 404, 348 ], "score": 1.0, "content": "Here are some additional requirements about generated questions and answers:", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 118, 357, 478, 397 ], "lines": [ { "bbox": [ 117, 356, 311, 369 ], "spans": [ { "bbox": [ 117, 356, 311, 369 ], "score": 1.0, "content": "1. Only include questions that have definite answers:", "type": "text" } ], "index": 19 }, { "bbox": [ 116, 365, 468, 379 ], "spans": [ { "bbox": [ 116, 365, 468, 379 ], "score": 1.0, "content": "(1) one can see the content in the image that the question asks about and can answer confidently;", "type": "text" } ], "index": 20 }, { "bbox": [ 116, 375, 478, 389 ], "spans": [ { "bbox": [ 116, 375, 478, 389 ], "score": 1.0, "content": "(2) one can determine confidently from the image that it is not in the image. Do not ask any", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 387, 288, 399 ], "spans": [ { "bbox": [ 116, 387, 288, 399 ], "score": 1.0, "content": "questions that cannot be answered confidently.", "type": "text" } ], "index": 22 } ], "index": 20.5 }, { "type": "text", "bbox": [ 119, 407, 478, 446 ], "lines": [ { "bbox": [ 116, 405, 479, 419 ], "spans": [ { "bbox": [ 116, 405, 479, 419 ], "score": 1.0, "content": "2. Also include complex questions that are relevant to the content in the image, for example, asking", "type": "text" } ], "index": 23 }, { "bbox": [ 117, 415, 479, 429 ], "spans": [ { "bbox": [ 117, 415, 479, 429 ], "score": 1.0, "content": "about background knowledge of the objects in the image, asking to discuss events happening in the", "type": "text" } ], "index": 24 }, { "bbox": [ 117, 426, 479, 438 ], "spans": [ { "bbox": [ 117, 426, 479, 438 ], "score": 1.0, "content": "image, asking about object actions in the context of entire images, etc. Again, do not ask about", "type": "text" } ], "index": 25 }, { "bbox": [ 117, 437, 182, 447 ], "spans": [ { "bbox": [ 117, 437, 182, 447 ], "score": 1.0, "content": "uncertain details.", "type": "text" } ], "index": 26 } ], "index": 24.5 }, { "type": "text", "bbox": [ 117, 456, 478, 487 ], "lines": [ { "bbox": [ 115, 454, 480, 469 ], "spans": [ { "bbox": [ 115, 454, 480, 469 ], "score": 1.0, "content": "3. Provide detailed answers when answering complex questions. For example, give detailed", "type": "text" } ], "index": 27 }, { "bbox": [ 116, 466, 479, 479 ], "spans": [ { "bbox": [ 116, 466, 479, 479 ], "score": 1.0, "content": "examples or reasoning steps to make the content more convincing and well-organized. You can", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 475, 267, 489 ], "spans": [ { "bbox": [ 116, 475, 267, 489 ], "score": 1.0, "content": "include multiple paragraphs if necessary.", "type": "text" } ], "index": 29 } ], "index": 28 }, { "type": "text", "bbox": [ 117, 496, 478, 527 ], "lines": [ { "bbox": [ 116, 495, 479, 508 ], "spans": [ { "bbox": [ 116, 495, 479, 508 ], "score": 1.0, "content": "4. In all samples, either in question or answer, you must mention bounding box coordinates to refer", "type": "text" } ], "index": 30 }, { "bbox": [ 116, 505, 479, 519 ], "spans": [ { "bbox": [ 116, 505, 479, 519 ], "score": 1.0, "content": "to the object or regions instead of directly saying the object name or describing the regions in text.", "type": "text" } ], "index": 31 }, { "bbox": [ 116, 515, 324, 528 ], "spans": [ { "bbox": [ 116, 515, 324, 528 ], "score": 1.0, "content": "In answer, explain the region in the context of the scene.", "type": "text" } ], "index": 32 } ], "index": 31 }, { "type": "text", "bbox": [ 115, 536, 477, 556 ], "lines": [ { "bbox": [ 116, 534, 478, 549 ], "spans": [ { "bbox": [ 116, 534, 478, 549 ], "score": 1.0, "content": "5. Do not mention that the information source is provided in the text/caption/region description.", "type": "text" } ], "index": 33 }, { "bbox": [ 117, 545, 330, 558 ], "spans": [ { "bbox": [ 117, 545, 330, 558 ], "score": 1.0, "content": "Always answer as if you are directly looking at the image.", "type": "text" } ], "index": 34 } ], "index": 33.5 }, { "type": "text", "bbox": [ 119, 566, 481, 616 ], "lines": [ { "bbox": [ 117, 565, 480, 578 ], "spans": [ { "bbox": [ 117, 565, 480, 578 ], "score": 1.0, "content": "6. Make the question as diverse as possible. Include questions asking about the visual content of", "type": "text" } ], "index": 35 }, { "bbox": [ 117, 575, 480, 588 ], "spans": [ { "bbox": [ 117, 575, 480, 588 ], "score": 1.0, "content": "the image, including the object types, counting the objects, object actions, object locations, relative", "type": "text" } ], "index": 36 }, { "bbox": [ 117, 585, 479, 599 ], "spans": [ { "bbox": [ 117, 585, 479, 599 ], "score": 1.0, "content": "positions between objects, object selection, object functions, etc. Make the question challenging by", "type": "text" } ], "index": 37 }, { "bbox": [ 117, 594, 331, 608 ], "spans": [ { "bbox": [ 117, 594, 331, 608 ], "score": 1.0, "content": "less including the visual content details in the question.”’}", "type": "text" } ], "index": 38 } ], "index": 36.5 }, { "type": "text", "bbox": [ 125, 617, 491, 656 ], "lines": [ { "bbox": [ 124, 615, 281, 626 ], "spans": [ { "bbox": [ 124, 618, 136, 625 ], "score": 1.0, "content": "or", "type": "text" }, { "bbox": [ 141, 616, 178, 626 ], "score": 1.0, "content": "sample", "type": "text" }, { "bbox": [ 180, 617, 193, 624 ], "score": 1.0, "content": "in", "type": "text" }, { "bbox": [ 195, 615, 281, 626 ], "score": 1.0, "content": "fewshot samples:", "type": "text" } ], "index": 39 }, { "bbox": [ 131, 624, 462, 638 ], "spans": [ { "bbox": [ 131, 624, 462, 638 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":sample[‘context’]})", "type": "text" } ], "index": 40 }, { "bbox": [ 131, 634, 487, 647 ], "spans": [ { "bbox": [ 131, 634, 487, 647 ], "score": 1.0, "content": "messages.append({\"role\":\"assistant\", \"content\":sample[‘response’]}", "type": "text" } ], "index": 41 } ], "index": 40 }, { "type": "text", "bbox": [ 116, 657, 438, 668 ], "lines": [ { "bbox": [ 116, 655, 439, 669 ], "spans": [ { "bbox": [ 116, 655, 439, 669 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":‘\\n’.join(query)})", "type": "text" } ], "index": 42 } ], "index": 42 } ], "page_idx": 17, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 763 ], "spans": [ { "bbox": [ 299, 750, 313, 763 ], "score": 1.0, "content": "18", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 82, 400, 94 ], "lines": [ { "bbox": [ 105, 82, 402, 95 ], "spans": [ { "bbox": [ 105, 82, 402, 95 ], "score": 1.0, "content": "C.3 EXAMPLES FOR GENERATING REFER-AND-GROUND DATASETS", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 102, 505, 147 ], "lines": [ { "bbox": [ 105, 102, 505, 115 ], "spans": [ { "bbox": [ 105, 102, 505, 115 ], "score": 1.0, "content": "We provide some example prompts to generate refer-and-ground from ChatGPT/GPT-4. Prompt", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 114, 505, 126 ], "spans": [ { "bbox": [ 105, 114, 505, 126 ], "score": 1.0, "content": "and the in-context example of multiple-round visual conversation data are shown in Table 11 and", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 125, 505, 138 ], "spans": [ { "bbox": [ 105, 125, 505, 138 ], "score": 1.0, "content": "Table 12. Prompt and the in-context example of one-round reasoning data are shown in Table 13", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 136, 163, 147 ], "spans": [ { "bbox": [ 106, 136, 163, 147 ], "score": 1.0, "content": "and Table 14.", "type": "text" } ], "index": 3 } ], "index": 1.5, "bbox_fs": [ 105, 102, 505, 147 ] }, { "type": "text", "bbox": [ 106, 155, 504, 178 ], "lines": [ { "bbox": [ 105, 154, 505, 168 ], "spans": [ { "bbox": [ 105, 154, 505, 168 ], "score": 1.0, "content": "Table 11: In this example, we provide the prompt used to generate the conversation response for", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 167, 462, 179 ], "spans": [ { "bbox": [ 105, 167, 462, 179 ], "score": 1.0, "content": "refer-and-ground instruction tuning, following the practice of LLaVA (Liu et al., 2023b).", "type": "text" } ], "index": 5 } ], "index": 4.5, "bbox_fs": [ 105, 154, 505, 179 ] }, { "type": "text", "bbox": [ 119, 197, 478, 288 ], "lines": [ { "bbox": [ 116, 197, 479, 208 ], "spans": [ { "bbox": [ 116, 197, 163, 208 ], "score": 1.0, "content": "messages", "type": "text" }, { "bbox": [ 163, 198, 218, 208 ], "score": 0.38, "content": "\\mathbf { \\Sigma } = [ \\mathbf { \\Sigma } \\{ { \\ \" } \\Sigma \\circ \\mathrm { { 1 } } \\mathsf { e \" \\Sigma }", "type": "inline_equation" }, { "bbox": [ 218, 197, 479, 208 ], "score": 1.0, "content": ":\"system\", \"content\": f”’You are an AI visual assistant that", "type": "text" } ], "index": 6 }, { "bbox": [ 117, 206, 479, 219 ], "spans": [ { "bbox": [ 117, 206, 479, 219 ], "score": 1.0, "content": "can analyze a single image. You receive five global captions, each describing the same image you", "type": "text" } ], "index": 7 }, { "bbox": [ 117, 217, 479, 229 ], "spans": [ { "bbox": [ 117, 217, 479, 229 ], "score": 1.0, "content": "are observing. In addition, specific object locations within the image are given, along with detailed", "type": "text" } ], "index": 8 }, { "bbox": [ 117, 227, 478, 239 ], "spans": [ { "bbox": [ 117, 227, 478, 239 ], "score": 1.0, "content": "coordinates. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2)", "type": "text" } ], "index": 9 }, { "bbox": [ 117, 236, 479, 249 ], "spans": [ { "bbox": [ 117, 236, 479, 249 ], "score": 1.0, "content": "with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y,", "type": "text" } ], "index": 10 }, { "bbox": [ 117, 246, 479, 259 ], "spans": [ { "bbox": [ 117, 246, 165, 259 ], "score": 1.0, "content": "bottom right", "type": "text" }, { "bbox": [ 166, 249, 172, 256 ], "score": 0.28, "content": "\\mathbf { X }", "type": "inline_equation" }, { "bbox": [ 172, 246, 479, 259 ], "score": 1.0, "content": ", and bottom right y. Also, the relationships between pairs of objects are provided in", "type": "text" } ], "index": 11 }, { "bbox": [ 117, 257, 479, 269 ], "spans": [ { "bbox": [ 117, 257, 194, 269 ], "score": 1.0, "content": "the format of object", "type": "text" }, { "bbox": [ 194, 258, 206, 267 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 207, 257, 252, 269 ], "score": 1.0, "content": "relationship", "type": "text" }, { "bbox": [ 252, 258, 264, 267 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 264, 257, 479, 269 ], "score": 1.0, "content": "subject, where the object/subject are indexed by object id", "type": "text" } ], "index": 12 }, { "bbox": [ 117, 266, 479, 279 ], "spans": [ { "bbox": [ 117, 266, 479, 279 ], "score": 1.0, "content": "from previous object lists as well as the object names. Also, several region descriptions are given,", "type": "text" } ], "index": 13 }, { "bbox": [ 117, 277, 368, 289 ], "spans": [ { "bbox": [ 117, 277, 368, 289 ], "score": 1.0, "content": "each describing a box region of the image, with detailed coordinates.", "type": "text" } ], "index": 14 } ], "index": 10, "bbox_fs": [ 116, 197, 479, 289 ] }, { "type": "text", "bbox": [ 118, 297, 478, 327 ], "lines": [ { "bbox": [ 117, 295, 480, 309 ], "spans": [ { "bbox": [ 117, 295, 480, 309 ], "score": 1.0, "content": "Design a conversation between you and a person asking about this photo. Ask diverse questions and", "type": "text" } ], "index": 15 }, { "bbox": [ 117, 306, 479, 320 ], "spans": [ { "bbox": [ 117, 306, 479, 320 ], "score": 1.0, "content": "give corresponding answers. The answers should be in a tone that a visual AI assistant is seeing the", "type": "text" } ], "index": 16 }, { "bbox": [ 117, 316, 245, 329 ], "spans": [ { "bbox": [ 117, 316, 245, 329 ], "score": 1.0, "content": "image and answering the question.", "type": "text" } ], "index": 17 } ], "index": 16, "bbox_fs": [ 117, 295, 480, 329 ] }, { "type": "text", "bbox": [ 119, 336, 403, 347 ], "lines": [ { "bbox": [ 117, 336, 404, 348 ], "spans": [ { "bbox": [ 117, 336, 404, 348 ], "score": 1.0, "content": "Here are some additional requirements about generated questions and answers:", "type": "text" } ], "index": 18 } ], "index": 18, "bbox_fs": [ 117, 336, 404, 348 ] }, { "type": "list", "bbox": [ 118, 357, 478, 397 ], "lines": [ { "bbox": [ 117, 356, 311, 369 ], "spans": [ { "bbox": [ 117, 356, 311, 369 ], "score": 1.0, "content": "1. Only include questions that have definite answers:", "type": "text" } ], "index": 19, "is_list_end_line": true }, { "bbox": [ 116, 365, 468, 379 ], "spans": [ { "bbox": [ 116, 365, 468, 379 ], "score": 1.0, "content": "(1) one can see the content in the image that the question asks about and can answer confidently;", "type": "text" } ], "index": 20, "is_list_start_line": true }, { "bbox": [ 116, 375, 478, 389 ], "spans": [ { "bbox": [ 116, 375, 478, 389 ], "score": 1.0, "content": "(2) one can determine confidently from the image that it is not in the image. Do not ask any", "type": "text" } ], "index": 21 }, { "bbox": [ 116, 387, 288, 399 ], "spans": [ { "bbox": [ 116, 387, 288, 399 ], "score": 1.0, "content": "questions that cannot be answered confidently.", "type": "text" } ], "index": 22, "is_list_end_line": true } ], "index": 20.5, "bbox_fs": [ 116, 356, 478, 399 ] }, { "type": "text", "bbox": [ 119, 407, 478, 446 ], "lines": [ { "bbox": [ 116, 405, 479, 419 ], "spans": [ { "bbox": [ 116, 405, 479, 419 ], "score": 1.0, "content": "2. Also include complex questions that are relevant to the content in the image, for example, asking", "type": "text" } ], "index": 23 }, { "bbox": [ 117, 415, 479, 429 ], "spans": [ { "bbox": [ 117, 415, 479, 429 ], "score": 1.0, "content": "about background knowledge of the objects in the image, asking to discuss events happening in the", "type": "text" } ], "index": 24 }, { "bbox": [ 117, 426, 479, 438 ], "spans": [ { "bbox": [ 117, 426, 479, 438 ], "score": 1.0, "content": "image, asking about object actions in the context of entire images, etc. Again, do not ask about", "type": "text" } ], "index": 25 }, { "bbox": [ 117, 437, 182, 447 ], "spans": [ { "bbox": [ 117, 437, 182, 447 ], "score": 1.0, "content": "uncertain details.", "type": "text" } ], "index": 26 } ], "index": 24.5, "bbox_fs": [ 116, 405, 479, 447 ] }, { "type": "text", "bbox": [ 117, 456, 478, 487 ], "lines": [ { "bbox": [ 115, 454, 480, 469 ], "spans": [ { "bbox": [ 115, 454, 480, 469 ], "score": 1.0, "content": "3. Provide detailed answers when answering complex questions. For example, give detailed", "type": "text" } ], "index": 27 }, { "bbox": [ 116, 466, 479, 479 ], "spans": [ { "bbox": [ 116, 466, 479, 479 ], "score": 1.0, "content": "examples or reasoning steps to make the content more convincing and well-organized. You can", "type": "text" } ], "index": 28 }, { "bbox": [ 116, 475, 267, 489 ], "spans": [ { "bbox": [ 116, 475, 267, 489 ], "score": 1.0, "content": "include multiple paragraphs if necessary.", "type": "text" } ], "index": 29 } ], "index": 28, "bbox_fs": [ 115, 454, 480, 489 ] }, { "type": "text", "bbox": [ 117, 496, 478, 527 ], "lines": [ { "bbox": [ 116, 495, 479, 508 ], "spans": [ { "bbox": [ 116, 495, 479, 508 ], "score": 1.0, "content": "4. In all samples, either in question or answer, you must mention bounding box coordinates to refer", "type": "text" } ], "index": 30 }, { "bbox": [ 116, 505, 479, 519 ], "spans": [ { "bbox": [ 116, 505, 479, 519 ], "score": 1.0, "content": "to the object or regions instead of directly saying the object name or describing the regions in text.", "type": "text" } ], "index": 31 }, { "bbox": [ 116, 515, 324, 528 ], "spans": [ { "bbox": [ 116, 515, 324, 528 ], "score": 1.0, "content": "In answer, explain the region in the context of the scene.", "type": "text" } ], "index": 32 } ], "index": 31, "bbox_fs": [ 116, 495, 479, 528 ] }, { "type": "list", "bbox": [ 115, 536, 477, 556 ], "lines": [ { "bbox": [ 116, 534, 478, 549 ], "spans": [ { "bbox": [ 116, 534, 478, 549 ], "score": 1.0, "content": "5. Do not mention that the information source is provided in the text/caption/region description.", "type": "text" } ], "index": 33, "is_list_end_line": true }, { "bbox": [ 117, 545, 330, 558 ], "spans": [ { "bbox": [ 117, 545, 330, 558 ], "score": 1.0, "content": "Always answer as if you are directly looking at the image.", "type": "text" } ], "index": 34, "is_list_start_line": true, "is_list_end_line": true } ], "index": 33.5, "bbox_fs": [ 116, 534, 478, 558 ] }, { "type": "text", "bbox": [ 119, 566, 481, 616 ], "lines": [ { "bbox": [ 117, 565, 480, 578 ], "spans": [ { "bbox": [ 117, 565, 480, 578 ], "score": 1.0, "content": "6. Make the question as diverse as possible. Include questions asking about the visual content of", "type": "text" } ], "index": 35 }, { "bbox": [ 117, 575, 480, 588 ], "spans": [ { "bbox": [ 117, 575, 480, 588 ], "score": 1.0, "content": "the image, including the object types, counting the objects, object actions, object locations, relative", "type": "text" } ], "index": 36 }, { "bbox": [ 117, 585, 479, 599 ], "spans": [ { "bbox": [ 117, 585, 479, 599 ], "score": 1.0, "content": "positions between objects, object selection, object functions, etc. Make the question challenging by", "type": "text" } ], "index": 37 }, { "bbox": [ 117, 594, 331, 608 ], "spans": [ { "bbox": [ 117, 594, 331, 608 ], "score": 1.0, "content": "less including the visual content details in the question.”’}", "type": "text" } ], "index": 38 } ], "index": 36.5, "bbox_fs": [ 117, 565, 480, 608 ] }, { "type": "text", "bbox": [ 125, 617, 491, 656 ], "lines": [ { "bbox": [ 124, 615, 281, 626 ], "spans": [ { "bbox": [ 124, 618, 136, 625 ], "score": 1.0, "content": "or", "type": "text" }, { "bbox": [ 141, 616, 178, 626 ], "score": 1.0, "content": "sample", "type": "text" }, { "bbox": [ 180, 617, 193, 624 ], "score": 1.0, "content": "in", "type": "text" }, { "bbox": [ 195, 615, 281, 626 ], "score": 1.0, "content": "fewshot samples:", "type": "text" } ], "index": 39 }, { "bbox": [ 131, 624, 462, 638 ], "spans": [ { "bbox": [ 131, 624, 462, 638 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":sample[‘context’]})", "type": "text" } ], "index": 40 }, { "bbox": [ 131, 634, 487, 647 ], "spans": [ { "bbox": [ 131, 634, 487, 647 ], "score": 1.0, "content": "messages.append({\"role\":\"assistant\", \"content\":sample[‘response’]}", "type": "text" } ], "index": 41 }, { "bbox": [ 116, 655, 439, 669 ], "spans": [ { "bbox": [ 116, 655, 439, 669 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":‘\\n’.join(query)})", "type": "text" } ], "index": 42 } ], "index": 40, "bbox_fs": [ 124, 615, 487, 647 ] }, { "type": "text", "bbox": [ 116, 657, 438, 668 ], "lines": [], "index": 42, "bbox_fs": [ 116, 655, 439, 669 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 129, 206, 500, 622 ], "blocks": [ { "type": "image_caption", "bbox": [ 106, 174, 505, 198 ], "group_id": 0, "lines": [ { "bbox": [ 106, 173, 505, 187 ], "spans": [ { "bbox": [ 106, 173, 505, 187 ], "score": 1.0, "content": "Table 12: One example used in in-context learning to construct GPT-Assisted Refer-and-Ground", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 186, 497, 199 ], "spans": [ { "bbox": [ 105, 186, 497, 199 ], "score": 1.0, "content": "Instructon-Tuning. We encourage readers to refer to the codebase for the prompts for other types.