diff --git "a/parse/dev/b9APFSTylGT/b9APFSTylGT_middle.json" "b/parse/dev/b9APFSTylGT/b9APFSTylGT_middle.json" new file mode 100644--- /dev/null +++ "b/parse/dev/b9APFSTylGT/b9APFSTylGT_middle.json" @@ -0,0 +1,41469 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 137, + 97, + 474, + 137 + ], + "lines": [ + { + "bbox": [ + 136, + 97, + 476, + 119 + ], + "spans": [ + { + "bbox": [ + 136, + 97, + 476, + 119 + ], + "score": 1.0, + "content": "Prompt Learning with Optimal Transport for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 212, + 116, + 399, + 139 + ], + "spans": [ + { + "bbox": [ + 212, + 116, + 399, + 139 + ], + "score": 1.0, + "content": "Vision-Language Models", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 259, + 179, + 354, + 223 + ], + "lines": [ + { + "bbox": [ + 258, + 178, + 356, + 191 + ], + "spans": [ + { + "bbox": [ + 258, + 178, + 356, + 191 + ], + "score": 1.0, + "content": "Anonymous Author(s)", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 283, + 189, + 328, + 202 + ], + "spans": [ + { + "bbox": [ + 283, + 189, + 328, + 202 + ], + "score": 1.0, + "content": "Affiliation", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 286, + 200, + 324, + 213 + ], + "spans": [ + { + "bbox": [ + 286, + 200, + 324, + 213 + ], + "score": 1.0, + "content": "Address", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 290, + 213, + 320, + 222 + ], + "spans": [ + { + "bbox": [ + 290, + 213, + 320, + 222 + ], + "score": 1.0, + "content": "email", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "title", + "bbox": [ + 283, + 252, + 328, + 265 + ], + "lines": [ + { + "bbox": [ + 281, + 251, + 331, + 267 + ], + "spans": [ + { + "bbox": [ + 281, + 251, + 331, + 267 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 91, + 276, + 469, + 430 + ], + "lines": [ + { + "bbox": [ + 93, + 276, + 470, + 289 + ], + "spans": [ + { + "bbox": [ + 93, + 280, + 99, + 287 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 141, + 276, + 470, + 289 + ], + "score": 1.0, + "content": "With the increasing attention to large vision-language models such as CLIP, there", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 287, + 471, + 300 + ], + "spans": [ + { + "bbox": [ + 93, + 290, + 99, + 298 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 141, + 287, + 471, + 300 + ], + "score": 1.0, + "content": "has been a significant amount of effort dedicated to building efficient prompts.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 297, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 93, + 301, + 99, + 309 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 141, + 297, + 470, + 312 + ], + "score": 1.0, + "content": "Unlike conventional methods of only learning one single prompt, we propose", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 92, + 309, + 470, + 322 + ], + "spans": [ + { + "bbox": [ + 92, + 312, + 99, + 321 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 141, + 309, + 470, + 322 + ], + "score": 1.0, + "content": "to learn multiple comprehensive prompts to describe diverse characteristics of", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 93, + 320, + 470, + 333 + ], + "spans": [ + { + "bbox": [ + 93, + 323, + 99, + 331 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 141, + 320, + 470, + 333 + ], + "score": 1.0, + "content": "categories such as intrinsic attributes or extrinsic contexts. However, directly", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 93, + 331, + 470, + 344 + ], + "spans": [ + { + "bbox": [ + 93, + 334, + 99, + 342 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 141, + 331, + 470, + 344 + ], + "score": 1.0, + "content": "matching each prompt to the same visual feature is problematic, as it pushes the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 91, + 342, + 469, + 355 + ], + "spans": [ + { + "bbox": [ + 91, + 344, + 99, + 353 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 141, + 342, + 469, + 355 + ], + "score": 1.0, + "content": "prompts to converge to one point. To solve this problem, we propose to apply", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 92, + 354, + 470, + 366 + ], + "spans": [ + { + "bbox": [ + 92, + 356, + 99, + 364 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 142, + 354, + 470, + 366 + ], + "score": 1.0, + "content": "optimal transport to match the vision and text modalities. Specifically, we first", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 92, + 363, + 470, + 376 + ], + "spans": [ + { + "bbox": [ + 92, + 366, + 100, + 376 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 141, + 363, + 470, + 376 + ], + "score": 1.0, + "content": "model images and the categories with visual and textual feature sets. Then, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 90, + 374, + 470, + 388 + ], + "spans": [ + { + "bbox": [ + 90, + 376, + 100, + 387 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 141, + 374, + 470, + 388 + ], + "score": 1.0, + "content": "apply a two-stage optimization strategy to learn the prompts. In the inner loop, we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 90, + 385, + 470, + 398 + ], + "spans": [ + { + "bbox": [ + 90, + 388, + 99, + 397 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 141, + 385, + 470, + 398 + ], + "score": 1.0, + "content": "optimize the optimal transport distance to align visual features