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "image_body", "bbox": [ 129, 206, 500, 622 ], "group_id": 0, "lines": [ { "bbox": [ 129, 206, 500, 622 ], "spans": [ { "bbox": [ 129, 206, 500, 622 ], "score": 0.628, "type": "image", "image_path": "974759c4e2150a7b8073ec1aaf24a0920dfea17100ea9f30e316230ebda2c1f5.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 129, 206, 500, 344.66666666666663 ], "spans": [], "index": 2 }, { "bbox": [ 129, 344.66666666666663, 500, 483.33333333333326 ], "spans": [], "index": 3 }, { "bbox": [ 129, 483.33333333333326, 500, 621.9999999999999 ], "spans": [], "index": 4 } ] } ], "index": 1.75 } ], "page_idx": 18, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "19", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 129, 206, 500, 622 ], "blocks": [ { "type": "image_caption", "bbox": [ 106, 174, 505, 198 ], "group_id": 0, "lines": [ { "bbox": [ 106, 173, 505, 187 ], "spans": [ { "bbox": [ 106, 173, 505, 187 ], "score": 1.0, "content": "Table 12: One example used in in-context learning to construct GPT-Assisted Refer-and-Ground", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 186, 497, 199 ], "spans": [ { "bbox": [ 105, 186, 497, 199 ], "score": 1.0, "content": "Instructon-Tuning. We encourage readers to refer to the codebase for the prompts for other types.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "image_body", "bbox": [ 129, 206, 500, 622 ], "group_id": 0, "lines": [ { "bbox": [ 129, 206, 500, 622 ], "spans": [ { "bbox": [ 129, 206, 500, 622 ], "score": 0.628, "type": "image", "image_path": "974759c4e2150a7b8073ec1aaf24a0920dfea17100ea9f30e316230ebda2c1f5.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 129, 206, 500, 344.66666666666663 ], "spans": [], "index": 2 }, { "bbox": [ 129, 344.66666666666663, 500, 483.33333333333326 ], "spans": [], "index": 3 }, { "bbox": [ 129, 483.33333333333326, 500, 621.9999999999999 ], "spans": [], "index": 4 } ] } ], "index": 1.75 } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 169, 503, 192 ], "lines": [ { "bbox": [ 106, 169, 505, 182 ], "spans": [ { "bbox": [ 106, 169, 505, 182 ], "score": 1.0, "content": "Table 13: In this example, we provide the prompt used to generate the reasoning response for refer-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 180, 441, 193 ], "spans": [ { "bbox": [ 106, 180, 441, 193 ], "score": 1.0, "content": "and-ground instruction tuning, following the practice of LLaVA (Liu et al., 2023b).", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "text", "bbox": [ 119, 210, 478, 301 ], "lines": [ { "bbox": [ 116, 210, 479, 222 ], "spans": [ { "bbox": [ 116, 210, 162, 221 ], "score": 1.0, "content": "messages", "type": "text" }, { "bbox": [ 163, 210, 218, 222 ], "score": 0.3, "content": "\\mathbf { \\Sigma } = [ \\mathbf { \\Sigma } \\{ { \\ \" } \\Sigma \\circ \\mathrm { { 1 } } \\mathsf { e \" \\Sigma }", "type": "inline_equation" }, { "bbox": [ 218, 210, 479, 221 ], "score": 1.0, "content": ":\"system\", \"content\": f”’You are an AI visual assistant that", "type": "text" } ], "index": 2 }, { "bbox": [ 117, 220, 479, 233 ], "spans": [ { "bbox": [ 117, 220, 479, 233 ], "score": 1.0, "content": "can analyze a single image. You receive five global captions, each describing the same image you", "type": "text" } ], "index": 3 }, { "bbox": [ 117, 230, 479, 242 ], "spans": [ { "bbox": [ 117, 230, 479, 242 ], "score": 1.0, "content": "are observing. In addition, specific object locations within the image are given, along with detailed", "type": "text" } ], "index": 4 }, { "bbox": [ 117, 240, 479, 253 ], "spans": [ { "bbox": [ 117, 240, 479, 253 ], "score": 1.0, "content": "coordinates. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2)", "type": "text" } ], "index": 5 }, { "bbox": [ 117, 249, 479, 264 ], "spans": [ { "bbox": [ 117, 249, 479, 264 ], "score": 1.0, "content": "with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y,", "type": "text" } ], "index": 6 }, { "bbox": [ 118, 261, 479, 272 ], "spans": [ { "bbox": [ 118, 261, 479, 272 ], "score": 1.0, "content": "bottom right x, and bottom right y. Also, the relationships between pairs of objects are provided, in", "type": "text" } ], "index": 7 }, { "bbox": [ 117, 270, 479, 282 ], "spans": [ { "bbox": [ 117, 270, 194, 282 ], "score": 1.0, "content": "the format of object", "type": "text" }, { "bbox": [ 194, 272, 206, 280 ], "score": 0.83, "content": "", "type": "inline_equation" }, { "bbox": [ 207, 270, 252, 282 ], "score": 1.0, "content": "relationship", "type": "text" }, { "bbox": [ 252, 272, 264, 280 ], "score": 0.83, "content": "", "type": "inline_equation" }, { "bbox": [ 264, 270, 479, 282 ], "score": 1.0, "content": "subject, where the object/subject are indexed by object id", "type": "text" } ], "index": 8 }, { "bbox": [ 117, 280, 479, 293 ], "spans": [ { "bbox": [ 117, 280, 479, 293 ], "score": 1.0, "content": "from previous object lists as well as the object names. Also, several region descriptions are given,", "type": "text" } ], "index": 9 }, { "bbox": [ 117, 290, 368, 302 ], "spans": [ { "bbox": [ 117, 290, 368, 302 ], "score": 1.0, "content": "each describing a box region of the image, with detailed coordinates.", "type": "text" } ], "index": 10 } ], "index": 6 }, { "type": "text", "bbox": [ 119, 310, 477, 340 ], "lines": [ { "bbox": [ 117, 308, 478, 323 ], "spans": [ { "bbox": [ 117, 308, 478, 323 ], "score": 1.0, "content": "The task is to use the provided image information (objects, attribute, relationship, region description,", "type": "text" } ], "index": 11 }, { "bbox": [ 117, 320, 479, 333 ], "spans": [ { "bbox": [ 117, 320, 479, 333 ], "score": 1.0, "content": "captions), create a plausible and challenging question about the image, and provide the answer in", "type": "text" } ], "index": 12 }, { "bbox": [ 117, 329, 144, 342 ], "spans": [ { "bbox": [ 117, 329, 144, 342 ], "score": 1.0, "content": "detail.", "type": "text" } ], "index": 13 } ], "index": 12 }, { "type": "text", "bbox": [ 118, 350, 477, 370 ], "lines": [ { "bbox": [ 117, 349, 478, 362 ], "spans": [ { "bbox": [ 117, 349, 478, 362 ], "score": 1.0, "content": "Create complex questions that mention specific regions of the image, but the question should require", "type": "text" } ], "index": 14 }, { "bbox": [ 117, 359, 450, 372 ], "spans": [ { "bbox": [ 117, 359, 450, 372 ], "score": 1.0, "content": "some knowledge-aware or high-level commonsense reasoning beyond describing the scene.", "type": "text" } ], "index": 15 } ], "index": 14.5 }, { "type": "text", "bbox": [ 119, 380, 478, 420 ], "lines": [ { "bbox": [ 118, 379, 479, 392 ], "spans": [ { "bbox": [ 118, 379, 479, 392 ], "score": 1.0, "content": "To answer such questions, one should first understand the visual content, then based on the", "type": "text" } ], "index": 16 }, { "bbox": [ 117, 390, 479, 402 ], "spans": [ { "bbox": [ 117, 390, 479, 402 ], "score": 1.0, "content": "background knowledge or reasoning, either explain why the things are happening that way or", "type": "text" } ], "index": 17 }, { "bbox": [ 117, 399, 479, 412 ], "spans": [ { "bbox": [ 117, 399, 479, 412 ], "score": 1.0, "content": "provide guides and help to the user’s request. Make the question challenging by not including the", "type": "text" } ], "index": 18 }, { "bbox": [ 118, 410, 422, 421 ], "spans": [ { "bbox": [ 118, 410, 422, 421 ], "score": 1.0, "content": "visual content details in the question so that the user needs to reason about that first.", "type": "text" } ], "index": 19 } ], "index": 17.5 }, { "type": "text", "bbox": [ 119, 430, 403, 440 ], "lines": [ { "bbox": [ 117, 429, 404, 441 ], "spans": [ { "bbox": [ 117, 429, 404, 441 ], "score": 1.0, "content": "Here are some additional requirements about generated questions and answers:", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 119, 450, 478, 490 ], "lines": [ { "bbox": [ 118, 449, 479, 461 ], "spans": [ { "bbox": [ 118, 449, 479, 461 ], "score": 1.0, "content": "1. In question or answer, you must mention bounding box coordinates to refer to the object or", "type": "text" } ], "index": 21 }, { "bbox": [ 117, 460, 479, 472 ], "spans": [ { "bbox": [ 117, 460, 479, 472 ], "score": 1.0, "content": "regions, instead of directly say the object name or describing the regions in text. In answers, explain", "type": "text" } ], "index": 22 }, { "bbox": [ 118, 470, 479, 482 ], "spans": [ { "bbox": [ 118, 470, 479, 482 ], "score": 1.0, "content": "the region in the context of scene. Include details like object counts, position of the objects, relative", "type": "text" } ], "index": 23 }, { "bbox": [ 118, 480, 225, 491 ], "spans": [ { "bbox": [ 118, 480, 225, 491 ], "score": 1.0, "content": "position between the objects.", "type": "text" } ], "index": 24 } ], "index": 22.5 }, { "type": "text", "bbox": [ 114, 500, 478, 520 ], "lines": [ { "bbox": [ 115, 498, 479, 513 ], "spans": [ { "bbox": [ 115, 498, 479, 513 ], "score": 1.0, "content": "2. Don’t ask the question you are not confident to answer. Only include question that have definite", "type": "text" } ], "index": 25 }, { "bbox": [ 117, 510, 148, 520 ], "spans": [ { "bbox": [ 117, 510, 148, 520 ], "score": 1.0, "content": "answer.", "type": "text" } ], "index": 26 } ], "index": 25.5 }, { "type": "text", "bbox": [ 116, 529, 477, 550 ], "lines": [ { "bbox": [ 115, 527, 478, 542 ], "spans": [ { "bbox": [ 115, 527, 478, 542 ], "score": 1.0, "content": "3. Do not mention that the information source is provided in text/catpion/region description. Always", "type": "text" } ], "index": 27 }, { "bbox": [ 116, 538, 301, 552 ], "spans": [ { "bbox": [ 116, 538, 301, 552 ], "score": 1.0, "content": "answer as if you are directly looking at the image.", "type": "text" } ], "index": 28 } ], "index": 27.5 }, { "type": "text", "bbox": [ 117, 559, 456, 580 ], "lines": [ { "bbox": [ 115, 557, 457, 572 ], "spans": [ { "bbox": [ 115, 557, 457, 572 ], "score": 1.0, "content": "4. Make the question as diverse as possible and as complex-reasoning required as possible.”’}", "type": "text" } ], "index": 29 }, { "bbox": [ 117, 570, 124, 578 ], "spans": [ { "bbox": [ 117, 570, 124, 578 ], "score": 1.0, "content": "]", "type": "text" } ], "index": 30 } ], "index": 29.5 }, { "type": "text", "bbox": [ 124, 581, 490, 619 ], "lines": [ { "bbox": [ 123, 579, 281, 590 ], "spans": [ { "bbox": [ 123, 581, 137, 589 ], "score": 1.0, "content": "or", "type": "text" }, { "bbox": [ 141, 580, 178, 590 ], "score": 1.0, "content": "sample", "type": "text" }, { "bbox": [ 179, 581, 193, 589 ], "score": 1.0, "content": "in", "type": "text" }, { "bbox": [ 195, 579, 281, 590 ], "score": 1.0, "content": "fewshot samples:", "type": "text" } ], "index": 31 }, { "bbox": [ 131, 587, 462, 602 ], "spans": [ { "bbox": [ 131, 587, 462, 602 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":sample[‘context’]})", "type": "text" } ], "index": 32 }, { "bbox": [ 131, 597, 490, 612 ], "spans": [ { "bbox": [ 131, 597, 490, 612 ], "score": 1.0, "content": "messages.append({\"role\":\"assistant\", \"content\":sample[‘response’]}", "type": "text" } ], "index": 33 } ], "index": 32 }, { "type": "text", "bbox": [ 117, 621, 439, 632 ], "lines": [ { "bbox": [ 116, 619, 440, 633 ], "spans": [ { "bbox": [ 116, 619, 440, 633 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":‘\\n’.join(query)})", "type": "text" } ], "index": 34 } ], "index": 34 } ], "page_idx": 19, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 749, 313, 764 ], "spans": [ { "bbox": [ 298, 749, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 169, 503, 192 ], "lines": [ { "bbox": [ 106, 169, 505, 182 ], "spans": [ { "bbox": [ 106, 169, 505, 182 ], "score": 1.0, "content": "Table 13: In this example, we provide the prompt used to generate the reasoning response for refer-", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 180, 441, 193 ], "spans": [ { "bbox": [ 106, 180, 441, 193 ], "score": 1.0, "content": "and-ground instruction tuning, following the practice of LLaVA (Liu et al., 2023b).", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 106, 169, 505, 193 ] }, { "type": "text", "bbox": [ 119, 210, 478, 301 ], "lines": [ { "bbox": [ 116, 210, 479, 222 ], "spans": [ { "bbox": [ 116, 210, 162, 221 ], "score": 1.0, "content": "messages", "type": "text" }, { "bbox": [ 163, 210, 218, 222 ], "score": 0.3, "content": "\\mathbf { \\Sigma } = [ \\mathbf { \\Sigma } \\{ { \\ \" } \\Sigma \\circ \\mathrm { { 1 } } \\mathsf { e \" \\Sigma }", "type": "inline_equation" }, { "bbox": [ 218, 210, 479, 221 ], "score": 1.0, "content": ":\"system\", \"content\": f”’You are an AI visual assistant that", "type": "text" } ], "index": 2 }, { "bbox": [ 117, 220, 479, 233 ], "spans": [ { "bbox": [ 117, 220, 479, 233 ], "score": 1.0, "content": "can analyze a single image. You receive five global captions, each describing the same image you", "type": "text" } ], "index": 3 }, { "bbox": [ 117, 230, 479, 242 ], "spans": [ { "bbox": [ 117, 230, 479, 242 ], "score": 1.0, "content": "are observing. In addition, specific object locations within the image are given, along with detailed", "type": "text" } ], "index": 4 }, { "bbox": [ 117, 240, 479, 253 ], "spans": [ { "bbox": [ 117, 240, 479, 253 ], "score": 1.0, "content": "coordinates. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2)", "type": "text" } ], "index": 5 }, { "bbox": [ 117, 249, 479, 264 ], "spans": [ { "bbox": [ 117, 249, 479, 264 ], "score": 1.0, "content": "with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y,", "type": "text" } ], "index": 6 }, { "bbox": [ 118, 261, 479, 272 ], "spans": [ { "bbox": [ 118, 261, 479, 272 ], "score": 1.0, "content": "bottom right x, and bottom right y. Also, the relationships between pairs of objects are provided, in", "type": "text" } ], "index": 7 }, { "bbox": [ 117, 270, 479, 282 ], "spans": [ { "bbox": [ 117, 270, 194, 282 ], "score": 1.0, "content": "the format of object", "type": "text" }, { "bbox": [ 194, 272, 206, 280 ], "score": 0.83, "content": "", "type": "inline_equation" }, { "bbox": [ 207, 270, 252, 282 ], "score": 1.0, "content": "relationship", "type": "text" }, { "bbox": [ 252, 272, 264, 280 ], "score": 0.83, "content": "", "type": "inline_equation" }, { "bbox": [ 264, 270, 479, 282 ], "score": 1.0, "content": "subject, where the object/subject are indexed by object id", "type": "text" } ], "index": 8 }, { "bbox": [ 117, 280, 479, 293 ], "spans": [ { "bbox": [ 117, 280, 479, 293 ], "score": 1.0, "content": "from previous object lists as well as the object names. Also, several region descriptions are given,", "type": "text" } ], "index": 9 }, { "bbox": [ 117, 290, 368, 302 ], "spans": [ { "bbox": [ 117, 290, 368, 302 ], "score": 1.0, "content": "each describing a box region of the image, with detailed coordinates.", "type": "text" } ], "index": 10 } ], "index": 6, "bbox_fs": [ 116, 210, 479, 302 ] }, { "type": "text", "bbox": [ 119, 310, 477, 340 ], "lines": [ { "bbox": [ 117, 308, 478, 323 ], "spans": [ { "bbox": [ 117, 308, 478, 323 ], "score": 1.0, "content": "The task is to use the provided image information (objects, attribute, relationship, region description,", "type": "text" } ], "index": 11 }, { "bbox": [ 117, 320, 479, 333 ], "spans": [ { "bbox": [ 117, 320, 479, 333 ], "score": 1.0, "content": "captions), create a plausible and challenging question about the image, and provide the answer in", "type": "text" } ], "index": 12 }, { "bbox": [ 117, 329, 144, 342 ], "spans": [ { "bbox": [ 117, 329, 144, 342 ], "score": 1.0, "content": "detail.", "type": "text" } ], "index": 13 } ], "index": 12, "bbox_fs": [ 117, 308, 479, 342 ] }, { "type": "text", "bbox": [ 118, 350, 477, 370 ], "lines": [ { "bbox": [ 117, 349, 478, 362 ], "spans": [ { "bbox": [ 117, 349, 478, 362 ], "score": 1.0, "content": "Create complex questions that mention specific regions of the image, but the question should require", "type": "text" } ], "index": 14 }, { "bbox": [ 117, 359, 450, 372 ], "spans": [ { "bbox": [ 117, 359, 450, 372 ], "score": 1.0, "content": "some knowledge-aware or high-level commonsense reasoning beyond describing the scene.", "type": "text" } ], "index": 15 } ], "index": 14.5, "bbox_fs": [ 117, 349, 478, 372 ] }, { "type": "text", "bbox": [ 119, 380, 478, 420 ], "lines": [ { "bbox": [ 118, 379, 479, 392 ], "spans": [ { "bbox": [ 118, 379, 479, 392 ], "score": 1.0, "content": "To answer such questions, one should first understand the visual content, then based on the", "type": "text" } ], "index": 16 }, { "bbox": [ 117, 390, 479, 402 ], "spans": [ { "bbox": [ 117, 390, 479, 402 ], "score": 1.0, "content": "background knowledge or reasoning, either explain why the things are happening that way or", "type": "text" } ], "index": 17 }, { "bbox": [ 117, 399, 479, 412 ], "spans": [ { "bbox": [ 117, 399, 479, 412 ], "score": 1.0, "content": "provide guides and help to the user’s request. Make the question challenging by not including the", "type": "text" } ], "index": 18 }, { "bbox": [ 118, 410, 422, 421 ], "spans": [ { "bbox": [ 118, 410, 422, 421 ], "score": 1.0, "content": "visual content details in the question so that the user needs to reason about that first.", "type": "text" } ], "index": 19 } ], "index": 17.5, "bbox_fs": [ 117, 379, 479, 421 ] }, { "type": "text", "bbox": [ 119, 430, 403, 440 ], "lines": [ { "bbox": [ 117, 429, 404, 441 ], "spans": [ { "bbox": [ 117, 429, 404, 441 ], "score": 1.0, "content": "Here are some additional requirements about generated questions and answers:", "type": "text" } ], "index": 20 } ], "index": 20, "bbox_fs": [ 117, 429, 404, 441 ] }, { "type": "text", "bbox": [ 119, 450, 478, 490 ], "lines": [ { "bbox": [ 118, 449, 479, 461 ], "spans": [ { "bbox": [ 118, 449, 479, 461 ], "score": 1.0, "content": "1. In question or answer, you must mention bounding box coordinates to refer to the object or", "type": "text" } ], "index": 21 }, { "bbox": [ 117, 460, 479, 472 ], "spans": [ { "bbox": [ 117, 460, 479, 472 ], "score": 1.0, "content": "regions, instead of directly say the object name or describing the regions in text. In answers, explain", "type": "text" } ], "index": 22 }, { "bbox": [ 118, 470, 479, 482 ], "spans": [ { "bbox": [ 118, 470, 479, 482 ], "score": 1.0, "content": "the region in the context of scene. Include details like object counts, position of the objects, relative", "type": "text" } ], "index": 23 }, { "bbox": [ 118, 480, 225, 491 ], "spans": [ { "bbox": [ 118, 480, 225, 491 ], "score": 1.0, "content": "position between the objects.", "type": "text" } ], "index": 24 } ], "index": 22.5, "bbox_fs": [ 117, 449, 479, 491 ] }, { "type": "text", "bbox": [ 114, 500, 478, 520 ], "lines": [ { "bbox": [ 115, 498, 479, 513 ], "spans": [ { "bbox": [ 115, 498, 479, 513 ], "score": 1.0, "content": "2. Don’t ask the question you are not confident to answer. Only include question that have definite", "type": "text" } ], "index": 25 }, { "bbox": [ 117, 510, 148, 520 ], "spans": [ { "bbox": [ 117, 510, 148, 520 ], "score": 1.0, "content": "answer.", "type": "text" } ], "index": 26 } ], "index": 25.5, "bbox_fs": [ 115, 498, 479, 520 ] }, { "type": "text", "bbox": [ 116, 529, 477, 550 ], "lines": [ { "bbox": [ 115, 527, 478, 542 ], "spans": [ { "bbox": [ 115, 527, 478, 542 ], "score": 1.0, "content": "3. Do not mention that the information source is provided in text/catpion/region description. Always", "type": "text" } ], "index": 27 }, { "bbox": [ 116, 538, 301, 552 ], "spans": [ { "bbox": [ 116, 538, 301, 552 ], "score": 1.0, "content": "answer as if you are directly looking at the image.", "type": "text" } ], "index": 28 } ], "index": 27.5, "bbox_fs": [ 115, 527, 478, 552 ] }, { "type": "text", "bbox": [ 117, 559, 456, 580 ], "lines": [ { "bbox": [ 115, 557, 457, 572 ], "spans": [ { "bbox": [ 115, 557, 457, 572 ], "score": 1.0, "content": "4. Make the question as diverse as possible and as complex-reasoning required as possible.”’}", "type": "text" } ], "index": 29 }, { "bbox": [ 117, 570, 124, 578 ], "spans": [ { "bbox": [ 117, 570, 124, 578 ], "score": 1.0, "content": "]", "type": "text" } ], "index": 30 } ], "index": 29.5, "bbox_fs": [ 115, 557, 457, 578 ] }, { "type": "text", "bbox": [ 124, 581, 490, 619 ], "lines": [ { "bbox": [ 123, 579, 281, 590 ], "spans": [ { "bbox": [ 123, 581, 137, 589 ], "score": 1.0, "content": "or", "type": "text" }, { "bbox": [ 141, 580, 178, 590 ], "score": 1.0, "content": "sample", "type": "text" }, { "bbox": [ 179, 581, 193, 589 ], "score": 1.0, "content": "in", "type": "text" }, { "bbox": [ 195, 579, 281, 590 ], "score": 1.0, "content": "fewshot samples:", "type": "text" } ], "index": 31 }, { "bbox": [ 131, 587, 462, 602 ], "spans": [ { "bbox": [ 131, 587, 462, 602 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":sample[‘context’]})", "type": "text" } ], "index": 32 }, { "bbox": [ 131, 597, 490, 612 ], "spans": [ { "bbox": [ 131, 597, 490, 612 ], "score": 1.0, "content": "messages.append({\"role\":\"assistant\", \"content\":sample[‘response’]}", "type": "text" } ], "index": 33 }, { "bbox": [ 116, 619, 440, 633 ], "spans": [ { "bbox": [ 116, 619, 440, 633 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":‘\\n’.join(query)})", "type": "text" } ], "index": 34 } ], "index": 32, "bbox_fs": [ 123, 579, 490, 612 ] }, { "type": "text", "bbox": [ 117, 621, 439, 632 ], "lines": [], "index": 34, "bbox_fs": [ 116, 619, 440, 633 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 80, 505, 102 ], "lines": [ { "bbox": [ 106, 80, 505, 93 ], "spans": [ { "bbox": [ 106, 80, 505, 93 ], "score": 1.0, "content": "Table 14: One example used in in-context learning to construct GPT-Assisted Refer-and-Ground", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 90, 497, 105 ], "spans": [ { "bbox": [ 105, 90, 497, 105 ], "score": 1.0, "content": "Instructon-Tuning. We encourage readers to refer to the codebase for the prompts for other types.", "type": "text" } ], "index": 1 } ], "index": 0.5 }, { "type": "image", "bbox": [ 128, 113, 497, 318 ], "blocks": [ { "type": "image_body", "bbox": [ 128, 113, 497, 318 ], "group_id": 0, "lines": [ { "bbox": [ 128, 113, 497, 318 ], "spans": [ { "bbox": [ 128, 113, 497, 318 ], "score": 0.366, "type": "image", "image_path": "0e8156dbb8eadd840c8b89f6200ec25e9d9dbc2a31a688f0a675b058d6961a2c.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 128, 113, 497, 181.33333333333331 ], "spans": [], "index": 2 }, { "bbox": [ 128, 181.33333333333331, 497, 249.66666666666663 ], "spans": [], "index": 3 }, { "bbox": [ 128, 249.66666666666663, 497, 317.99999999999994 ], "spans": [], "index": 4 } ] } ], "index": 3 }, { "type": "title", "bbox": [ 107, 344, 372, 358 ], "lines": [ { "bbox": [ 106, 344, 373, 359 ], "spans": [ { "bbox": [ 106, 344, 373, 359 ], "score": 1.0, "content": "D EXAMPLES AND PROMPTS FOR FERRET-BENCH", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 107, 370, 505, 425 ], "lines": [ { "bbox": [ 106, 369, 505, 383 ], "spans": [ { "bbox": [ 106, 369, 505, 383 ], "score": 1.0, "content": "We leverage GPT-4 to generate three kinds of region-based questions evaluating referring and", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 381, 505, 394 ], "spans": [ { "bbox": [ 105, 381, 505, 394 ], "score": 1.0, "content": "grounding capability: (i) Referring Description, (ii) Referring Reasoning, and (iii) Grounding in", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 392, 505, 405 ], "spans": [ { "bbox": [ 105, 392, 505, 405 ], "score": 1.0, "content": "Conversation. Here, we only provide the prompt in Table 15 used to generate the referring descrip-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 403, 505, 415 ], "spans": [ { "bbox": [ 105, 403, 505, 415 ], "score": 1.0, "content": "tion response. One example of GPT-4 answers is shown in Table 16. We recommend readers check", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 414, 248, 426 ], "spans": [ { "bbox": [ 105, 414, 248, 426 ], "score": 1.0, "content": "out more examples in Appendix E.", "type": "text" } ], "index": 10 } ], "index": 8 } ], "page_idx": 20, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 310, 760 ], "lines": [ { "bbox": [ 298, 750, 312, 765 ], "spans": [ { "bbox": [ 298, 750, 312, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 80, 505, 102 ], "lines": [ { "bbox": [ 106, 80, 505, 93 ], "spans": [ { "bbox": [ 106, 80, 505, 93 ], "score": 1.0, "content": "Table 14: One example used in in-context learning to construct GPT-Assisted Refer-and-Ground", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 90, 497, 105 ], "spans": [ { "bbox": [ 105, 90, 497, 105 ], "score": 1.0, "content": "Instructon-Tuning. We encourage readers to refer to the codebase for the prompts for other types.", "type": "text" } ], "index": 1 } ], "index": 0.5, "bbox_fs": [ 105, 80, 505, 105 ] }, { "type": "image", "bbox": [ 128, 113, 497, 318 ], "blocks": [ { "type": "image_body", "bbox": [ 128, 113, 497, 318 ], "group_id": 0, "lines": [ { "bbox": [ 128, 113, 497, 318 ], "spans": [ { "bbox": [ 128, 113, 497, 318 ], "score": 0.366, "type": "image", "image_path": "0e8156dbb8eadd840c8b89f6200ec25e9d9dbc2a31a688f0a675b058d6961a2c.jpg" } ] } ], "index": 3, "virtual_lines": [ { "bbox": [ 128, 113, 497, 181.33333333333331 ], "spans": [], "index": 2 }, { "bbox": [ 128, 181.33333333333331, 497, 249.66666666666663 ], "spans": [], "index": 3 }, { "bbox": [ 128, 249.66666666666663, 497, 317.99999999999994 ], "spans": [], "index": 4 } ] } ], "index": 3 }, { "type": "title", "bbox": [ 107, 344, 372, 358 ], "lines": [ { "bbox": [ 106, 344, 373, 359 ], "spans": [ { "bbox": [ 106, 344, 373, 359 ], "score": 1.0, "content": "D EXAMPLES AND PROMPTS FOR FERRET-BENCH", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 107, 370, 505, 425 ], "lines": [ { "bbox": [ 106, 369, 505, 383 ], "spans": [ { "bbox": [ 106, 369, 505, 383 ], "score": 1.0, "content": "We leverage GPT-4 to generate three kinds of region-based questions evaluating referring and", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 381, 505, 394 ], "spans": [ { "bbox": [ 105, 381, 505, 394 ], "score": 1.0, "content": "grounding capability: (i) Referring Description, (ii) Referring Reasoning, and (iii) Grounding in", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 392, 505, 405 ], "spans": [ { "bbox": [ 105, 392, 505, 405 ], "score": 1.0, "content": "Conversation. Here, we only provide the prompt in Table 15 used to generate the referring descrip-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 403, 505, 415 ], "spans": [ { "bbox": [ 105, 403, 505, 415 ], "score": 1.0, "content": "tion response. One example of GPT-4 answers is shown in Table 16. We recommend readers check", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 414, 248, 426 ], "spans": [ { "bbox": [ 105, 414, 248, 426 ], "score": 1.0, "content": "out more examples in Appendix E.", "type": "text" } ], "index": 10 } ], "index": 8, "bbox_fs": [ 105, 369, 505, 426 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 105, 174, 503, 186 ], "lines": [ { "bbox": [ 105, 172, 505, 189 ], "spans": [ { "bbox": [ 105, 172, 505, 189 ], "score": 1.0, "content": "Table 15: In this example, we provide the prompt used to generate the referring description response.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 119, 205, 478, 295 ], "lines": [ { "bbox": [ 116, 205, 480, 216 ], "spans": [ { "bbox": [ 116, 205, 163, 216 ], "score": 1.0, "content": "messages", "type": "text" }, { "bbox": [ 164, 206, 173, 214 ], "score": 0.42, "content": "=", "type": "inline_equation" }, { "bbox": [ 173, 205, 480, 216 ], "score": 1.0, "content": "[ {\"role\":\"system\", \"content\": f”’You are an AI visual assistant that", "type": "text" } ], "index": 1 }, { "bbox": [ 117, 214, 479, 227 ], "spans": [ { "bbox": [ 117, 214, 479, 227 ], "score": 1.0, "content": "can analyze a single image. You receive five global captions, each describing the same image you", "type": "text" } ], "index": 2 }, { "bbox": [ 117, 225, 479, 237 ], "spans": [ { "bbox": [ 117, 225, 479, 237 ], "score": 1.0, "content": "are observing. In addition, specific object locations within the image are given, along with detailed", "type": "text" } ], "index": 3 }, { "bbox": [ 117, 235, 479, 247 ], "spans": [ { "bbox": [ 117, 235, 479, 247 ], "score": 1.0, "content": "coordinates. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2)", "type": "text" } ], "index": 4 }, { "bbox": [ 117, 244, 479, 257 ], "spans": [ { "bbox": [ 117, 244, 479, 257 ], "score": 1.0, "content": "with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y,", "type": "text" } ], "index": 5 }, { "bbox": [ 117, 254, 479, 267 ], "spans": [ { "bbox": [ 117, 254, 165, 267 ], "score": 1.0, "content": "bottom right", "type": "text" }, { "bbox": [ 165, 257, 171, 264 ], "score": 0.34, "content": "\\mathbf { X }", "type": "inline_equation" }, { "bbox": [ 172, 254, 479, 267 ], "score": 1.0, "content": ", and bottom right y. Also, the relationships between pairs of objects are provided, in", "type": "text" } ], "index": 6 }, { "bbox": [ 117, 265, 479, 277 ], "spans": [ { "bbox": [ 117, 265, 194, 277 ], "score": 1.0, "content": "the format of object", "type": "text" }, { "bbox": [ 194, 266, 207, 274 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 207, 265, 252, 277 ], "score": 1.0, "content": "relationship", "type": "text" }, { "bbox": [ 252, 266, 264, 275 ], "score": 0.83, "content": "", "type": "inline_equation" }, { "bbox": [ 264, 265, 479, 277 ], "score": 1.0, "content": "subject, where the object/subject are indexed by object id", "type": "text" } ], "index": 7 }, { "bbox": [ 117, 274, 479, 287 ], "spans": [ { "bbox": [ 117, 274, 479, 287 ], "score": 1.0, "content": "from previous object lists as well as the object names. Also, several region description are given,", "type": "text" } ], "index": 8 }, { "bbox": [ 117, 285, 355, 297 ], "spans": [ { "bbox": [ 117, 285, 355, 297 ], "score": 1.0, "content": "each describing a box region of image, with detailed coordinates.", "type": "text" } ], "index": 9 } ], "index": 5 }, { "type": "text", "bbox": [ 118, 305, 478, 335 ], "lines": [ { "bbox": [ 117, 303, 478, 317 ], "spans": [ { "bbox": [ 117, 303, 478, 317 ], "score": 1.0, "content": "The task is to use the provided image information (objects, attribute, relationship, region description,", "type": "text" } ], "index": 10 }, { "bbox": [ 117, 314, 479, 327 ], "spans": [ { "bbox": [ 117, 314, 479, 327 ], "score": 1.0, "content": "captions), create a plausible and challenging question about the image, and provide the answer in", "type": "text" } ], "index": 11 }, { "bbox": [ 117, 324, 144, 336 ], "spans": [ { "bbox": [ 117, 324, 144, 336 ], "score": 1.0, "content": "detail.", "type": "text" } ], "index": 12 } ], "index": 11 }, { "type": "text", "bbox": [ 119, 344, 477, 365 ], "lines": [ { "bbox": [ 118, 344, 478, 356 ], "spans": [ { "bbox": [ 118, 344, 478, 356 ], "score": 1.0, "content": "Create questions that refer to coordinates of some objects or regions without describing it, and ask", "type": "text" } ], "index": 13 }, { "bbox": [ 117, 353, 313, 366 ], "spans": [ { "bbox": [ 117, 353, 313, 366 ], "score": 1.0, "content": "about its interaction with surrounding/nearby objects.", "type": "text" } ], "index": 14 } ], "index": 13.5 }, { "type": "text", "bbox": [ 119, 374, 477, 395 ], "lines": [ { "bbox": [ 118, 374, 478, 386 ], "spans": [ { "bbox": [ 118, 374, 478, 386 ], "score": 1.0, "content": "To answer such questions, one should require first understanding the visual content, then based on", "type": "text" } ], "index": 15 }, { "bbox": [ 118, 384, 238, 396 ], "spans": [ { "bbox": [ 118, 384, 238, 396 ], "score": 1.0, "content": "the spatial information provided.", "type": "text" } ], "index": 16 } ], "index": 15.5 }, { "type": "text", "bbox": [ 119, 404, 403, 415 ], "lines": [ { "bbox": [ 117, 403, 404, 415 ], "spans": [ { "bbox": [ 117, 403, 404, 415 ], "score": 1.0, "content": "Here are some additional requirements about generated questions and answers:", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 119, 424, 478, 465 ], "lines": [ { "bbox": [ 118, 424, 479, 436 ], "spans": [ { "bbox": [ 118, 424, 479, 436 ], "score": 1.0, "content": "1. In question, you must mention bounding box coordinates to refer to the object or regions, instead", "type": "text" } ], "index": 18 }, { "bbox": [ 117, 433, 479, 447 ], "spans": [ { "bbox": [ 117, 433, 479, 447 ], "score": 1.0, "content": "of directly say the object name or describing the regions in text. In answers, explain the region in", "type": "text" } ], "index": 19 }, { "bbox": [ 117, 443, 479, 457 ], "spans": [ { "bbox": [ 117, 443, 479, 457 ], "score": 1.0, "content": "the context of scene. Include details like object counts, position of the objects, relative position", "type": "text" } ], "index": 20 }, { "bbox": [ 117, 453, 195, 467 ], "spans": [ { "bbox": [ 117, 453, 195, 467 ], "score": 1.0, "content": "between the objects.", "type": "text" } ], "index": 21 } ], "index": 19.5 }, { "type": "text", "bbox": [ 117, 474, 478, 495 ], "lines": [ { "bbox": [ 116, 473, 479, 487 ], "spans": [ { "bbox": [ 116, 473, 479, 487 ], "score": 1.0, "content": "2. Don’t ask the question you are not confident to answer. Only include question that have definite", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 485, 147, 495 ], "spans": [ { "bbox": [ 116, 485, 147, 495 ], "score": 1.0, "content": "answer.", "type": "text" } ], "index": 23 } ], "index": 22.5 }, { "type": "text", "bbox": [ 118, 504, 478, 525 ], "lines": [ { "bbox": [ 116, 502, 479, 517 ], "spans": [ { "bbox": [ 116, 502, 479, 517 ], "score": 1.0, "content": "3. Do not mention that the information source is provided in text/catpion/region description. Always", "type": "text" } ], "index": 24 }, { "bbox": [ 116, 512, 300, 527 ], "spans": [ { "bbox": [ 116, 512, 300, 527 ], "score": 1.0, "content": "answer as if you are directly looking at the image.", "type": "text" } ], "index": 25 } ], "index": 24.5 }, { "type": "text", "bbox": [ 118, 534, 315, 545 ], "lines": [ { "bbox": [ 117, 533, 317, 545 ], "spans": [ { "bbox": [ 117, 533, 317, 545 ], "score": 1.0, "content": "4. Don’t mention additional coordinates in the answer.", "type": "text" } ], "index": 26 } ], "index": 26 }, { "type": "text", "bbox": [ 117, 554, 472, 574 ], "lines": [ { "bbox": [ 115, 551, 473, 566 ], "spans": [ { "bbox": [ 115, 551, 473, 566 ], "score": 1.0, "content": "5. Question should be explicitly ask about context/surrounding/nearby information/interaction.”’}", "type": "text" } ], "index": 27 } ], "index": 27 }, { "type": "text", "bbox": [ 123, 575, 492, 614 ], "lines": [ { "bbox": [ 121, 574, 281, 585 ], "spans": [ { "bbox": [ 121, 574, 138, 584 ], "score": 1.0, "content": "for", "type": "text" }, { "bbox": [ 141, 574, 179, 585 ], "score": 1.0, "content": "sample", "type": "text" }, { "bbox": [ 180, 574, 194, 584 ], "score": 1.0, "content": "in", "type": "text" }, { "bbox": [ 195, 574, 281, 585 ], "score": 1.0, "content": "fewshot samples:", "type": "text" } ], "index": 28 }, { "bbox": [ 131, 582, 462, 596 ], "spans": [ { "bbox": [ 131, 582, 462, 596 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":sample[‘context’]})", "type": "text" } ], "index": 29 }, { "bbox": [ 131, 591, 490, 606 ], "spans": [ { "bbox": [ 131, 591, 490, 606 ], "score": 1.0, "content": "messages.append({\"role\":\"assistant\", \"content\":sample[‘response’]}", "type": "text" } ], "index": 30 } ], "index": 29 }, { "type": "text", "bbox": [ 118, 615, 439, 626 ], "lines": [ { "bbox": [ 116, 613, 441, 628 ], "spans": [ { "bbox": [ 116, 613, 441, 628 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":‘\\n’.join(query)})", "type": "text" } ], "index": 31 } ], "index": 31 } ], "page_idx": 21, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 765 ], "spans": [ { "bbox": [ 298, 750, 313, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 105, 174, 503, 186 ], "lines": [ { "bbox": [ 105, 172, 505, 189 ], "spans": [ { "bbox": [ 105, 172, 505, 189 ], "score": 1.0, "content": "Table 15: In this example, we provide the prompt used to generate the referring description response.", "type": "text" } ], "index": 0 } ], "index": 0, "bbox_fs": [ 105, 172, 505, 189 ] }, { "type": "text", "bbox": [ 119, 205, 478, 295 ], "lines": [ { "bbox": [ 116, 205, 480, 216 ], "spans": [ { "bbox": [ 116, 205, 163, 216 ], "score": 1.0, "content": "messages", "type": "text" }, { "bbox": [ 164, 206, 173, 214 ], "score": 0.42, "content": "=", "type": "inline_equation" }, { "bbox": [ 173, 205, 480, 216 ], "score": 1.0, "content": "[ {\"role\":\"system\", \"content\": f”’You are an AI visual assistant that", "type": "text" } ], "index": 1 }, { "bbox": [ 117, 214, 479, 227 ], "spans": [ { "bbox": [ 117, 214, 479, 227 ], "score": 1.0, "content": "can analyze a single image. You receive five global captions, each describing the same image you", "type": "text" } ], "index": 2 }, { "bbox": [ 117, 225, 479, 237 ], "spans": [ { "bbox": [ 117, 225, 479, 237 ], "score": 1.0, "content": "are observing. In addition, specific object locations within the image are given, along with detailed", "type": "text" } ], "index": 3 }, { "bbox": [ 117, 235, 479, 247 ], "spans": [ { "bbox": [ 117, 235, 479, 247 ], "score": 1.0, "content": "coordinates. These coordinates are in the form of bounding boxes, represented as (x1, y1, x2, y2)", "type": "text" } ], "index": 4 }, { "bbox": [ 117, 244, 479, 257 ], "spans": [ { "bbox": [ 117, 244, 479, 257 ], "score": 1.0, "content": "with floating numbers ranging from 0 to 1. These values correspond to the top left x, top left y,", "type": "text" } ], "index": 5 }, { "bbox": [ 117, 254, 479, 267 ], "spans": [ { "bbox": [ 117, 254, 165, 267 ], "score": 1.0, "content": "bottom right", "type": "text" }, { "bbox": [ 165, 257, 171, 264 ], "score": 0.34, "content": "\\mathbf { X }", "type": "inline_equation" }, { "bbox": [ 172, 254, 479, 267 ], "score": 1.0, "content": ", and bottom right y. Also, the relationships between pairs of objects are provided, in", "type": "text" } ], "index": 6 }, { "bbox": [ 117, 265, 479, 277 ], "spans": [ { "bbox": [ 117, 265, 194, 277 ], "score": 1.0, "content": "the format of object", "type": "text" }, { "bbox": [ 194, 266, 207, 274 ], "score": 0.82, "content": "", "type": "inline_equation" }, { "bbox": [ 207, 265, 252, 277 ], "score": 1.0, "content": "relationship", "type": "text" }, { "bbox": [ 252, 266, 264, 275 ], "score": 0.83, "content": "", "type": "inline_equation" }, { "bbox": [ 264, 265, 479, 277 ], "score": 1.0, "content": "subject, where the object/subject are indexed by object id", "type": "text" } ], "index": 7 }, { "bbox": [ 117, 274, 479, 287 ], "spans": [ { "bbox": [ 117, 274, 479, 287 ], "score": 1.0, "content": "from previous object lists as well as the object names. Also, several region description are given,", "type": "text" } ], "index": 8 }, { "bbox": [ 117, 285, 355, 297 ], "spans": [ { "bbox": [ 117, 285, 355, 297 ], "score": 1.0, "content": "each describing a box region of image, with detailed coordinates.", "type": "text" } ], "index": 9 } ], "index": 5, "bbox_fs": [ 116, 205, 480, 297 ] }, { "type": "text", "bbox": [ 118, 305, 478, 335 ], "lines": [ { "bbox": [ 117, 303, 478, 317 ], "spans": [ { "bbox": [ 117, 303, 478, 317 ], "score": 1.0, "content": "The task is to use the provided image information (objects, attribute, relationship, region description,", "type": "text" } ], "index": 10 }, { "bbox": [ 117, 314, 479, 327 ], "spans": [ { "bbox": [ 117, 314, 479, 327 ], "score": 1.0, "content": "captions), create a plausible and challenging question about the image, and provide the answer in", "type": "text" } ], "index": 11 }, { "bbox": [ 117, 324, 144, 336 ], "spans": [ { "bbox": [ 117, 324, 144, 336 ], "score": 1.0, "content": "detail.", "type": "text" } ], "index": 12 } ], "index": 11, "bbox_fs": [ 117, 303, 479, 336 ] }, { "type": "text", "bbox": [ 119, 344, 477, 365 ], "lines": [ { "bbox": [ 118, 344, 478, 356 ], "spans": [ { "bbox": [ 118, 344, 478, 356 ], "score": 1.0, "content": "Create questions that refer to coordinates of some objects or regions without describing it, and ask", "type": "text" } ], "index": 13 }, { "bbox": [ 117, 353, 313, 366 ], "spans": [ { "bbox": [ 117, 353, 313, 366 ], "score": 1.0, "content": "about its interaction with surrounding/nearby objects.", "type": "text" } ], "index": 14 } ], "index": 13.5, "bbox_fs": [ 117, 344, 478, 366 ] }, { "type": "text", "bbox": [ 119, 374, 477, 395 ], "lines": [ { "bbox": [ 118, 374, 478, 386 ], "spans": [ { "bbox": [ 118, 374, 478, 386 ], "score": 1.0, "content": "To answer such questions, one should require first understanding the visual content, then based on", "type": "text" } ], "index": 15 }, { "bbox": [ 118, 384, 238, 396 ], "spans": [ { "bbox": [ 118, 384, 238, 396 ], "score": 1.0, "content": "the spatial information provided.", "type": "text" } ], "index": 16 } ], "index": 15.5, "bbox_fs": [ 118, 374, 478, 396 ] }, { "type": "text", "bbox": [ 119, 404, 403, 415 ], "lines": [ { "bbox": [ 117, 403, 404, 415 ], "spans": [ { "bbox": [ 117, 403, 404, 415 ], "score": 1.0, "content": "Here are some additional requirements about generated questions and answers:", "type": "text" } ], "index": 17 } ], "index": 17, "bbox_fs": [ 117, 403, 404, 415 ] }, { "type": "text", "bbox": [ 119, 424, 478, 465 ], "lines": [ { "bbox": [ 118, 424, 479, 436 ], "spans": [ { "bbox": [ 118, 424, 479, 436 ], "score": 1.0, "content": "1. In question, you must mention bounding box coordinates to refer to the object or regions, instead", "type": "text" } ], "index": 18 }, { "bbox": [ 117, 433, 479, 447 ], "spans": [ { "bbox": [ 117, 433, 479, 447 ], "score": 1.0, "content": "of directly say the object name or describing the regions in text. In answers, explain the region in", "type": "text" } ], "index": 19 }, { "bbox": [ 117, 443, 479, 457 ], "spans": [ { "bbox": [ 117, 443, 479, 457 ], "score": 1.0, "content": "the context of scene. Include details like object counts, position of the objects, relative position", "type": "text" } ], "index": 20 }, { "bbox": [ 117, 453, 195, 467 ], "spans": [ { "bbox": [ 117, 453, 195, 467 ], "score": 1.0, "content": "between the objects.", "type": "text" } ], "index": 21 } ], "index": 19.5, "bbox_fs": [ 117, 424, 479, 467 ] }, { "type": "text", "bbox": [ 117, 474, 478, 495 ], "lines": [ { "bbox": [ 116, 473, 479, 487 ], "spans": [ { "bbox": [ 116, 473, 479, 487 ], "score": 1.0, "content": "2. Don’t ask the question you are not confident to answer. Only include question that have definite", "type": "text" } ], "index": 22 }, { "bbox": [ 116, 485, 147, 495 ], "spans": [ { "bbox": [ 116, 485, 147, 495 ], "score": 1.0, "content": "answer.", "type": "text" } ], "index": 23 } ], "index": 22.5, "bbox_fs": [ 116, 473, 479, 495 ] }, { "type": "text", "bbox": [ 118, 504, 478, 525 ], "lines": [ { "bbox": [ 116, 502, 479, 517 ], "spans": [ { "bbox": [ 116, 502, 479, 517 ], "score": 1.0, "content": "3. Do not mention that the information source is provided in text/catpion/region description. Always", "type": "text" } ], "index": 24 }, { "bbox": [ 116, 512, 300, 527 ], "spans": [ { "bbox": [ 116, 512, 300, 527 ], "score": 1.0, "content": "answer as if you are directly looking at the image.", "type": "text" } ], "index": 25 } ], "index": 24.5, "bbox_fs": [ 116, 502, 479, 527 ] }, { "type": "text", "bbox": [ 118, 534, 315, 545 ], "lines": [ { "bbox": [ 117, 533, 317, 545 ], "spans": [ { "bbox": [ 117, 533, 317, 545 ], "score": 1.0, "content": "4. Don’t mention additional coordinates in the answer.", "type": "text" } ], "index": 26 } ], "index": 26, "bbox_fs": [ 117, 533, 317, 545 ] }, { "type": "text", "bbox": [ 117, 554, 472, 574 ], "lines": [ { "bbox": [ 115, 551, 473, 566 ], "spans": [ { "bbox": [ 115, 551, 473, 566 ], "score": 1.0, "content": "5. Question should be explicitly ask about context/surrounding/nearby information/interaction.”’}", "type": "text" } ], "index": 27 }, { "bbox": [ 121, 574, 281, 585 ], "spans": [ { "bbox": [ 121, 574, 138, 584 ], "score": 1.0, "content": "for", "type": "text" }, { "bbox": [ 141, 574, 179, 585 ], "score": 1.0, "content": "sample", "type": "text" }, { "bbox": [ 180, 574, 194, 584 ], "score": 1.0, "content": "in", "type": "text" }, { "bbox": [ 195, 574, 281, 585 ], "score": 1.0, "content": "fewshot samples:", "type": "text" } ], "index": 28 }, { "bbox": [ 131, 582, 462, 596 ], "spans": [ { "bbox": [ 131, 582, 462, 596 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":sample[‘context’]})", "type": "text" } ], "index": 29 }, { "bbox": [ 131, 591, 490, 606 ], "spans": [ { "bbox": [ 131, 591, 490, 606 ], "score": 1.0, "content": "messages.append({\"role\":\"assistant\", \"content\":sample[‘response’]}", "type": "text" } ], "index": 30 }, { "bbox": [ 116, 613, 441, 628 ], "spans": [ { "bbox": [ 116, 613, 441, 628 ], "score": 1.0, "content": "messages.append({\"role\":\"user\", \"content\":‘\\n’.join(query)})", "type": "text" } ], "index": 31 } ], "index": 27, "bbox_fs": [ 115, 551, 473, 566 ] }, { "type": "text", "bbox": [ 123, 575, 492, 614 ], "lines": [], "index": 29, "bbox_fs": [ 121, 574, 490, 606 ], "lines_deleted": true }, { "type": "text", "bbox": [ 118, 615, 439, 626 ], "lines": [], "index": 31, "bbox_fs": [ 116, 613, 441, 628 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 164, 80, 447, 92 ], "lines": [ { "bbox": [ 162, 78, 448, 95 ], "spans": [ { "bbox": [ 162, 78, 448, 95 ], "score": 1.0, "content": "Table 16: One example of GPT4’s answers about referring description.", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "title", "bbox": [ 122, 111, 163, 119 ], "lines": [ { "bbox": [ 121, 111, 163, 120 ], "spans": [ { "bbox": [ 121, 111, 163, 120 ], "score": 1.0, "content": "Conversation", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 122, 120, 150, 127 ], "lines": [ { "bbox": [ 121, 119, 151, 128 ], "spans": [ { "bbox": [ 121, 119, 151, 128 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 2 } ], "index": 2 }, { "type": "text", "bbox": [ 119, 127, 364, 143 ], "lines": [ { "bbox": [ 119, 125, 365, 139 ], "spans": [ { "bbox": [ 119, 125, 365, 139 ], "score": 1.0, "content": "Describe the region [0.102, 0.230, 0.578, 0.761] in the context of surrounding objects.", "type": "text" } ], "index": 3 }, { "bbox": [ 119, 133, 148, 145 ], "spans": [ { "bbox": [ 119, 133, 148, 145 ], "score": 1.0, "content": "Answer:", "type": "text" } ], "index": 4 } ], "index": 3.5 }, { "type": "text", "bbox": [ 120, 143, 481, 160 ], "lines": [ { "bbox": [ 119, 142, 482, 154 ], "spans": [ { "bbox": [ 119, 142, 482, 154 ], "score": 1.0, "content": "It’s a man in the region, who is standing on top of a skateboard, performing a rail grind over a metal hand rail. There is a bench", "type": "text" } ], "index": 5 }, { "bbox": [ 119, 151, 287, 161 ], "spans": [ { "bbox": [ 119, 151, 287, 161 ], "score": 1.0, "content": "and a tree near him. Also, he is in front of some buildings.", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "image", "bbox": [ 251, 174, 351, 318 ], "blocks": [ { "type": "image_body", "bbox": [ 251, 174, 351, 318 ], "group_id": 1, "lines": [ { "bbox": [ 251, 174, 351, 318 ], "spans": [ { "bbox": [ 251, 174, 351, 318 ], "score": 0.866, "type": "image", "image_path": "003d9dbaa4cb96d2188c97db064f7e3bd61d785608dd4d1772094affd036c242.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 251, 174, 351, 318 ], "spans": [], "index": 7 } ] } ], "index": 7 }, { "type": "title", "bbox": [ 107, 361, 245, 374 ], "lines": [ { "bbox": [ 105, 360, 246, 376 ], "spans": [ { "bbox": [ 105, 360, 246, 376 ], "score": 1.0, "content": "E MORE VISUALIZATION", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 107, 386, 504, 409 ], "lines": [ { "bbox": [ 106, 385, 506, 398 ], "spans": [ { "bbox": [ 106, 385, 506, 398 ], "score": 1.0, "content": "We provide more quantitative results of the predictions under various tasks from Ferret to indicate", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 396, 249, 409 ], "spans": [ { "bbox": [ 105, 396, 249, 409 ], "score": 1.0, "content": "the model’s strength and capability.", "type": "text" } ], "index": 10 } ], "index": 9.5 }, { "type": "text", "bbox": [ 106, 417, 506, 541 ], "lines": [ { "bbox": [ 105, 416, 505, 432 ], "spans": [ { "bbox": [ 105, 416, 505, 432 ], "score": 1.0, "content": "• Please refer to Figure 5 for Referring Object Classification on LVIS with different referring for-", "type": "text" } ], "index": 11 }, { "bbox": [ 114, 430, 188, 441 ], "spans": [ { "bbox": [ 114, 430, 188, 441 ], "score": 1.0, "content": "mats (point/box/).", "type": "text" } ], "index": 12 }, { "bbox": [ 103, 442, 506, 459 ], "spans": [ { "bbox": [ 103, 442, 506, 459 ], "score": 1.0, "content": "• Please refer to Figure 6 for Visual Grounding on Flickr30k Entities and Referring Expression", "type": "text" } ], "index": 13 }, { "bbox": [ 112, 454, 340, 469 ], "spans": [ { "bbox": [ 112, 454, 280, 469 ], "score": 1.0, "content": "Comprehension on RefCOCO/RefCOCO", "type": "text" }, { "bbox": [ 280, 456, 286, 464 ], "score": 0.4, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 287, 454, 340, 469 ], "score": 1.0, "content": "/RefCOCOg.", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 470, 428, 482 ], "spans": [ { "bbox": [ 105, 470, 428, 482 ], "score": 1.0, "content": "• Please refer to Figure 7 for Grounded Captioning on Flickr30k Karpathy split.", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 484, 468, 498 ], "spans": [ { "bbox": [ 105, 484, 468, 498 ], "score": 1.0, "content": "• Please refer to Figure 8 for Evaluating Object Hallucination (POPE) on COCO val split.", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 498, 382, 512 ], "spans": [ { "bbox": [ 105, 498, 382, 512 ], "score": 1.0, "content": "• Please refer to Table 17 for Referring Description in Ferret-Bench.", "type": "text" } ], "index": 17 }, { "bbox": [ 104, 514, 374, 527 ], "spans": [ { "bbox": [ 104, 514, 374, 527 ], "score": 1.0, "content": "• Please refer to Table 18 for Referring Resoning in Ferret-Bench.", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 528, 403, 542 ], "spans": [ { "bbox": [ 105, 528, 403, 542 ], "score": 1.0, "content": "• Please refer to Table 19 for Grounding in Conversation in Ferret-Bench.", "type": "text" } ], "index": 19 } ], "index": 15 }, { "type": "title", "bbox": [ 111, 559, 248, 568 ], "lines": [ { "bbox": [ 111, 558, 249, 569 ], "spans": [ { "bbox": [ 111, 558, 249, 569 ], "score": 1.0, "content": "Referring Object Classification (LVIS) – point", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "image", "bbox": [ 107, 569, 551, 695 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 569, 551, 695 ], "group_id": 0, "lines": [ { "bbox": [ 107, 569, 551, 695 ], "spans": [ { "bbox": [ 107, 569, 551, 695 ], "score": 0.834, "type": "image", "image_path": "e084c90795f3cf70ed3f220ce71f783628efd1cede1582b7a3b021fa34951999.jpg" } ] } ], "index": 22, "virtual_lines": [ { "bbox": [ 107, 569, 551, 611.0 ], "spans": [], "index": 21 }, { "bbox": [ 107, 611.0, 551, 653.0 ], "spans": [], "index": 22 }, { "bbox": [ 107, 653.0, 551, 695.0 ], "spans": [], "index": 23 } ] }, { "type": "image_caption", "bbox": [ 106, 702, 507, 726 ], "group_id": 0, "lines": [ { "bbox": [ 105, 701, 506, 716 ], "spans": [ { "bbox": [ 105, 701, 506, 716 ], "score": 1.0, "content": "Figure 5: Referring Object Classification on LVIS. The task aims to classify specific region(s) in", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 713, 308, 726 ], "spans": [ { "bbox": [ 106, 713, 308, 726 ], "score": 1.0, "content": "an image given by point/box/segmentation inputs.", "type": "text" } ], "index": 25 } ], "index": 24.5 } ], "index": 23.25 } ], "page_idx": 22, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 764 ], "spans": [ { "bbox": [ 298, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 15 } ] } ] }, { "type": "discarded", "bbox": [ 325, 559, 462, 568 ], "lines": [ { "bbox": [ 324, 558, 463, 569 ], "spans": [ { "bbox": [ 324, 558, 463, 569 ], "score": 1.0, "content": "Referring Object Classification (LVIS) – point", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 164, 80, 447, 92 ], "lines": [ { "bbox": [ 162, 78, 448, 95 ], "spans": [ { "bbox": [ 162, 78, 448, 95 ], "score": 1.0, "content": "Table 16: One example of GPT4’s answers about referring description.", "type": "text" } ], "index": 0 } ], "index": 0, "bbox_fs": [ 162, 78, 448, 95 ] }, { "type": "title", "bbox": [ 122, 111, 163, 119 ], "lines": [ { "bbox": [ 121, 111, 163, 120 ], "spans": [ { "bbox": [ 121, 111, 163, 120 ], "score": 1.0, "content": "Conversation", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 122, 120, 150, 127 ], "lines": [ { "bbox": [ 121, 119, 151, 128 ], "spans": [ { "bbox": [ 121, 119, 151, 128 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 2 } ], "index": 2, "bbox_fs": [ 121, 119, 151, 128 ] }, { "type": "text", "bbox": [ 119, 127, 364, 143 ], "lines": [ { "bbox": [ 119, 125, 365, 139 ], "spans": [ { "bbox": [ 119, 125, 365, 139 ], "score": 1.0, "content": "Describe the region [0.102, 0.230, 0.578, 0.761] in the context of surrounding objects.", "type": "text" } ], "index": 3 }, { "bbox": [ 119, 133, 148, 145 ], "spans": [ { "bbox": [ 119, 133, 148, 145 ], "score": 1.0, "content": "Answer:", "type": "text" } ], "index": 4 } ], "index": 3.5, "bbox_fs": [ 119, 125, 365, 145 ] }, { "type": "text", "bbox": [ 120, 143, 481, 160 ], "lines": [ { "bbox": [ 119, 142, 482, 154 ], "spans": [ { "bbox": [ 119, 142, 482, 154 ], "score": 1.0, "content": "It’s a man in the region, who is standing on top of a skateboard, performing a rail grind over a metal hand rail. There is a bench", "type": "text" } ], "index": 5 }, { "bbox": [ 119, 151, 287, 161 ], "spans": [ { "bbox": [ 119, 151, 287, 161 ], "score": 1.0, "content": "and a tree near him. Also, he is in front of some buildings.", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 119, 142, 482, 161 ] }, { "type": "image", "bbox": [ 251, 174, 351, 318 ], "blocks": [ { "type": "image_body", "bbox": [ 251, 174, 351, 318 ], "group_id": 1, "lines": [ { "bbox": [ 251, 174, 351, 318 ], "spans": [ { "bbox": [ 251, 174, 351, 318 ], "score": 0.866, "type": "image", "image_path": "003d9dbaa4cb96d2188c97db064f7e3bd61d785608dd4d1772094affd036c242.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 251, 174, 351, 318 ], "spans": [], "index": 7 } ] } ], "index": 7 }, { "type": "title", "bbox": [ 107, 361, 245, 374 ], "lines": [ { "bbox": [ 105, 360, 246, 376 ], "spans": [ { "bbox": [ 105, 360, 246, 376 ], "score": 1.0, "content": "E MORE VISUALIZATION", "type": "text" } ], "index": 8 } ], "index": 8 }, { "type": "text", "bbox": [ 107, 386, 504, 409 ], "lines": [ { "bbox": [ 106, 385, 506, 398 ], "spans": [ { "bbox": [ 106, 385, 506, 398 ], "score": 1.0, "content": "We provide more quantitative results of the predictions under various tasks from Ferret to indicate", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 396, 249, 409 ], "spans": [ { "bbox": [ 105, 396, 249, 409 ], "score": 1.0, "content": "the model’s strength and capability.", "type": "text" } ], "index": 10 } ], "index": 9.5, "bbox_fs": [ 105, 385, 506, 409 ] }, { "type": "list", "bbox": [ 106, 417, 506, 541 ], "lines": [ { "bbox": [ 105, 416, 505, 432 ], "spans": [ { "bbox": [ 105, 416, 505, 432 ], "score": 1.0, "content": "• Please refer to Figure 5 for Referring Object Classification on LVIS with different referring for-", "type": "text" } ], "index": 11, "is_list_start_line": true }, { "bbox": [ 114, 430, 188, 441 ], "spans": [ { "bbox": [ 114, 430, 188, 441 ], "score": 1.0, "content": "mats (point/box/).", "type": "text" } ], "index": 12, "is_list_end_line": true }, { "bbox": [ 103, 442, 506, 459 ], "spans": [ { "bbox": [ 103, 442, 506, 459 ], "score": 1.0, "content": "• Please refer to Figure 6 for Visual Grounding on Flickr30k Entities and Referring Expression", "type": "text" } ], "index": 13, "is_list_start_line": true }, { "bbox": [ 112, 454, 340, 469 ], "spans": [ { "bbox": [ 112, 454, 280, 469 ], "score": 1.0, "content": "Comprehension on RefCOCO/RefCOCO", "type": "text" }, { "bbox": [ 280, 456, 286, 464 ], "score": 0.4, "content": "^ +", "type": "inline_equation" }, { "bbox": [ 287, 454, 340, 469 ], "score": 1.0, "content": "/RefCOCOg.", "type": "text" } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 105, 470, 428, 482 ], "spans": [ { "bbox": [ 105, 470, 428, 482 ], "score": 1.0, "content": "• Please refer to Figure 7 for Grounded Captioning on Flickr30k Karpathy split.", "type": "text" } ], "index": 15, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 484, 468, 498 ], "spans": [ { "bbox": [ 105, 484, 468, 498 ], "score": 1.0, "content": "• Please refer to Figure 8 for Evaluating Object Hallucination (POPE) on COCO val split.", "type": "text" } ], "index": 16, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 498, 382, 512 ], "spans": [ { "bbox": [ 105, 498, 382, 512 ], "score": 1.0, "content": "• Please refer to Table 17 for Referring Description in Ferret-Bench.", "type": "text" } ], "index": 17, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 104, 514, 374, 527 ], "spans": [ { "bbox": [ 104, 514, 374, 527 ], "score": 1.0, "content": "• Please refer to Table 18 for Referring Resoning in Ferret-Bench.", "type": "text" } ], "index": 18, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 528, 403, 542 ], "spans": [ { "bbox": [ 105, 528, 403, 542 ], "score": 1.0, "content": "• Please refer to Table 19 for Grounding in Conversation in Ferret-Bench.", "type": "text" } ], "index": 19, "is_list_start_line": true, "is_list_end_line": true } ], "index": 15, "bbox_fs": [ 103, 416, 506, 542 ] }, { "type": "title", "bbox": [ 111, 559, 248, 568 ], "lines": [ { "bbox": [ 111, 558, 249, 569 ], "spans": [ { "bbox": [ 111, 558, 249, 569 ], "score": 1.0, "content": "Referring Object Classification (LVIS) – point", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "image", "bbox": [ 107, 569, 551, 695 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 569, 551, 695 ], "group_id": 0, "lines": [ { "bbox": [ 107, 569, 551, 695 ], "spans": [ { "bbox": [ 107, 569, 551, 695 ], "score": 0.834, "type": "image", "image_path": "e084c90795f3cf70ed3f220ce71f783628efd1cede1582b7a3b021fa34951999.jpg" } ] } ], "index": 22, "virtual_lines": [ { "bbox": [ 107, 569, 551, 611.0 ], "spans": [], "index": 21 }, { "bbox": [ 107, 611.0, 551, 653.0 ], "spans": [], "index": 22 }, { "bbox": [ 107, 653.0, 551, 695.0 ], "spans": [], "index": 23 } ] }, { "type": "image_caption", "bbox": [ 106, 702, 507, 726 ], "group_id": 0, "lines": [ { "bbox": [ 105, 701, 506, 716 ], "spans": [ { "bbox": [ 105, 701, 506, 716 ], "score": 1.0, "content": "Figure 5: Referring Object Classification on LVIS. The task aims to classify specific region(s) in", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 713, 308, 726 ], "spans": [ { "bbox": [ 106, 713, 308, 726 ], "score": 1.0, "content": "an image given by point/box/segmentation inputs.", "type": "text" } ], "index": 25 } ], "index": 24.5 } ], "index": 23.25 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 109, 119, 546, 420 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 119, 546, 420 ], "group_id": 0, "lines": [ { "bbox": [ 109, 119, 546, 420 ], "spans": [ { "bbox": [ 109, 119, 546, 420 ], "score": 0.971, "type": "image", "image_path": "d3be54eee4ba06b26beca0e4d7cde52a0f7e855a6695a767e2825d478beb54ef.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 109, 119, 546, 219.33333333333331 ], "spans": [], "index": 0 }, { "bbox": [ 109, 219.33333333333331, 546, 319.66666666666663 ], "spans": [], "index": 1 }, { "bbox": [ 109, 319.66666666666663, 546, 419.99999999999994 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 426, 507, 460 ], "group_id": 0, "lines": [ { "bbox": [ 106, 426, 505, 438 ], "spans": [ { "bbox": [ 106, 426, 505, 438 ], "score": 1.0, "content": "Figure 6: Phrase Grounding on Flickr30k Entities and Referring Expression Comprehension", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 436, 505, 451 ], "spans": [ { "bbox": [ 105, 436, 208, 451 ], "score": 1.0, "content": "on RefCOCO/RefCOCO", "type": "text" }, { "bbox": [ 208, 438, 215, 447 ], "score": 0.38, "content": "+ \\prime", "type": "inline_equation" }, { "bbox": [ 215, 436, 505, 451 ], "score": 1.0, "content": "/RefCOCOg. The tasks aim to localize specific object(s) in an image", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 448, 301, 462 ], "spans": [ { "bbox": [ 106, 448, 301, 462 ], "score": 1.0, "content": "described by a referring expression/given entity.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "image", "bbox": [ 109, 536, 543, 672 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 536, 543, 672 ], "group_id": 1, "lines": [ { "bbox": [ 109, 536, 543, 672 ], "spans": [ { "bbox": [ 109, 536, 543, 672 ], "score": 0.974, "type": "image", "image_path": "870b2a6c1e6881026c4762c6c05bfb8fec21839573871875dc62bd3c66bea534.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 109, 536, 543, 581.3333333333334 ], "spans": [], "index": 6 }, { "bbox": [ 109, 581.3333333333334, 543, 626.6666666666667 ], "spans": [], "index": 7 }, { "bbox": [ 109, 626.6666666666667, 543, 672.0000000000001 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 106, 680, 506, 703 ], "group_id": 1, "lines": [ { "bbox": [ 105, 678, 505, 694 ], "spans": [ { "bbox": [ 105, 678, 505, 694 ], "score": 1.0, "content": "Figure 7: Grounded Captioning on Flickr30k. The task aims to generate a caption about the image", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 691, 334, 704 ], "spans": [ { "bbox": [ 106, 691, 334, 704 ], "score": 1.0, "content": "and ground all generated noun phrases to image regions.", "type": "text" } ], "index": 10 } ], "index": 9.5 } ], "index": 8.25 } ], "page_idx": 23, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 108, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 764 ], "spans": [ { "bbox": [ 298, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 15 } ] } ] }, { "type": "discarded", "bbox": [ 324, 111, 471, 119 ], "lines": [ { "bbox": [ 323, 110, 471, 121 ], "spans": [ { "bbox": [ 323, 110, 471, 121 ], "score": 1.0, "content": "Referring Expression Comprehension (RefCOCO)", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 112, 524, 254, 532 ], "lines": [ { "bbox": [ 111, 522, 255, 534 ], "spans": [ { "bbox": [ 111, 522, 255, 534 ], "score": 1.0, "content": "Grounded Captioning (Flickr30k Karpathy split)", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 111, 111, 225, 119 ], "lines": [ { "bbox": [ 111, 110, 226, 120 ], "spans": [ { "bbox": [ 111, 110, 226, 120 ], "score": 1.0, "content": "Phrase Grounding (Flickr30k Entities)", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 109, 119, 546, 420 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 119, 546, 420 ], "group_id": 0, "lines": [ { "bbox": [ 109, 119, 546, 420 ], "spans": [ { "bbox": [ 109, 119, 546, 420 ], "score": 0.971, "type": "image", "image_path": "d3be54eee4ba06b26beca0e4d7cde52a0f7e855a6695a767e2825d478beb54ef.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 109, 119, 546, 219.33333333333331 ], "spans": [], "index": 0 }, { "bbox": [ 109, 219.33333333333331, 546, 319.66666666666663 ], "spans": [], "index": 1 }, { "bbox": [ 109, 319.66666666666663, 546, 419.99999999999994 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 426, 507, 460 ], "group_id": 0, "lines": [ { "bbox": [ 106, 426, 505, 438 ], "spans": [ { "bbox": [ 106, 426, 505, 438 ], "score": 1.0, "content": "Figure 6: Phrase Grounding on Flickr30k Entities and Referring Expression Comprehension", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 436, 505, 451 ], "spans": [ { "bbox": [ 105, 436, 208, 451 ], "score": 1.0, "content": "on RefCOCO/RefCOCO", "type": "text" }, { "bbox": [ 208, 438, 215, 447 ], "score": 0.38, "content": "+ \\prime", "type": "inline_equation" }, { "bbox": [ 215, 436, 505, 451 ], "score": 1.0, "content": "/RefCOCOg. The tasks aim to localize specific object(s) in an image", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 448, 301, 462 ], "spans": [ { "bbox": [ 106, 448, 301, 462 ], "score": 1.0, "content": "described by a referring expression/given entity.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "image", "bbox": [ 109, 536, 543, 672 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 536, 543, 672 ], "group_id": 1, "lines": [ { "bbox": [ 109, 536, 543, 672 ], "spans": [ { "bbox": [ 109, 536, 543, 672 ], "score": 0.974, "type": "image", "image_path": "870b2a6c1e6881026c4762c6c05bfb8fec21839573871875dc62bd3c66bea534.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 109, 536, 543, 581.3333333333334 ], "spans": [], "index": 6 }, { "bbox": [ 109, 581.3333333333334, 543, 626.6666666666667 ], "spans": [], "index": 7 }, { "bbox": [ 109, 626.6666666666667, 543, 672.0000000000001 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 106, 680, 506, 703 ], "group_id": 1, "lines": [ { "bbox": [ 105, 678, 505, 694 ], "spans": [ { "bbox": [ 105, 678, 505, 694 ], "score": 1.0, "content": "Figure 7: Grounded Captioning on Flickr30k. The task aims to generate a caption about the image", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 691, 334, 704 ], "spans": [ { "bbox": [ 106, 691, 334, 704 ], "score": 1.0, "content": "and ground all generated noun phrases to image regions.", "type": "text" } ], "index": 10 } ], "index": 9.5 } ], "index": 8.25 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 109, 240, 545, 542 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 240, 545, 542 ], "group_id": 0, "lines": [ { "bbox": [ 109, 240, 545, 542 ], "spans": [ { "bbox": [ 109, 240, 545, 542 ], "score": 0.975, "type": "image", "image_path": "2e2366c809f7208ab1082371ea7d8a51b03d593121c1fa81bebb26409b4c30ce.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 109, 240, 545, 340.6666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 109, 340.6666666666667, 545, 441.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 109, 441.33333333333337, 545, 542.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 549, 507, 583 ], "group_id": 0, "lines": [ { "bbox": [ 105, 547, 505, 563 ], "spans": [ { "bbox": [ 105, 547, 505, 563 ], "score": 1.0, "content": "Figure 8: Object Hallucination Evaluation (POPE) on COCO. The task aims to input a query", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 560, 505, 573 ], "spans": [ { "bbox": [ 106, 560, 505, 573 ], "score": 1.0, "content": "inquiring about the existence of an object, and the model is expected to generate a response in the", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 571, 205, 583 ], "spans": [ { "bbox": [ 106, 571, 205, 583 ], "score": 1.0, "content": "form of either “yes/no”.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ], "page_idx": 24, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 749, 313, 764 ], "spans": [ { "bbox": [ 298, 749, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] }, { "type": "discarded", "bbox": [ 111, 231, 189, 239 ], "lines": [ { "bbox": [ 111, 230, 190, 241 ], "spans": [ { "bbox": [ 111, 230, 190, 241 ], "score": 1.0, "content": "POPE (COCO adversarial)", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 109, 240, 545, 542 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 240, 545, 542 ], "group_id": 0, "lines": [ { "bbox": [ 109, 240, 545, 542 ], "spans": [ { "bbox": [ 109, 240, 545, 542 ], "score": 0.975, "type": "image", "image_path": "2e2366c809f7208ab1082371ea7d8a51b03d593121c1fa81bebb26409b4c30ce.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 109, 240, 545, 340.6666666666667 ], "spans": [], "index": 0 }, { "bbox": [ 109, 340.6666666666667, 545, 441.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 109, 441.33333333333337, 545, 542.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 549, 507, 583 ], "group_id": 0, "lines": [ { "bbox": [ 105, 547, 505, 563 ], "spans": [ { "bbox": [ 105, 547, 505, 563 ], "score": 1.0, "content": "Figure 8: Object Hallucination Evaluation (POPE) on COCO. The task aims to input a query", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 560, 505, 573 ], "spans": [ { "bbox": [ 106, 560, 505, 573 ], "score": 1.0, "content": "inquiring about the existence of an object, and the model is expected to generate a response in the", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 571, 205, 583 ], "spans": [ { "bbox": [ 106, 571, 205, 583 ], "score": 1.0, "content": "form of either “yes/no”.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 139, 505, 172 ], "lines": [ { "bbox": [ 105, 138, 505, 152 ], "spans": [ { "bbox": [ 105, 138, 505, 152 ], "score": 1.0, "content": "Table 17: Referring Description in Ferret-Bench. Qualitative examples to illustrate the difference", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 149, 505, 162 ], "spans": [ { "bbox": [ 105, 149, 505, 162 ], "score": 1.0, "content": "between various models (LLaVA vs. Kosmos-2 vs. Shikra vs. Ferret (Ours)). For clarity, we have", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 161, 363, 173 ], "spans": [ { "bbox": [ 106, 161, 363, 173 ], "score": 1.0, "content": "omitted the bounding box outputs in the textual representations.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "title", "bbox": [ 128, 192, 167, 201 ], "lines": [ { "bbox": [ 127, 191, 168, 202 ], "spans": [ { "bbox": [ 127, 191, 168, 202 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 128, 202, 316, 222 ], "lines": [ { "bbox": [ 127, 200, 317, 214 ], "spans": [ { "bbox": [ 127, 200, 317, 214 ], "score": 1.0, "content": "What is the interaction between the object [region0]", "type": "text" } ], "index": 4 }, { "bbox": [ 127, 210, 232, 222 ], "spans": [ { "bbox": [ 127, 210, 232, 222 ], "score": 1.0, "content": "and its surrounding objects?", "type": "text" } ], "index": 5 } ], "index": 4.5 }, { "type": "title", "bbox": [ 128, 232, 186, 241 ], "lines": [ { "bbox": [ 127, 231, 188, 242 ], "spans": [ { "bbox": [ 127, 231, 188, 242 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 128, 243, 329, 292 ], "lines": [ { "bbox": [ 127, 241, 295, 253 ], "spans": [ { "bbox": [ 127, 241, 295, 253 ], "score": 1.0, "content": "In the region, there is a metal support column.", "type": "text" } ], "index": 7 }, { "bbox": [ 127, 250, 328, 264 ], "spans": [ { "bbox": [ 127, 250, 328, 264 ], "score": 1.0, "content": "This column is providing support for a bridge above it,", "type": "text" } ], "index": 8 }, { "bbox": [ 127, 261, 329, 273 ], "spans": [ { "bbox": [ 127, 261, 329, 273 ], "score": 1.0, "content": "which a train is passing over. The column also features", "type": "text" } ], "index": 9 }, { "bbox": [ 127, 270, 330, 283 ], "spans": [ { "bbox": [ 127, 270, 330, 283 ], "score": 1.0, "content": "a red line on it. This region is part of a larger scene that", "type": "text" } ], "index": 10 }, { "bbox": [ 127, 280, 287, 293 ], "spans": [ { "bbox": [ 127, 280, 287, 293 ], "score": 1.0, "content": "includes a train track on an elevated bridge.", "type": "text" } ], "index": 11 } ], "index": 9 }, { "type": "image", "bbox": [ 348, 192, 493, 300 ], "blocks": [ { "type": "image_body", "bbox": [ 348, 192, 493, 300 ], "group_id": 0, "lines": [ { "bbox": [ 348, 192, 493, 300 ], "spans": [ { "bbox": [ 348, 192, 493, 300 ], "score": 0.967, "type": "image", "image_path": "a5caab1234f881f9eb4b4a3d3fbd0908b25430d28ccfaa8cab52ff319b8d3417.jpg" } ] } ], "index": 12.5, "virtual_lines": [ { "bbox": [ 348, 192, 493, 246.0 ], "spans": [], "index": 12 }, { "bbox": [ 348, 246.0, 493, 300.0 ], "spans": [], "index": 13 } ] } ], "index": 12.5 }, { "type": "title", "bbox": [ 128, 302, 160, 311 ], "lines": [ { "bbox": [ 127, 301, 162, 312 ], "spans": [ { "bbox": [ 127, 301, 162, 312 ], "score": 1.0, "content": "LLaVA:", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 128, 312, 485, 341 ], "lines": [ { "bbox": [ 127, 311, 485, 323 ], "spans": [ { "bbox": [ 127, 311, 485, 323 ], "score": 1.0, "content": "The object is a bridge which appears to be a part of a train station, as there are traffic lights present in", "type": "text" } ], "index": 15 }, { "bbox": [ 126, 320, 485, 335 ], "spans": [ { "bbox": [ 126, 320, 485, 335 ], "score": 1.0, "content": "the scene. The traffic lights are likely used to control the train’s movement and ensure safe passage", "type": "text" } ], "index": 16 }, { "bbox": [ 126, 330, 187, 344 ], "spans": [ { "bbox": [ 126, 330, 187, 344 ], "score": 1.0, "content": "over the bridge.", "type": "text" } ], "index": 17 } ], "index": 16 }, { "type": "title", "bbox": [ 129, 342, 171, 351 ], "lines": [ { "bbox": [ 127, 341, 172, 351 ], "spans": [ { "bbox": [ 127, 341, 172, 351 ], "score": 1.0, "content": "Kosmos-2:", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 127, 351, 468, 362 ], "lines": [ { "bbox": [ 126, 349, 469, 363 ], "spans": [ { "bbox": [ 126, 349, 469, 363 ], "score": 1.0, "content": "The interaction between the object and its surroundings is that of a train passing over a bridge.", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "title", "bbox": [ 128, 362, 159, 371 ], "lines": [ { "bbox": [ 127, 360, 161, 372 ], "spans": [ { "bbox": [ 127, 360, 161, 372 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 128, 371, 483, 391 ], "lines": [ { "bbox": [ 126, 369, 486, 384 ], "spans": [ { "bbox": [ 126, 369, 486, 384 ], "score": 1.0, "content": "The metal pole is standing tall and straight, supporting the train that is traveling on the elevated", "type": "text" } ], "index": 21 }, { "bbox": [ 126, 380, 150, 392 ], "spans": [ { "bbox": [ 126, 380, 150, 392 ], "score": 1.0, "content": "track.", "type": "text" } ], "index": 22 } ], "index": 21.5 }, { "type": "title", "bbox": [ 129, 392, 185, 401 ], "lines": [ { "bbox": [ 127, 391, 186, 402 ], "spans": [ { "bbox": [ 127, 391, 186, 402 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 128, 401, 484, 461 ], "lines": [ { "bbox": [ 126, 399, 486, 413 ], "spans": [ { "bbox": [ 126, 399, 486, 413 ], "score": 1.0, "content": "The object is a metal post that is part of a bridge. This post is essential because it supports the bridge", "type": "text" } ], "index": 24 }, { "bbox": [ 127, 411, 486, 423 ], "spans": [ { "bbox": [ 127, 411, 486, 423 ], "score": 1.0, "content": "structure, ensuring its stability for the train to safely traverse the bridge. Other elements such as", "type": "text" } ], "index": 25 }, { "bbox": [ 127, 421, 485, 433 ], "spans": [ { "bbox": [ 127, 421, 485, 433 ], "score": 1.0, "content": "the train, and the overhead guard rail, depend on this post for support and stability. Without this", "type": "text" } ], "index": 26 }, { "bbox": [ 126, 430, 486, 443 ], "spans": [ { "bbox": [ 126, 430, 486, 443 ], "score": 1.0, "content": "post, the bridge wouldn’t be able to withstand the train’s weight and motion, potentially leading to", "type": "text" } ], "index": 27 }, { "bbox": [ 126, 440, 485, 454 ], "spans": [ { "bbox": [ 126, 440, 485, 454 ], "score": 1.0, "content": "a structural collapse. Therefore, this post plays a critical role in maintaining the bridge’s integrity", "type": "text" } ], "index": 28 }, { "bbox": [ 127, 450, 176, 463 ], "spans": [ { "bbox": [ 127, 450, 176, 463 ], "score": 1.0, "content": "and stability.", "type": "text" } ], "index": 29 } ], "index": 26.5 }, { "type": "title", "bbox": [ 128, 482, 167, 490 ], "lines": [ { "bbox": [ 127, 480, 168, 491 ], "spans": [ { "bbox": [ 127, 480, 168, 491 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 128, 491, 320, 511 ], "lines": [ { "bbox": [ 127, 490, 321, 502 ], "spans": [ { "bbox": [ 127, 490, 321, 502 ], "score": 1.0, "content": "What is the relationship between the object [region0]", "type": "text" } ], "index": 31 }, { "bbox": [ 127, 500, 307, 512 ], "spans": [ { "bbox": [ 127, 500, 307, 512 ], "score": 1.0, "content": "and object [region1] and its surrounding objects?", "type": "text" } ], "index": 32 } ], "index": 31.5 }, { "type": "title", "bbox": [ 128, 521, 187, 530 ], "lines": [ { "bbox": [ 127, 520, 189, 531 ], "spans": [ { "bbox": [ 127, 520, 189, 531 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 33 } ], "index": 33 }, { "type": "text", "bbox": [ 128, 531, 295, 560 ], "lines": [ { "bbox": [ 127, 530, 297, 542 ], "spans": [ { "bbox": [ 127, 530, 297, 542 ], "score": 1.0, "content": "The object is a vase, and the object is a flower.", "type": "text" } ], "index": 34 }, { "bbox": [ 127, 540, 284, 552 ], "spans": [ { "bbox": [ 127, 540, 284, 552 ], "score": 1.0, "content": "The flower is in the vase, suggesting it is a", "type": "text" } ], "index": 35 }, { "bbox": [ 127, 549, 260, 561 ], "spans": [ { "bbox": [ 127, 549, 260, 561 ], "score": 1.0, "content": "decorative element within the room.", "type": "text" } ], "index": 37 } ], "index": 35 }, { "type": "image", "bbox": [ 373, 481, 469, 623 ], "blocks": [ { "type": "image_body", "bbox": [ 373, 481, 469, 623 ], "group_id": 1, "lines": [ { "bbox": [ 373, 481, 469, 623 ], "spans": [ { "bbox": [ 373, 481, 469, 623 ], "score": 0.929, "type": "image", "image_path": "6e13e83f57ea100a0ff264f473ee397938aaa9d3d964567cd54050ab0cc2cac0.jpg" } ] } ], "index": 36, "virtual_lines": [ { "bbox": [ 373, 481, 469, 623 ], "spans": [], "index": 36 } ] } ], "index": 36 }, { "type": "text", "bbox": [ 127, 571, 299, 632 ], "lines": [ { "bbox": [ 126, 568, 163, 582 ], "spans": [ { "bbox": [ 126, 568, 163, 582 ], "score": 1.0, "content": "LLaVA:", "type": "text" } ], "index": 38 }, { "bbox": [ 126, 578, 300, 592 ], "spans": [ { "bbox": [ 126, 578, 300, 592 ], "score": 1.0, "content": "A wooden table with a green bowl placed on it.", "type": "text" } ], "index": 39 }, { "bbox": [ 126, 588, 173, 601 ], "spans": [ { "bbox": [ 126, 588, 173, 601 ], "score": 1.0, "content": "Kosmos-2:", "type": "text" } ], "index": 40 }, { "bbox": [ 126, 598, 269, 613 ], "spans": [ { "bbox": [ 126, 598, 269, 613 ], "score": 1.0, "content": "The object is a vase with a green base.", "type": "text" } ], "index": 41 }, { "bbox": [ 128, 609, 161, 621 ], "spans": [ { "bbox": [ 128, 609, 161, 621 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 42 }, { "bbox": [ 126, 618, 212, 632 ], "spans": [ { "bbox": [ 126, 618, 212, 632 ], "score": 1.0, "content": "The answer is flowers.", "type": "text" } ], "index": 43 } ], "index": 40.5 }, { "type": "title", "bbox": [ 129, 631, 185, 640 ], "lines": [ { "bbox": [ 127, 630, 186, 641 ], "spans": [ { "bbox": [ 127, 630, 186, 641 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 44 } ], "index": 44 }, { "type": "text", "bbox": [ 127, 640, 483, 661 ], "lines": [ { "bbox": [ 127, 639, 484, 653 ], "spans": [ { "bbox": [ 127, 639, 484, 653 ], "score": 1.0, "content": "The object and object are a vase and a flower, respectively. The vase is holding the flower, suggest-", "type": "text" } ], "index": 45 }, { "bbox": [ 127, 649, 255, 663 ], "spans": [ { "bbox": [ 127, 649, 255, 663 ], "score": 1.0, "content": "ing it’s displaying or supporting it.", "type": "text" } ], "index": 46 } ], "index": 45.5 } ], "page_idx": 25, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 763 ], "spans": [ { "bbox": [ 298, 750, 313, 763 ], "score": 1.0, "content": "26", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 139, 505, 172 ], "lines": [ { "bbox": [ 105, 138, 505, 152 ], "spans": [ { "bbox": [ 105, 138, 505, 152 ], "score": 1.0, "content": "Table 17: Referring Description in Ferret-Bench. Qualitative examples to illustrate the difference", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 149, 505, 162 ], "spans": [ { "bbox": [ 105, 149, 505, 162 ], "score": 1.0, "content": "between various models (LLaVA vs. Kosmos-2 vs. Shikra vs. Ferret (Ours)). For clarity, we have", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 161, 363, 173 ], "spans": [ { "bbox": [ 106, 161, 363, 173 ], "score": 1.0, "content": "omitted the bounding box outputs in the textual representations.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 105, 138, 505, 173 ] }, { "type": "title", "bbox": [ 128, 192, 167, 201 ], "lines": [ { "bbox": [ 127, 191, 168, 202 ], "spans": [ { "bbox": [ 127, 191, 168, 202 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 128, 202, 316, 222 ], "lines": [ { "bbox": [ 127, 200, 317, 214 ], "spans": [ { "bbox": [ 127, 200, 317, 214 ], "score": 1.0, "content": "What is the interaction between the object [region0]", "type": "text" } ], "index": 4 }, { "bbox": [ 127, 210, 232, 222 ], "spans": [ { "bbox": [ 127, 210, 232, 222 ], "score": 1.0, "content": "and its surrounding objects?", "type": "text" } ], "index": 5 } ], "index": 4.5, "bbox_fs": [ 127, 200, 317, 222 ] }, { "type": "title", "bbox": [ 128, 232, 186, 241 ], "lines": [ { "bbox": [ 127, 231, 188, 242 ], "spans": [ { "bbox": [ 127, 231, 188, 242 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 128, 243, 329, 292 ], "lines": [ { "bbox": [ 127, 241, 295, 253 ], "spans": [ { "bbox": [ 127, 241, 295, 253 ], "score": 1.0, "content": "In the region, there is a metal support column.", "type": "text" } ], "index": 7 }, { "bbox": [ 127, 250, 328, 264 ], "spans": [ { "bbox": [ 127, 250, 328, 264 ], "score": 1.0, "content": "This column is providing support for a bridge above it,", "type": "text" } ], "index": 8 }, { "bbox": [ 127, 261, 329, 273 ], "spans": [ { "bbox": [ 127, 261, 329, 273 ], "score": 1.0, "content": "which a train is passing over. The column also features", "type": "text" } ], "index": 9 }, { "bbox": [ 127, 270, 330, 283 ], "spans": [ { "bbox": [ 127, 270, 330, 283 ], "score": 1.0, "content": "a red line on it. This region is part of a larger scene that", "type": "text" } ], "index": 10 }, { "bbox": [ 127, 280, 287, 293 ], "spans": [ { "bbox": [ 127, 280, 287, 293 ], "score": 1.0, "content": "includes a train track on an elevated bridge.", "type": "text" } ], "index": 11 } ], "index": 9, "bbox_fs": [ 127, 241, 330, 293 ] }, { "type": "image", "bbox": [ 348, 192, 493, 300 ], "blocks": [ { "type": "image_body", "bbox": [ 348, 192, 493, 300 ], "group_id": 0, "lines": [ { "bbox": [ 348, 192, 493, 300 ], "spans": [ { "bbox": [ 348, 192, 493, 300 ], "score": 0.967, "type": "image", "image_path": "a5caab1234f881f9eb4b4a3d3fbd0908b25430d28ccfaa8cab52ff319b8d3417.jpg" } ] } ], "index": 12.5, "virtual_lines": [ { "bbox": [ 348, 192, 493, 246.0 ], "spans": [], "index": 12 }, { "bbox": [ 348, 246.0, 493, 300.0 ], "spans": [], "index": 13 } ] } ], "index": 12.5 }, { "type": "title", "bbox": [ 128, 302, 160, 311 ], "lines": [ { "bbox": [ 127, 301, 162, 312 ], "spans": [ { "bbox": [ 127, 301, 162, 312 ], "score": 1.0, "content": "LLaVA:", "type": "text" } ], "index": 14 } ], "index": 14 }, { "type": "text", "bbox": [ 128, 312, 485, 341 ], "lines": [ { "bbox": [ 127, 311, 485, 323 ], "spans": [ { "bbox": [ 127, 311, 485, 323 ], "score": 1.0, "content": "The object is a bridge which appears to be a part of a train station, as there are traffic lights present in", "type": "text" } ], "index": 15 }, { "bbox": [ 126, 320, 485, 335 ], "spans": [ { "bbox": [ 126, 320, 485, 335 ], "score": 1.0, "content": "the scene. The traffic lights are likely used to control the train’s movement and ensure safe passage", "type": "text" } ], "index": 16 }, { "bbox": [ 126, 330, 187, 344 ], "spans": [ { "bbox": [ 126, 330, 187, 344 ], "score": 1.0, "content": "over the bridge.", "type": "text" } ], "index": 17 } ], "index": 16, "bbox_fs": [ 126, 311, 485, 344 ] }, { "type": "title", "bbox": [ 129, 342, 171, 351 ], "lines": [ { "bbox": [ 127, 341, 172, 351 ], "spans": [ { "bbox": [ 127, 341, 172, 351 ], "score": 1.0, "content": "Kosmos-2:", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 127, 351, 468, 362 ], "lines": [ { "bbox": [ 126, 349, 469, 363 ], "spans": [ { "bbox": [ 126, 349, 469, 363 ], "score": 1.0, "content": "The interaction between the object and its surroundings is that of a train passing over a bridge.", "type": "text" } ], "index": 19 } ], "index": 19, "bbox_fs": [ 126, 349, 469, 363 ] }, { "type": "title", "bbox": [ 128, 362, 159, 371 ], "lines": [ { "bbox": [ 127, 360, 161, 372 ], "spans": [ { "bbox": [ 127, 360, 161, 372 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 128, 371, 483, 391 ], "lines": [ { "bbox": [ 126, 369, 486, 384 ], "spans": [ { "bbox": [ 126, 369, 486, 384 ], "score": 1.0, "content": "The metal pole is standing tall and straight, supporting the train that is traveling on the elevated", "type": "text" } ], "index": 21 }, { "bbox": [ 126, 380, 150, 392 ], "spans": [ { "bbox": [ 126, 380, 150, 392 ], "score": 1.0, "content": "track.", "type": "text" } ], "index": 22 } ], "index": 21.5, "bbox_fs": [ 126, 369, 486, 392 ] }, { "type": "title", "bbox": [ 129, 392, 185, 401 ], "lines": [ { "bbox": [ 127, 391, 186, 402 ], "spans": [ { "bbox": [ 127, 391, 186, 402 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 128, 401, 484, 461 ], "lines": [ { "bbox": [ 126, 399, 486, 413 ], "spans": [ { "bbox": [ 126, 399, 486, 413 ], "score": 1.0, "content": "The object is a metal post that is part of a bridge. This post is essential because it supports the bridge", "type": "text" } ], "index": 24 }, { "bbox": [ 127, 411, 486, 423 ], "spans": [ { "bbox": [ 127, 411, 486, 423 ], "score": 1.0, "content": "structure, ensuring its stability for the train to safely traverse the bridge. Other elements such as", "type": "text" } ], "index": 25 }, { "bbox": [ 127, 421, 485, 433 ], "spans": [ { "bbox": [ 127, 421, 485, 433 ], "score": 1.0, "content": "the train, and the overhead guard rail, depend on this post for support and stability. Without this", "type": "text" } ], "index": 26 }, { "bbox": [ 126, 430, 486, 443 ], "spans": [ { "bbox": [ 126, 430, 486, 443 ], "score": 1.0, "content": "post, the bridge wouldn’t be able to withstand the train’s weight and motion, potentially leading to", "type": "text" } ], "index": 27 }, { "bbox": [ 126, 440, 485, 454 ], "spans": [ { "bbox": [ 126, 440, 485, 454 ], "score": 1.0, "content": "a structural collapse. Therefore, this post plays a critical role in maintaining the bridge’s integrity", "type": "text" } ], "index": 28 }, { "bbox": [ 127, 450, 176, 463 ], "spans": [ { "bbox": [ 127, 450, 176, 463 ], "score": 1.0, "content": "and stability.", "type": "text" } ], "index": 29 } ], "index": 26.5, "bbox_fs": [ 126, 399, 486, 463 ] }, { "type": "title", "bbox": [ 128, 482, 167, 490 ], "lines": [ { "bbox": [ 127, 480, 168, 491 ], "spans": [ { "bbox": [ 127, 480, 168, 491 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 30 } ], "index": 30 }, { "type": "text", "bbox": [ 128, 491, 320, 511 ], "lines": [ { "bbox": [ 127, 490, 321, 502 ], "spans": [ { "bbox": [ 127, 490, 321, 502 ], "score": 1.0, "content": "What is the relationship between the object [region0]", "type": "text" } ], "index": 31 }, { "bbox": [ 127, 500, 307, 512 ], "spans": [ { "bbox": [ 127, 500, 307, 512 ], "score": 1.0, "content": "and object [region1] and its surrounding objects?", "type": "text" } ], "index": 32 } ], "index": 31.5, "bbox_fs": [ 127, 490, 321, 512 ] }, { "type": "title", "bbox": [ 128, 521, 187, 530 ], "lines": [ { "bbox": [ 127, 520, 189, 531 ], "spans": [ { "bbox": [ 127, 520, 189, 531 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 33 } ], "index": 33 }, { "type": "text", "bbox": [ 128, 531, 295, 560 ], "lines": [ { "bbox": [ 127, 530, 297, 542 ], "spans": [ { "bbox": [ 127, 530, 297, 542 ], "score": 1.0, "content": "The object is a vase, and the object is a flower.", "type": "text" } ], "index": 34 }, { "bbox": [ 127, 540, 284, 552 ], "spans": [ { "bbox": [ 127, 540, 284, 552 ], "score": 1.0, "content": "The flower is in the vase, suggesting it is a", "type": "text" } ], "index": 35 }, { "bbox": [ 127, 549, 260, 561 ], "spans": [ { "bbox": [ 127, 549, 260, 561 ], "score": 1.0, "content": "decorative element within the room.", "type": "text" } ], "index": 37 } ], "index": 35, "bbox_fs": [ 127, 530, 297, 561 ] }, { "type": "image", "bbox": [ 373, 481, 469, 623 ], "blocks": [ { "type": "image_body", "bbox": [ 373, 481, 469, 623 ], "group_id": 1, "lines": [ { "bbox": [ 373, 481, 469, 623 ], "spans": [ { "bbox": [ 373, 481, 469, 623 ], "score": 0.929, "type": "image", "image_path": "6e13e83f57ea100a0ff264f473ee397938aaa9d3d964567cd54050ab0cc2cac0.jpg" } ] } ], "index": 36, "virtual_lines": [ { "bbox": [ 373, 481, 469, 623 ], "spans": [], "index": 36 } ] } ], "index": 36 }, { "type": "list", "bbox": [ 127, 571, 299, 632 ], "lines": [ { "bbox": [ 126, 568, 163, 582 ], "spans": [ { "bbox": [ 126, 568, 163, 582 ], "score": 1.0, "content": "LLaVA:", "type": "text" } ], "index": 38, "is_list_end_line": true }, { "bbox": [ 126, 578, 300, 592 ], "spans": [ { "bbox": [ 126, 578, 300, 592 ], "score": 1.0, "content": "A wooden table with a green bowl placed on it.", "type": "text" } ], "index": 39, "is_list_start_line": true }, { "bbox": [ 126, 588, 173, 601 ], "spans": [ { "bbox": [ 126, 588, 173, 601 ], "score": 1.0, "content": "Kosmos-2:", "type": "text" } ], "index": 40, "is_list_end_line": true }, { "bbox": [ 126, 598, 269, 613 ], "spans": [ { "bbox": [ 126, 598, 269, 613 ], "score": 1.0, "content": "The object is a vase with a green base.", "type": "text" } ], "index": 41, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 128, 609, 161, 621 ], "spans": [ { "bbox": [ 128, 609, 161, 621 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 42, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 126, 618, 212, 632 ], "spans": [ { "bbox": [ 126, 618, 212, 632 ], "score": 1.0, "content": "The answer is flowers.", "type": "text" } ], "index": 43, "is_list_start_line": true, "is_list_end_line": true } ], "index": 40.5, "bbox_fs": [ 126, 568, 300, 632 ] }, { "type": "title", "bbox": [ 129, 631, 185, 640 ], "lines": [ { "bbox": [ 127, 630, 186, 641 ], "spans": [ { "bbox": [ 127, 630, 186, 641 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 44 } ], "index": 44 }, { "type": "text", "bbox": [ 127, 640, 483, 661 ], "lines": [ { "bbox": [ 127, 639, 484, 653 ], "spans": [ { "bbox": [ 127, 639, 484, 653 ], "score": 1.0, "content": "The object and object are a vase and a flower, respectively. The vase is holding the flower, suggest-", "type": "text" } ], "index": 45 }, { "bbox": [ 127, 649, 255, 663 ], "spans": [ { "bbox": [ 127, 649, 255, 663 ], "score": 1.0, "content": "ing it’s displaying or supporting it.", "type": "text" } ], "index": 46 } ], "index": 45.5, "bbox_fs": [ 127, 639, 484, 663 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 153, 505, 187 ], "lines": [ { "bbox": [ 105, 153, 505, 166 ], "spans": [ { "bbox": [ 105, 153, 505, 166 ], "score": 1.0, "content": "Table 18: Referring Reasoning in Ferret-Bench. Qualitative examples to illustrate the difference", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 165, 505, 177 ], "spans": [ { "bbox": [ 106, 165, 505, 177 ], "score": 1.0, "content": "between various models (LLaVA vs. Kosmos-2 vs. Shikra vs. Ferret (Ours)). For clarity, we have", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 176, 363, 188 ], "spans": [ { "bbox": [ 106, 176, 363, 188 ], "score": 1.0, "content": "omitted the bounding box outputs in the textual representations.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "title", "bbox": [ 128, 207, 167, 216 ], "lines": [ { "bbox": [ 127, 205, 168, 217 ], "spans": [ { "bbox": [ 127, 205, 168, 217 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 130, 217, 288, 227 ], "lines": [ { "bbox": [ 128, 216, 289, 228 ], "spans": [ { "bbox": [ 128, 216, 289, 228 ], "score": 1.0, "content": "What is the purpose of the object [region0]?", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "title", "bbox": [ 128, 237, 187, 246 ], "lines": [ { "bbox": [ 128, 236, 187, 246 ], "spans": [ { "bbox": [ 128, 236, 187, 246 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 128, 247, 329, 277 ], "lines": [ { "bbox": [ 127, 245, 312, 258 ], "spans": [ { "bbox": [ 127, 245, 312, 258 ], "score": 1.0, "content": "The object is a bottle. Bottles are typically used to", "type": "text" } ], "index": 6 }, { "bbox": [ 126, 255, 330, 269 ], "spans": [ { "bbox": [ 126, 255, 330, 269 ], "score": 1.0, "content": "hold and store different types of liquids. In this context,", "type": "text" } ], "index": 7 }, { "bbox": [ 127, 266, 305, 278 ], "spans": [ { "bbox": [ 127, 266, 305, 278 ], "score": 1.0, "content": "it might be used to store a beverage for the meal.", "type": "text" } ], "index": 8 } ], "index": 7 }, { "type": "image", "bbox": [ 348, 208, 498, 307 ], "blocks": [ { "type": "image_body", "bbox": [ 348, 208, 498, 307 ], "group_id": 0, "lines": [ { "bbox": [ 348, 208, 498, 307 ], "spans": [ { "bbox": [ 348, 208, 498, 307 ], "score": 0.971, "type": "image", "image_path": "c6836caac403445c88c4fb683267e9ac1e0024968f18020c9fedb9af5f282cb6.jpg" } ] } ], "index": 9.5, "virtual_lines": [ { "bbox": [ 348, 208, 498, 257.5 ], "spans": [], "index": 9 }, { "bbox": [ 348, 257.5, 498, 307.0 ], "spans": [], "index": 10 } ] } ], "index": 9.5 }, { "type": "title", "bbox": [ 128, 287, 160, 296 ], "lines": [ { "bbox": [ 127, 286, 162, 296 ], "spans": [ { "bbox": [ 127, 286, 162, 296 ], "score": 1.0, "content": "LLaVA:", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 130, 297, 329, 307 ], "lines": [ { "bbox": [ 129, 296, 330, 307 ], "spans": [ { "bbox": [ 129, 296, 330, 307 ], "score": 1.0, "content": "A knife and a fork, indicating that it is ready for a meal.", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 130, 317, 340, 327 ], "lines": [ { "bbox": [ 129, 316, 342, 327 ], "spans": [ { "bbox": [ 129, 316, 342, 327 ], "score": 1.0, "content": "The purpose of the object is to attract the birds to the table.", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 128, 327, 241, 346 ], "lines": [ { "bbox": [ 127, 325, 162, 338 ], "spans": [ { "bbox": [ 127, 325, 162, 338 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 14 }, { "bbox": [ 127, 334, 243, 348 ], "spans": [ { "bbox": [ 127, 334, 243, 348 ], "score": 1.0, "content": "The answer is keep birds away.", "type": "text" } ], "index": 15 } ], "index": 14.5 }, { "type": "text", "bbox": [ 129, 347, 184, 356 ], "lines": [ { "bbox": [ 127, 345, 186, 357 ], "spans": [ { "bbox": [ 127, 345, 186, 357 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "text", "bbox": [ 128, 357, 483, 376 ], "lines": [ { "bbox": [ 127, 354, 484, 369 ], "spans": [ { "bbox": [ 127, 354, 484, 369 ], "score": 1.0, "content": "The object is a bottle, generally used to store and dispense liquids like water, juice, or other bever-", "type": "text" } ], "index": 17 }, { "bbox": [ 127, 367, 149, 378 ], "spans": [ { "bbox": [ 127, 367, 149, 378 ], "score": 1.0, "content": "ages.", "type": "text" } ], "index": 18 } ], "index": 17.5 }, { "type": "title", "bbox": [ 128, 397, 167, 406 ], "lines": [ { "bbox": [ 127, 395, 169, 407 ], "spans": [ { "bbox": [ 127, 395, 169, 407 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 128, 406, 330, 417 ], "lines": [ { "bbox": [ 127, 405, 332, 417 ], "spans": [ { "bbox": [ 127, 405, 332, 417 ], "score": 1.0, "content": "What is the purpose of the object [region0] on the bike?", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "title", "bbox": [ 128, 426, 186, 435 ], "lines": [ { "bbox": [ 128, 426, 187, 436 ], "spans": [ { "bbox": [ 128, 426, 187, 436 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 128, 437, 315, 516 ], "lines": [ { "bbox": [ 127, 435, 284, 447 ], "spans": [ { "bbox": [ 127, 435, 284, 447 ], "score": 1.0, "content": "The object is a shock absorber on the bike.", "type": "text" } ], "index": 22 }, { "bbox": [ 127, 446, 313, 457 ], "spans": [ { "bbox": [ 127, 446, 313, 457 ], "score": 1.0, "content": "Its purpose is to absorb or dampen shock impulses.", "type": "text" } ], "index": 23 }, { "bbox": [ 127, 455, 298, 467 ], "spans": [ { "bbox": [ 127, 455, 298, 467 ], "score": 1.0, "content": "It does this by converting the kinetic energy of", "type": "text" } ], "index": 24 }, { "bbox": [ 127, 465, 316, 477 ], "spans": [ { "bbox": [ 127, 465, 316, 477 ], "score": 1.0, "content": "the shock into another form of energy which is then", "type": "text" } ], "index": 25 }, { "bbox": [ 128, 476, 288, 487 ], "spans": [ { "bbox": [ 128, 476, 288, 487 ], "score": 1.0, "content": "dissipated. In the context of the motorcycle,", "type": "text" } ], "index": 26 }, { "bbox": [ 127, 486, 303, 497 ], "spans": [ { "bbox": [ 127, 486, 303, 497 ], "score": 1.0, "content": "it is particularly useful in providing comfort and", "type": "text" } ], "index": 27 }, { "bbox": [ 127, 493, 303, 509 ], "spans": [ { "bbox": [ 127, 493, 303, 509 ], "score": 1.0, "content": "stability for the rider, especially when travelling", "type": "text" } ], "index": 28 }, { "bbox": [ 128, 506, 238, 516 ], "spans": [ { "bbox": [ 128, 506, 238, 516 ], "score": 1.0, "content": "over uneven or rough terrains.", "type": "text" } ], "index": 29 } ], "index": 25.5 }, { "type": "image", "bbox": [ 350, 396, 495, 505 ], "blocks": [ { "type": "image_body", "bbox": [ 350, 396, 495, 505 ], "group_id": 1, "lines": [ { "bbox": [ 350, 396, 495, 505 ], "spans": [ { "bbox": [ 350, 396, 495, 505 ], "score": 0.966, "type": "image", "image_path": "374730542b5bbe01698e33da5f84294f1c9442fbaeaf8dc837c63230cf8b2069.jpg" } ] } ], "index": 30.5, "virtual_lines": [ { "bbox": [ 350, 396, 495, 450.5 ], "spans": [], "index": 30 }, { "bbox": [ 350, 450.5, 495, 505.0 ], "spans": [], "index": 31 } ] } ], "index": 30.5 }, { "type": "title", "bbox": [ 128, 526, 160, 535 ], "lines": [ { "bbox": [ 127, 525, 162, 535 ], "spans": [ { "bbox": [ 127, 525, 162, 535 ], "score": 1.0, "content": "LLaVA:", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 127, 536, 484, 556 ], "lines": [ { "bbox": [ 126, 533, 486, 547 ], "spans": [ { "bbox": [ 126, 533, 486, 547 ], "score": 1.0, "content": "A black and silver Kawasaki motorcycle which is likely used for transportation, commuting, leisure", "type": "text" } ], "index": 33 }, { "bbox": [ 127, 544, 235, 557 ], "spans": [ { "bbox": [ 127, 544, 235, 557 ], "score": 1.0, "content": "rides, or long-distance travel.", "type": "text" } ], "index": 34 } ], "index": 33.5 }, { "type": "title", "bbox": [ 128, 556, 171, 565 ], "lines": [ { "bbox": [ 128, 555, 172, 565 ], "spans": [ { "bbox": [ 128, 555, 172, 565 ], "score": 1.0, "content": "Kosmos-2:", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 128, 565, 465, 576 ], "lines": [ { "bbox": [ 126, 564, 467, 577 ], "spans": [ { "bbox": [ 126, 564, 467, 577 ], "score": 1.0, "content": "The purpose of the object on the motorcycle is to store and protect the motorcycle’s fuel tank.", "type": "text" } ], "index": 36 } ], "index": 36 }, { "type": "text", "bbox": [ 128, 576, 194, 600 ], "lines": [ { "bbox": [ 127, 574, 160, 586 ], "spans": [ { "bbox": [ 127, 574, 160, 586 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 37 }, { "bbox": [ 126, 581, 195, 598 ], "spans": [ { "bbox": [ 126, 581, 195, 598 ], "score": 1.0, "content": "The answer is leg.", "type": "text" } ], "index": 38 } ], "index": 37.5 }, { "type": "title", "bbox": [ 129, 596, 185, 605 ], "lines": [ { "bbox": [ 127, 595, 186, 606 ], "spans": [ { "bbox": [ 127, 595, 186, 606 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 39 } ], "index": 39 }, { "type": "text", "bbox": [ 128, 606, 484, 646 ], "lines": [ { "bbox": [ 126, 604, 486, 618 ], "spans": [ { "bbox": [ 126, 604, 486, 618 ], "score": 1.0, "content": "The object is a shock absorber. This component is essential for the bike’s stability, providing a", "type": "text" } ], "index": 40 }, { "bbox": [ 127, 615, 486, 627 ], "spans": [ { "bbox": [ 127, 615, 486, 627 ], "score": 1.0, "content": "smoother ride by soaking up the impact from bumps and uneven surfaces. It significantly improves", "type": "text" } ], "index": 41 }, { "bbox": [ 127, 624, 486, 637 ], "spans": [ { "bbox": [ 127, 624, 486, 637 ], "score": 1.0, "content": "the overall performance and comfort of the bike, enabling the rider to maintain control and stability", "type": "text" } ], "index": 42 }, { "bbox": [ 126, 636, 213, 647 ], "spans": [ { "bbox": [ 126, 636, 213, 647 ], "score": 1.0, "content": "even on rough terrains.", "type": "text" } ], "index": 43 } ], "index": 41.5 } ], "page_idx": 26, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 765 ], "spans": [ { "bbox": [ 298, 750, 313, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 153, 505, 187 ], "lines": [ { "bbox": [ 105, 153, 505, 166 ], "spans": [ { "bbox": [ 105, 153, 505, 166 ], "score": 1.0, "content": "Table 18: Referring Reasoning in Ferret-Bench. Qualitative examples to illustrate the difference", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 165, 505, 177 ], "spans": [ { "bbox": [ 106, 165, 505, 177 ], "score": 1.0, "content": "between various models (LLaVA vs. Kosmos-2 vs. Shikra vs. Ferret (Ours)). For clarity, we have", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 176, 363, 188 ], "spans": [ { "bbox": [ 106, 176, 363, 188 ], "score": 1.0, "content": "omitted the bounding box outputs in the textual representations.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 105, 153, 505, 188 ] }, { "type": "title", "bbox": [ 128, 207, 167, 216 ], "lines": [ { "bbox": [ 127, 205, 168, 217 ], "spans": [ { "bbox": [ 127, 205, 168, 217 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 130, 217, 288, 227 ], "lines": [ { "bbox": [ 128, 216, 289, 228 ], "spans": [ { "bbox": [ 128, 216, 289, 228 ], "score": 1.0, "content": "What is the purpose of the object [region0]?", "type": "text" } ], "index": 4 } ], "index": 4, "bbox_fs": [ 128, 216, 289, 228 ] }, { "type": "title", "bbox": [ 128, 237, 187, 246 ], "lines": [ { "bbox": [ 128, 236, 187, 246 ], "spans": [ { "bbox": [ 128, 236, 187, 246 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 128, 247, 329, 277 ], "lines": [ { "bbox": [ 127, 245, 312, 258 ], "spans": [ { "bbox": [ 127, 245, 312, 258 ], "score": 1.0, "content": "The object is a bottle. Bottles are typically used to", "type": "text" } ], "index": 6 }, { "bbox": [ 126, 255, 330, 269 ], "spans": [ { "bbox": [ 126, 255, 330, 269 ], "score": 1.0, "content": "hold and store different types of liquids. In this context,", "type": "text" } ], "index": 7 }, { "bbox": [ 127, 266, 305, 278 ], "spans": [ { "bbox": [ 127, 266, 305, 278 ], "score": 1.0, "content": "it might be used to store a beverage for the meal.", "type": "text" } ], "index": 8 } ], "index": 7, "bbox_fs": [ 126, 245, 330, 278 ] }, { "type": "image", "bbox": [ 348, 208, 498, 307 ], "blocks": [ { "type": "image_body", "bbox": [ 348, 208, 498, 307 ], "group_id": 0, "lines": [ { "bbox": [ 348, 208, 498, 307 ], "spans": [ { "bbox": [ 348, 208, 498, 307 ], "score": 0.971, "type": "image", "image_path": "c6836caac403445c88c4fb683267e9ac1e0024968f18020c9fedb9af5f282cb6.jpg" } ] } ], "index": 9.5, "virtual_lines": [ { "bbox": [ 348, 208, 498, 257.5 ], "spans": [], "index": 9 }, { "bbox": [ 348, 257.5, 498, 307.0 ], "spans": [], "index": 10 } ] } ], "index": 9.5 }, { "type": "title", "bbox": [ 128, 287, 160, 296 ], "lines": [ { "bbox": [ 127, 286, 162, 296 ], "spans": [ { "bbox": [ 127, 286, 162, 296 ], "score": 1.0, "content": "LLaVA:", "type": "text" } ], "index": 11 } ], "index": 11 }, { "type": "text", "bbox": [ 130, 297, 329, 307 ], "lines": [ { "bbox": [ 129, 296, 330, 307 ], "spans": [ { "bbox": [ 129, 296, 330, 307 ], "score": 1.0, "content": "A knife and a fork, indicating that it is ready for a meal.", "type": "text" } ], "index": 12 } ], "index": 12, "bbox_fs": [ 129, 296, 330, 307 ] }, { "type": "text", "bbox": [ 130, 317, 340, 327 ], "lines": [ { "bbox": [ 129, 316, 342, 327 ], "spans": [ { "bbox": [ 129, 316, 342, 327 ], "score": 1.0, "content": "The purpose of the object is to attract the birds to the table.", "type": "text" } ], "index": 13 } ], "index": 13, "bbox_fs": [ 129, 316, 342, 327 ] }, { "type": "text", "bbox": [ 128, 327, 241, 346 ], "lines": [ { "bbox": [ 127, 325, 162, 338 ], "spans": [ { "bbox": [ 127, 325, 162, 338 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 14 }, { "bbox": [ 127, 334, 243, 348 ], "spans": [ { "bbox": [ 127, 334, 243, 348 ], "score": 1.0, "content": "The answer is keep birds away.", "type": "text" } ], "index": 15 } ], "index": 14.5, "bbox_fs": [ 127, 325, 243, 348 ] }, { "type": "text", "bbox": [ 129, 347, 184, 356 ], "lines": [ { "bbox": [ 127, 345, 186, 357 ], "spans": [ { "bbox": [ 127, 345, 186, 357 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 16 } ], "index": 16, "bbox_fs": [ 127, 345, 186, 357 ] }, { "type": "text", "bbox": [ 128, 357, 483, 376 ], "lines": [ { "bbox": [ 127, 354, 484, 369 ], "spans": [ { "bbox": [ 127, 354, 484, 369 ], "score": 1.0, "content": "The object is a bottle, generally used to store and dispense liquids like water, juice, or other bever-", "type": "text" } ], "index": 17 }, { "bbox": [ 127, 367, 149, 378 ], "spans": [ { "bbox": [ 127, 367, 149, 378 ], "score": 1.0, "content": "ages.", "type": "text" } ], "index": 18 } ], "index": 17.5, "bbox_fs": [ 127, 354, 484, 378 ] }, { "type": "title", "bbox": [ 128, 397, 167, 406 ], "lines": [ { "bbox": [ 127, 395, 169, 407 ], "spans": [ { "bbox": [ 127, 395, 169, 407 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 128, 406, 330, 417 ], "lines": [ { "bbox": [ 127, 405, 332, 417 ], "spans": [ { "bbox": [ 127, 405, 332, 417 ], "score": 1.0, "content": "What is the purpose of the object [region0] on the bike?", "type": "text" } ], "index": 20 } ], "index": 20, "bbox_fs": [ 127, 405, 332, 417 ] }, { "type": "title", "bbox": [ 128, 426, 186, 435 ], "lines": [ { "bbox": [ 128, 426, 187, 436 ], "spans": [ { "bbox": [ 128, 426, 187, 436 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 21 } ], "index": 21 }, { "type": "text", "bbox": [ 128, 437, 315, 516 ], "lines": [ { "bbox": [ 127, 435, 284, 447 ], "spans": [ { "bbox": [ 127, 435, 284, 447 ], "score": 1.0, "content": "The object is a shock absorber on the bike.", "type": "text" } ], "index": 22 }, { "bbox": [ 127, 446, 313, 457 ], "spans": [ { "bbox": [ 127, 446, 313, 457 ], "score": 1.0, "content": "Its purpose is to absorb or dampen shock impulses.", "type": "text" } ], "index": 23 }, { "bbox": [ 127, 455, 298, 467 ], "spans": [ { "bbox": [ 127, 455, 298, 467 ], "score": 1.0, "content": "It does this by converting the kinetic energy of", "type": "text" } ], "index": 24 }, { "bbox": [ 127, 465, 316, 477 ], "spans": [ { "bbox": [ 127, 465, 316, 477 ], "score": 1.0, "content": "the shock into another form of energy which is then", "type": "text" } ], "index": 25 }, { "bbox": [ 128, 476, 288, 487 ], "spans": [ { "bbox": [ 128, 476, 288, 487 ], "score": 1.0, "content": "dissipated. In the context of the motorcycle,", "type": "text" } ], "index": 26 }, { "bbox": [ 127, 486, 303, 497 ], "spans": [ { "bbox": [ 127, 486, 303, 497 ], "score": 1.0, "content": "it is particularly useful in providing comfort and", "type": "text" } ], "index": 27 }, { "bbox": [ 127, 493, 303, 509 ], "spans": [ { "bbox": [ 127, 493, 303, 509 ], "score": 1.0, "content": "stability for the rider, especially when travelling", "type": "text" } ], "index": 28 }, { "bbox": [ 128, 506, 238, 516 ], "spans": [ { "bbox": [ 128, 506, 238, 516 ], "score": 1.0, "content": "over uneven or rough terrains.", "type": "text" } ], "index": 29 } ], "index": 25.5, "bbox_fs": [ 127, 435, 316, 516 ] }, { "type": "image", "bbox": [ 350, 396, 495, 505 ], "blocks": [ { "type": "image_body", "bbox": [ 350, 396, 495, 505 ], "group_id": 1, "lines": [ { "bbox": [ 350, 396, 495, 505 ], "spans": [ { "bbox": [ 350, 396, 495, 505 ], "score": 0.966, "type": "image", "image_path": "374730542b5bbe01698e33da5f84294f1c9442fbaeaf8dc837c63230cf8b2069.jpg" } ] } ], "index": 30.5, "virtual_lines": [ { "bbox": [ 350, 396, 495, 450.5 ], "spans": [], "index": 30 }, { "bbox": [ 350, 450.5, 495, 505.0 ], "spans": [], "index": 31 } ] } ], "index": 30.5 }, { "type": "title", "bbox": [ 128, 526, 160, 535 ], "lines": [ { "bbox": [ 127, 525, 162, 535 ], "spans": [ { "bbox": [ 127, 525, 162, 535 ], "score": 1.0, "content": "LLaVA:", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 127, 536, 484, 556 ], "lines": [ { "bbox": [ 126, 533, 486, 547 ], "spans": [ { "bbox": [ 126, 533, 486, 547 ], "score": 1.0, "content": "A black and silver Kawasaki motorcycle which is likely used for transportation, commuting, leisure", "type": "text" } ], "index": 33 }, { "bbox": [ 127, 544, 235, 557 ], "spans": [ { "bbox": [ 127, 544, 235, 557 ], "score": 1.0, "content": "rides, or long-distance travel.", "type": "text" } ], "index": 34 } ], "index": 33.5, "bbox_fs": [ 126, 533, 486, 557 ] }, { "type": "title", "bbox": [ 128, 556, 171, 565 ], "lines": [ { "bbox": [ 128, 555, 172, 565 ], "spans": [ { "bbox": [ 128, 555, 172, 565 ], "score": 1.0, "content": "Kosmos-2:", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 128, 565, 465, 576 ], "lines": [ { "bbox": [ 126, 564, 467, 577 ], "spans": [ { "bbox": [ 126, 564, 467, 577 ], "score": 1.0, "content": "The purpose of the object on the motorcycle is to store and protect the motorcycle’s fuel tank.", "type": "text" } ], "index": 36 } ], "index": 36, "bbox_fs": [ 126, 564, 467, 577 ] }, { "type": "text", "bbox": [ 128, 576, 194, 600 ], "lines": [ { "bbox": [ 127, 574, 160, 586 ], "spans": [ { "bbox": [ 127, 574, 160, 586 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 37 }, { "bbox": [ 126, 581, 195, 598 ], "spans": [ { "bbox": [ 126, 581, 195, 598 ], "score": 1.0, "content": "The answer is leg.", "type": "text" } ], "index": 38 } ], "index": 37.5, "bbox_fs": [ 126, 574, 195, 598 ] }, { "type": "title", "bbox": [ 129, 596, 185, 605 ], "lines": [ { "bbox": [ 127, 595, 186, 606 ], "spans": [ { "bbox": [ 127, 595, 186, 606 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 39 } ], "index": 39 }, { "type": "text", "bbox": [ 128, 606, 484, 646 ], "lines": [ { "bbox": [ 126, 604, 486, 618 ], "spans": [ { "bbox": [ 126, 604, 486, 618 ], "score": 1.0, "content": "The object is a shock absorber. This component is essential for the bike’s stability, providing a", "type": "text" } ], "index": 40 }, { "bbox": [ 127, 615, 486, 627 ], "spans": [ { "bbox": [ 127, 615, 486, 627 ], "score": 1.0, "content": "smoother ride by soaking up the impact from bumps and uneven surfaces. It significantly improves", "type": "text" } ], "index": 41 }, { "bbox": [ 127, 624, 486, 637 ], "spans": [ { "bbox": [ 127, 624, 486, 637 ], "score": 1.0, "content": "the overall performance and comfort of the bike, enabling the rider to maintain control and stability", "type": "text" } ], "index": 42 }, { "bbox": [ 126, 636, 213, 647 ], "spans": [ { "bbox": [ 126, 636, 213, 647 ], "score": 1.0, "content": "even on rough terrains.", "type": "text" } ], "index": 43 } ], "index": 41.5, "bbox_fs": [ 126, 604, 486, 647 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 173, 505, 207 ], "lines": [ { "bbox": [ 106, 174, 504, 185 ], "spans": [ { "bbox": [ 106, 174, 504, 185 ], "score": 1.0, "content": "Table 19: Grounding in Conversation in Ferret-Bench. Qualitative examples to illustrate the", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 183, 504, 198 ], "spans": [ { "bbox": [ 106, 183, 504, 198 ], "score": 1.0, "content": "difference between various models (Kosmos-2 vs. Shikra vs. Ferret (Ours)). For clarity, we only", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 196, 330, 208 ], "spans": [ { "bbox": [ 106, 196, 330, 208 ], "score": 1.0, "content": "display the bounding box outputs in Ferret ’s responses.", "type": "text" } ], "index": 2 } ], "index": 1 }, { "type": "title", "bbox": [ 128, 227, 167, 236 ], "lines": [ { "bbox": [ 127, 226, 168, 237 ], "spans": [ { "bbox": [ 127, 226, 168, 237 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 128, 237, 295, 257 ], "lines": [ { "bbox": [ 128, 236, 254, 247 ], "spans": [ { "bbox": [ 128, 236, 254, 247 ], "score": 1.0, "content": "What is happening in the field and", "type": "text" } ], "index": 4 }, { "bbox": [ 127, 245, 297, 257 ], "spans": [ { "bbox": [ 127, 245, 297, 257 ], "score": 1.0, "content": "provide the coordinates of mentioned objects?", "type": "text" } ], "index": 5 } ], "index": 4.5 }, { "type": "title", "bbox": [ 128, 267, 186, 276 ], "lines": [ { "bbox": [ 127, 266, 188, 277 ], "spans": [ { "bbox": [ 127, 266, 188, 277 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 128, 277, 319, 316 ], "lines": [ { "bbox": [ 127, 276, 321, 288 ], "spans": [ { "bbox": [ 127, 276, 321, 288 ], "score": 1.0, "content": "In the image, there is a group of sheep [box0] [box1]", "type": "text" } ], "index": 7 }, { "bbox": [ 126, 286, 302, 297 ], "spans": [ { "bbox": [ 126, 286, 302, 297 ], "score": 1.0, "content": "grazing in a field [box2]. The field is filled with", "type": "text" } ], "index": 8 }, { "bbox": [ 127, 297, 307, 308 ], "spans": [ { "bbox": [ 127, 297, 307, 308 ], "score": 1.0, "content": "plant life [box3]. There is also a bush [box4] and", "type": "text" } ], "index": 10 }, { "bbox": [ 127, 306, 266, 317 ], "spans": [ { "bbox": [ 127, 306, 266, 317 ], "score": 1.0, "content": "some trees [box5] present in the field.", "type": "text" } ], "index": 11 } ], "index": 9.0 }, { "type": "image", "bbox": [ 368, 227, 474, 385 ], "blocks": [ { "type": "image_body", "bbox": [ 368, 227, 474, 385 ], "group_id": 0, "lines": [ { "bbox": [ 368, 227, 474, 385 ], "spans": [ { "bbox": [ 368, 227, 474, 385 ], "score": 0.965, "type": "image", "image_path": "9dc7c8f3f72a1f94103f23551847f05670ecec30399b83846c3717808b7c7ab6.jpg" } ] } ], "index": 9, "virtual_lines": [ { "bbox": [ 368, 227, 474, 385 ], "spans": [], "index": 9 } ] } ], "index": 9 }, { "type": "title", "bbox": [ 129, 327, 171, 336 ], "lines": [ { "bbox": [ 127, 325, 172, 336 ], "spans": [ { "bbox": [ 127, 325, 172, 336 ], "score": 1.0, "content": "Kosmos-2:", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 128, 336, 316, 356 ], "lines": [ { "bbox": [ 126, 335, 317, 348 ], "spans": [ { "bbox": [ 126, 335, 317, 348 ], "score": 1.0, "content": "In the field, a herd of sheep [box0] is grazing on the", "type": "text" } ], "index": 13 }, { "bbox": [ 127, 345, 217, 358 ], "spans": [ { "bbox": [ 127, 345, 217, 358 ], "score": 1.0, "content": "lush green grass [box1].", "type": "text" } ], "index": 14 } ], "index": 13.5 }, { "type": "title", "bbox": [ 128, 357, 159, 366 ], "lines": [ { "bbox": [ 128, 355, 161, 367 ], "spans": [ { "bbox": [ 128, 355, 161, 367 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 15 } ], "index": 15 }, { "type": "text", "bbox": [ 128, 366, 349, 386 ], "lines": [ { "bbox": [ 126, 365, 350, 379 ], "spans": [ { "bbox": [ 126, 365, 350, 379 ], "score": 1.0, "content": "A group of sheep [box0] [box1] [box2] [box3] [box4] [box5]", "type": "text" } ], "index": 16 }, { "bbox": [ 127, 376, 231, 387 ], "spans": [ { "bbox": [ 127, 376, 231, 387 ], "score": 1.0, "content": "grazing on lush green grass.", "type": "text" } ], "index": 17 } ], "index": 16.5 }, { "type": "title", "bbox": [ 129, 387, 185, 396 ], "lines": [ { "bbox": [ 127, 386, 186, 397 ], "spans": [ { "bbox": [ 127, 386, 186, 397 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 127, 396, 484, 407 ], "lines": [ { "bbox": [ 126, 395, 484, 408 ], "spans": [ { "bbox": [ 126, 395, 484, 408 ], "score": 1.0, "content": "In the field, there are sheep [box0] [box1] [box2] [box3] [box4] [box5] grazing on the grass [box6].", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "title", "bbox": [ 128, 426, 167, 435 ], "lines": [ { "bbox": [ 127, 425, 168, 437 ], "spans": [ { "bbox": [ 127, 425, 168, 437 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 128, 436, 345, 456 ], "lines": [ { "bbox": [ 127, 435, 346, 447 ], "spans": [ { "bbox": [ 127, 435, 346, 447 ], "score": 1.0, "content": "What is the man doing and what objects are in front of him?", "type": "text" } ], "index": 21 }, { "bbox": [ 127, 445, 300, 457 ], "spans": [ { "bbox": [ 127, 445, 300, 457 ], "score": 1.0, "content": "Please provide the coordinates of these objects.", "type": "text" } ], "index": 22 } ], "index": 21.5 }, { "type": "image", "bbox": [ 367, 426, 474, 568 ], "blocks": [ { "type": "image_body", "bbox": [ 367, 426, 474, 568 ], "group_id": 1, "lines": [ { "bbox": [ 367, 426, 474, 568 ], "spans": [ { "bbox": [ 367, 426, 474, 568 ], "score": 0.965, "type": "image", "image_path": "823220ca35cb50018bb7557cb0f1ceec3ef29a443701734e6c1ed5c576445fcf.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 367, 426, 474, 568 ], "spans": [], "index": 23 } ] } ], "index": 23 }, { "type": "title", "bbox": [ 128, 466, 186, 475 ], "lines": [ { "bbox": [ 127, 465, 188, 476 ], "spans": [ { "bbox": [ 127, 465, 188, 476 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 128, 476, 326, 506 ], "lines": [ { "bbox": [ 127, 474, 328, 488 ], "spans": [ { "bbox": [ 127, 474, 328, 488 ], "score": 1.0, "content": "The man [box0] is sitting at a table [box1] and looking", "type": "text" } ], "index": 25 }, { "bbox": [ 127, 486, 307, 497 ], "spans": [ { "bbox": [ 127, 486, 307, 497 ], "score": 1.0, "content": "at a plate of food [box2] . There is a glass [box3]", "type": "text" } ], "index": 26 }, { "bbox": [ 127, 495, 299, 507 ], "spans": [ { "bbox": [ 127, 495, 299, 507 ], "score": 1.0, "content": "and a bottle [box4] on the table in front of him.", "type": "text" } ], "index": 27 } ], "index": 26 }, { "type": "title", "bbox": [ 128, 516, 170, 525 ], "lines": [ { "bbox": [ 127, 515, 172, 526 ], "spans": [ { "bbox": [ 127, 515, 172, 526 ], "score": 1.0, "content": "Kosmos-2:", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 128, 526, 325, 546 ], "lines": [ { "bbox": [ 127, 525, 326, 536 ], "spans": [ { "bbox": [ 127, 525, 326, 536 ], "score": 1.0, "content": "The man is sitting at a dining table [box0] smiling and", "type": "text" } ], "index": 29 }, { "bbox": [ 127, 534, 302, 547 ], "spans": [ { "bbox": [ 127, 534, 302, 547 ], "score": 1.0, "content": "looking at the plate of food in front him [box1].", "type": "text" } ], "index": 30 } ], "index": 29.5 }, { "type": "title", "bbox": [ 128, 546, 159, 555 ], "lines": [ { "bbox": [ 128, 545, 161, 556 ], "spans": [ { "bbox": [ 128, 545, 161, 556 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "text", "bbox": [ 128, 556, 345, 586 ], "lines": [ { "bbox": [ 127, 555, 326, 567 ], "spans": [ { "bbox": [ 127, 555, 326, 567 ], "score": 1.0, "content": "The man [box0] is sitting at a table [box1] with a plate", "type": "text" } ], "index": 32 }, { "bbox": [ 127, 565, 318, 577 ], "spans": [ { "bbox": [ 127, 565, 318, 577 ], "score": 1.0, "content": "of food [box2] in front of him. He is looking up and", "type": "text" } ], "index": 33 }, { "bbox": [ 127, 575, 347, 587 ], "spans": [ { "bbox": [ 127, 575, 347, 587 ], "score": 1.0, "content": "smiling, possibly staring at a person [box0] in the restaurant", "type": "text" } ], "index": 34 } ], "index": 33 }, { "type": "title", "bbox": [ 128, 586, 185, 595 ], "lines": [ { "bbox": [ 127, 585, 187, 596 ], "spans": [ { "bbox": [ 127, 585, 187, 596 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 128, 596, 484, 626 ], "lines": [ { "bbox": [ 127, 594, 486, 607 ], "spans": [ { "bbox": [ 127, 594, 486, 607 ], "score": 1.0, "content": "The man [box0] is having a meal at a table [box1]. In front of him, there is a plate [box2] with a", "type": "text" } ], "index": 36 }, { "bbox": [ 127, 604, 486, 617 ], "spans": [ { "bbox": [ 127, 604, 486, 617 ], "score": 1.0, "content": "variety of food items such as fruit and eggs. There is also a glass [box3] filled with water and a", "type": "text" } ], "index": 37 }, { "bbox": [ 127, 614, 179, 627 ], "spans": [ { "bbox": [ 127, 614, 179, 627 ], "score": 1.0, "content": "bottle [box4].", "type": "text" } ], "index": 38 } ], "index": 37 } ], "page_idx": 27, "page_size": [ 612, 792 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 751, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 763 ], "spans": [ { "bbox": [ 298, 750, 313, 763 ], "score": 1.0, "content": "28", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 173, 505, 207 ], "lines": [ { "bbox": [ 106, 174, 504, 185 ], "spans": [ { "bbox": [ 106, 174, 504, 185 ], "score": 1.0, "content": "Table 19: Grounding in Conversation in Ferret-Bench. Qualitative examples to illustrate the", "type": "text" } ], "index": 0 }, { "bbox": [ 106, 183, 504, 198 ], "spans": [ { "bbox": [ 106, 183, 504, 198 ], "score": 1.0, "content": "difference between various models (Kosmos-2 vs. Shikra vs. Ferret (Ours)). For clarity, we only", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 196, 330, 208 ], "spans": [ { "bbox": [ 106, 196, 330, 208 ], "score": 1.0, "content": "display the bounding box outputs in Ferret ’s responses.", "type": "text" } ], "index": 2 } ], "index": 1, "bbox_fs": [ 106, 174, 504, 208 ] }, { "type": "title", "bbox": [ 128, 227, 167, 236 ], "lines": [ { "bbox": [ 127, 226, 168, 237 ], "spans": [ { "bbox": [ 127, 226, 168, 237 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 3 } ], "index": 3 }, { "type": "text", "bbox": [ 128, 237, 295, 257 ], "lines": [ { "bbox": [ 128, 236, 254, 247 ], "spans": [ { "bbox": [ 128, 236, 254, 247 ], "score": 1.0, "content": "What is happening in the field and", "type": "text" } ], "index": 4 }, { "bbox": [ 127, 245, 297, 257 ], "spans": [ { "bbox": [ 127, 245, 297, 257 ], "score": 1.0, "content": "provide the coordinates of mentioned objects?", "type": "text" } ], "index": 5 } ], "index": 4.5, "bbox_fs": [ 127, 236, 297, 257 ] }, { "type": "title", "bbox": [ 128, 267, 186, 276 ], "lines": [ { "bbox": [ 127, 266, 188, 277 ], "spans": [ { "bbox": [ 127, 266, 188, 277 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 6 } ], "index": 6 }, { "type": "text", "bbox": [ 128, 277, 319, 316 ], "lines": [ { "bbox": [ 127, 276, 321, 288 ], "spans": [ { "bbox": [ 127, 276, 321, 288 ], "score": 1.0, "content": "In the image, there is a group of sheep [box0] [box1]", "type": "text" } ], "index": 7 }, { "bbox": [ 126, 286, 302, 297 ], "spans": [ { "bbox": [ 126, 286, 302, 297 ], "score": 1.0, "content": "grazing in a field [box2]. The field is filled with", "type": "text" } ], "index": 8 }, { "bbox": [ 127, 297, 307, 308 ], "spans": [ { "bbox": [ 127, 297, 307, 308 ], "score": 1.0, "content": "plant life [box3]. There is also a bush [box4] and", "type": "text" } ], "index": 10 }, { "bbox": [ 127, 306, 266, 317 ], "spans": [ { "bbox": [ 127, 306, 266, 317 ], "score": 1.0, "content": "some trees [box5] present in the field.", "type": "text" } ], "index": 11 } ], "index": 9.0, "bbox_fs": [ 126, 276, 321, 317 ] }, { "type": "image", "bbox": [ 368, 227, 474, 385 ], "blocks": [ { "type": "image_body", "bbox": [ 368, 227, 474, 385 ], "group_id": 0, "lines": [ { "bbox": [ 368, 227, 474, 385 ], "spans": [ { "bbox": [ 368, 227, 474, 385 ], "score": 0.965, "type": "image", "image_path": "9dc7c8f3f72a1f94103f23551847f05670ecec30399b83846c3717808b7c7ab6.jpg" } ] } ], "index": 9, "virtual_lines": [ { "bbox": [ 368, 227, 474, 385 ], "spans": [], "index": 9 } ] } ], "index": 9 }, { "type": "title", "bbox": [ 129, 327, 171, 336 ], "lines": [ { "bbox": [ 127, 325, 172, 336 ], "spans": [ { "bbox": [ 127, 325, 172, 336 ], "score": 1.0, "content": "Kosmos-2:", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 128, 336, 316, 356 ], "lines": [ { "bbox": [ 126, 335, 317, 348 ], "spans": [ { "bbox": [ 126, 335, 317, 348 ], "score": 1.0, "content": "In the field, a herd of sheep [box0] is grazing on the", "type": "text" } ], "index": 13 }, { "bbox": [ 127, 345, 217, 358 ], "spans": [ { "bbox": [ 127, 345, 217, 358 ], "score": 1.0, "content": "lush green grass [box1].", "type": "text" } ], "index": 14 } ], "index": 13.5, "bbox_fs": [ 126, 335, 317, 358 ] }, { "type": "title", "bbox": [ 128, 357, 159, 366 ], "lines": [ { "bbox": [ 128, 355, 161, 367 ], "spans": [ { "bbox": [ 128, 355, 161, 367 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 15 } ], "index": 15 }, { "type": "text", "bbox": [ 128, 366, 349, 386 ], "lines": [ { "bbox": [ 126, 365, 350, 379 ], "spans": [ { "bbox": [ 126, 365, 350, 379 ], "score": 1.0, "content": "A group of sheep [box0] [box1] [box2] [box3] [box4] [box5]", "type": "text" } ], "index": 16 }, { "bbox": [ 127, 376, 231, 387 ], "spans": [ { "bbox": [ 127, 376, 231, 387 ], "score": 1.0, "content": "grazing on lush green grass.", "type": "text" } ], "index": 17 } ], "index": 16.5, "bbox_fs": [ 126, 365, 350, 387 ] }, { "type": "title", "bbox": [ 129, 387, 185, 396 ], "lines": [ { "bbox": [ 127, 386, 186, 397 ], "spans": [ { "bbox": [ 127, 386, 186, 397 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 127, 396, 484, 407 ], "lines": [ { "bbox": [ 126, 395, 484, 408 ], "spans": [ { "bbox": [ 126, 395, 484, 408 ], "score": 1.0, "content": "In the field, there are sheep [box0] [box1] [box2] [box3] [box4] [box5] grazing on the grass [box6].", "type": "text" } ], "index": 19 } ], "index": 19, "bbox_fs": [ 126, 395, 484, 408 ] }, { "type": "title", "bbox": [ 128, 426, 167, 435 ], "lines": [ { "bbox": [ 127, 425, 168, 437 ], "spans": [ { "bbox": [ 127, 425, 168, 437 ], "score": 1.0, "content": "Question:", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 128, 436, 345, 456 ], "lines": [ { "bbox": [ 127, 435, 346, 447 ], "spans": [ { "bbox": [ 127, 435, 346, 447 ], "score": 1.0, "content": "What is the man doing and what objects are in front of him?", "type": "text" } ], "index": 21 }, { "bbox": [ 127, 445, 300, 457 ], "spans": [ { "bbox": [ 127, 445, 300, 457 ], "score": 1.0, "content": "Please provide the coordinates of these objects.", "type": "text" } ], "index": 22 } ], "index": 21.5, "bbox_fs": [ 127, 435, 346, 457 ] }, { "type": "image", "bbox": [ 367, 426, 474, 568 ], "blocks": [ { "type": "image_body", "bbox": [ 367, 426, 474, 568 ], "group_id": 1, "lines": [ { "bbox": [ 367, 426, 474, 568 ], "spans": [ { "bbox": [ 367, 426, 474, 568 ], "score": 0.965, "type": "image", "image_path": "823220ca35cb50018bb7557cb0f1ceec3ef29a443701734e6c1ed5c576445fcf.jpg" } ] } ], "index": 23, "virtual_lines": [ { "bbox": [ 367, 426, 474, 568 ], "spans": [], "index": 23 } ] } ], "index": 23 }, { "type": "title", "bbox": [ 128, 466, 186, 475 ], "lines": [ { "bbox": [ 127, 465, 188, 476 ], "spans": [ { "bbox": [ 127, 465, 188, 476 ], "score": 1.0, "content": "Ground Truth:", "type": "text" } ], "index": 24 } ], "index": 24 }, { "type": "text", "bbox": [ 128, 476, 326, 506 ], "lines": [ { "bbox": [ 127, 474, 328, 488 ], "spans": [ { "bbox": [ 127, 474, 328, 488 ], "score": 1.0, "content": "The man [box0] is sitting at a table [box1] and looking", "type": "text" } ], "index": 25 }, { "bbox": [ 127, 486, 307, 497 ], "spans": [ { "bbox": [ 127, 486, 307, 497 ], "score": 1.0, "content": "at a plate of food [box2] . There is a glass [box3]", "type": "text" } ], "index": 26 }, { "bbox": [ 127, 495, 299, 507 ], "spans": [ { "bbox": [ 127, 495, 299, 507 ], "score": 1.0, "content": "and a bottle [box4] on the table in front of him.", "type": "text" } ], "index": 27 } ], "index": 26, "bbox_fs": [ 127, 474, 328, 507 ] }, { "type": "title", "bbox": [ 128, 516, 170, 525 ], "lines": [ { "bbox": [ 127, 515, 172, 526 ], "spans": [ { "bbox": [ 127, 515, 172, 526 ], "score": 1.0, "content": "Kosmos-2:", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 128, 526, 325, 546 ], "lines": [ { "bbox": [ 127, 525, 326, 536 ], "spans": [ { "bbox": [ 127, 525, 326, 536 ], "score": 1.0, "content": "The man is sitting at a dining table [box0] smiling and", "type": "text" } ], "index": 29 }, { "bbox": [ 127, 534, 302, 547 ], "spans": [ { "bbox": [ 127, 534, 302, 547 ], "score": 1.0, "content": "looking at the plate of food in front him [box1].", "type": "text" } ], "index": 30 } ], "index": 29.5, "bbox_fs": [ 127, 525, 326, 547 ] }, { "type": "title", "bbox": [ 128, 546, 159, 555 ], "lines": [ { "bbox": [ 128, 545, 161, 556 ], "spans": [ { "bbox": [ 128, 545, 161, 556 ], "score": 1.0, "content": "Shikra:", "type": "text" } ], "index": 31 } ], "index": 31 }, { "type": "text", "bbox": [ 128, 556, 345, 586 ], "lines": [ { "bbox": [ 127, 555, 326, 567 ], "spans": [ { "bbox": [ 127, 555, 326, 567 ], "score": 1.0, "content": "The man [box0] is sitting at a table [box1] with a plate", "type": "text" } ], "index": 32 }, { "bbox": [ 127, 565, 318, 577 ], "spans": [ { "bbox": [ 127, 565, 318, 577 ], "score": 1.0, "content": "of food [box2] in front of him. He is looking up and", "type": "text" } ], "index": 33 }, { "bbox": [ 127, 575, 347, 587 ], "spans": [ { "bbox": [ 127, 575, 347, 587 ], "score": 1.0, "content": "smiling, possibly staring at a person [box0] in the restaurant", "type": "text" } ], "index": 34 } ], "index": 33, "bbox_fs": [ 127, 555, 347, 587 ] }, { "type": "title", "bbox": [ 128, 586, 185, 595 ], "lines": [ { "bbox": [ 127, 585, 187, 596 ], "spans": [ { "bbox": [ 127, 585, 187, 596 ], "score": 1.0, "content": "Ferret (Ours):", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 128, 596, 484, 626 ], "lines": [ { "bbox": [ 127, 594, 486, 607 ], "spans": [ { "bbox": [ 127, 594, 486, 607 ], "score": 1.0, "content": "The man [box0] is having a meal at a table [box1]. In front of him, there is a plate [box2] with a", "type": "text" } ], "index": 36 }, { "bbox": [ 127, 604, 486, 617 ], "spans": [ { "bbox": [ 127, 604, 486, 617 ], "score": 1.0, "content": "variety of food items such as fruit and eggs. There is also a glass [box3] filled with water and a", "type": "text" } ], "index": 37 }, { "bbox": [ 127, 614, 179, 627 ], "spans": [ { "bbox": [ 127, 614, 179, 627 ], "score": 1.0, "content": "bottle [box4].", "type": "text" } ], "index": 38 } ], "index": 37, "bbox_fs": [ 127, 594, 486, 627 ] } ] } ], "_backend": "pipeline", "_version_name": "2.1.11" }