and prompts by the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 89, + 396, + 470, + 409 + ], + "spans": [ + { + "bbox": [ + 89, + 398, + 100, + 409 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 141, + 396, + 470, + 409 + ], + "score": 1.0, + "content": "Sinkhorn algorithm, while in the outer loop, we learn the prompts by this distance", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 90, + 407, + 470, + 420 + ], + "spans": [ + { + "bbox": [ + 90, + 410, + 99, + 419 + ], + "score": 1.0, + "content": "13", + "type": "text" + }, + { + "bbox": [ + 141, + 407, + 470, + 420 + ], + "score": 1.0, + "content": "from the supervised data. Extensive experiments are conducted on the few-shot", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 89, + 419, + 467, + 431 + ], + "spans": [ + { + "bbox": [ + 89, + 421, + 100, + 430 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 141, + 419, + 467, + 431 + ], + "score": 1.0, + "content": "recognition task and the improvement demonstrates the superiority of our method.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 91, + 444, + 191, + 457 + ], + "lines": [ + { + "bbox": [ + 87, + 443, + 192, + 459 + ], + "spans": [ + { + "bbox": [ + 87, + 443, + 192, + 459 + ], + "score": 1.0, + "content": "15 1 Introduction", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 90, + 463, + 507, + 518 + ], + "lines": [ + { + "bbox": [ + 90, + 464, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 90, + 465, + 100, + 474 + ], + "score": 1.0, + "content": "16", + "type": "text" + }, + { + "bbox": [ + 106, + 464, + 506, + 475 + ], + "score": 1.0, + "content": "In the past few years, large-scale vision-language pre-trained (VLP) models, such as CLIP [39],", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 90, + 473, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 90, + 477, + 100, + 486 + ], + "score": 1.0, + "content": "17", + "type": "text" + }, + { + "bbox": [ + 105, + 473, + 507, + 488 + ], + "score": 1.0, + "content": "ALIGN [17], and BLIP [23] have achieved remarkable success in open-world visual concept learning.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 90, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 90, + 488, + 99, + 496 + ], + "score": 1.0, + "content": "18", + "type": "text" + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "These methods have brought new light but also pose a new question: how to efficiently adapt the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 89, + 498, + 100, + 507 + ], + "score": 1.0, + "content": "19", + "type": "text" + }, + { + "bbox": [ + 105, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "knowledge from pretraining to the downstream tasks since these models are typical of massive sizes", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 89, + 507, + 308, + 519 + ], + "spans": [ + { + "bbox": [ + 89, + 509, + 100, + 518 + ], + "score": 1.0, + "content": "20", + "type": "text" + }, + { + "bbox": [ + 105, + 507, + 308, + 519 + ], + "score": 1.0, + "content": "which are not feasible for normal users to re-train.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 90, + 524, + 240, + 643 + ], + "lines": [ + { + "bbox": [ + 88, + 522, + 240, + 534 + ], + "spans": [ + { + "bbox": [ + 88, + 524, + 100, + 534 + ], + "score": 1.0, + "content": "21", + "type": "text" + }, + { + "bbox": [ + 105, + 522, + 240, + 534 + ], + "score": 1.0, + "content": "One of the conventional", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 89, + 534, + 240, + 546 + ], + "spans": [ + { + "bbox": [ + 89, + 536, + 100, + 546 + ], + "score": 1.0, + "content": "22", + "type": "text" + }, + { + "bbox": [ + 104, + 534, + 240, + 546 + ], + "score": 1.0, + "content": "paradigms of utilizing pretrained", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 88, + 543, + 242, + 558 + ], + "spans": [ + { + "bbox": [ + 88, + 547, + 100, + 557 + ], + "score": 1.0, + "content": "23", + "type": "text" + }, + { + "bbox": [ + 104, + 543, + 242, + 558 + ], + "score": 1.0, + "content": "knowledge is “pre-training,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 88, + 556, + 240, + 567 + ], + "spans": [ + { + "bbox": [ + 88, + 558, + 100, + 567 + ], + "score": 1.0, + "content": "24", + "type": "text" + }, + { + "bbox": [ + 105, + 556, + 240, + 567 + ], + "score": 1.0, + "content": "fine-tuning”, which fixes the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 88, + 567, + 240, + 578 + ], + "spans": [ + { + "bbox": [ + 88, + 569, + 100, + 578 + ], + "score": 1.0, + "content": "25", + "type": "text" + }, + { + "bbox": [ + 106, + 567, + 240, + 578 + ], + "score": 1.0, + "content": "architecture of the pre-trained", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 88, + 578, + 240, + 589 + ], + "spans": [ + { + "bbox": [ + 88, + 579, + 100, + 589 + ], + "score": 1.0, + "content": "26", + "type": "text" + }, + { + "bbox": [ + 105, + 578, + 240, + 589 + ], + "score": 1.0, + "content": "neural network and tunes its", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 88, + 589, + 240, + 600 + ], + "spans": [ + { + "bbox": [ + 88, + 590, + 100, + 600 + ], + "score": 1.0, + "content": "27", + "type": "text" + }, + { + "bbox": [ + 105, + 589, + 240, + 600 + ], + "score": 1.0, + "content": "parameters using task-specific", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 88, + 599, + 241, + 612 + ], + "spans": [ + { + "bbox": [ + 88, + 601, + 100, + 611 + ], + "score": 1.0, + "content": "28", + "type": "text" + }, + { + "bbox": [ + 105, + 599, + 241, + 612 + ], + "score": 1.0, + "content": "objective functions. Beyond", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 88, + 610, + 242, + 623 + ], + "spans": [ + { + "bbox": [ + 88, + 612, + 100, + 622 + ], + "score": 1.0, + "content": "29", + "type": "text" + }, + { + "bbox": [ + 105, + 610, + 242, + 623 + ], + "score": 1.0, + "content": "fine-tuning the parameters,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 621, + 241, + 633 + ], + "spans": [ + { + "bbox": [ + 89, + 624, + 99, + 632 + ], + "score": 1.0, + "content": "30", + "type": "text" + }, + { + "bbox": [ + 104, + 621, + 241, + 633 + ], + "score": 1.0, + "content": "many recent methods [63, 64]", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 88, + 631, + 241, + 645 + ], + "spans": [ + { + "bbox": [ + 88, + 634, + 100, + 644 + ], + "score": 1.0, + "content": "31", + "type": "text" + }, + { + "bbox": [ + 104, + 631, + 241, + 645 + ], + "score": 1.0, + "content": "introduce the concept of prompt", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 248, + 525, + 503, + 608 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 248, + 525, + 503, + 608 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 248, + 525, + 503, + 608 + ], + "spans": [ + { + "bbox": [ + 248, + 525, + 503, + 608 + ], + "score": 0.97, + "type": "image", + "image_path": "c57e9c1381d6db8b3dade63c3241e4fb4e845de8801b03d88426df4af12cc486.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 248, + 525, + 503, + 552.6666666666666 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 248, + 552.6666666666666, + 503, + 580.3333333333333 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 248, + 580.3333333333333, + 503, + 607.9999999999999 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 246, + 614, + 505, + 636 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 246, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 246, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "Figure 1: The motivation that one category can be complementar-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 246, + 623, + 498, + 637 + ], + "spans": [ + { + "bbox": [ + 246, + 623, + 498, + 637 + ], + "score": 1.0, + "content": "ily described in different views (An example of “Brambling”).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "index": 40.25 + }, + { + "type": "text", + "bbox": [ + 94, + 644, + 506, + 687 + ], + "lines": [ + { + "bbox": [ + 93, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 93, + 646, + 100, + 654 + ], + "score": 1.0, + "content": "32", + "type": "text" + }, + { + "bbox": [ + 105, + 643, + 505, + 656 + ], + "score": 1.0, + "content": "learning from the field of NLP to the vision domain and achieve striking performance gain for the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 93, + 653, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 93, + 658, + 99, + 665 + ], + "score": 1.0, + "content": "33", + "type": "text" + }, + { + "bbox": [ + 104, + 653, + 505, + 668 + ], + "score": 1.0, + "content": "few-shot visual classification. They fix the model parameters and instead learn suitable prompts", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 93, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 93, + 668, + 99, + 675 + ], + "score": 1.0, + "content": "34", + "type": "text" + }, + { + "bbox": [ + 104, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "by turning a template sentence into a set of learnable vectors. Then, these prompts are learned by", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 93, + 676, + 464, + 689 + ], + "spans": [ + { + "bbox": [ + 93, + 680, + 99, + 687 + ], + "score": 1.0, + "content": "35", + "type": "text" + }, + { + "bbox": [ + 105, + 676, + 464, + 689 + ], + "score": 1.0, + "content": "minimizing the distance between the visual features and prompt-based language features.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + }, + { + "type": "text", + "bbox": [ + 89, + 693, + 506, + 714 + ], + "lines": [ + { + "bbox": [ + 89, + 692, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 89, + 694, + 100, + 703 + ], + "score": 1.0, + "content": "36", + "type": "text" + }, + { + "bbox": [ + 105, + 692, + 505, + 705 + ], + "score": 1.0, + "content": "Despite significant improvements over manual prompts, learning only a sentence is intuitively", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 89, + 703, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 89, + 705, + 101, + 715 + ], + "score": 1.0, + "content": "37", + "type": "text" + }, + { + "bbox": [ + 104, + 703, + 505, + 716 + ], + "score": 1.0, + "content": "insufficient to represent a class. One class can be described by many intrinsic characteristics and", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 101, + 731, + 499, + 742 + ], + "lines": [ + { + "bbox": [ + 106, + 730, + 500, + 743 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 500, + 743 + ], + "score": 1.0, + "content": "Submitted to 36th Conference on Neural Information Processing Systems (NeurIPS 2022). Do not distribute.", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 137, + 97, + 474, + 137 + ], + "lines": [ + { + "bbox": [ + 136, + 97, + 476, + 119 + ], + "spans": [ + { + "bbox": [ + 136, + 97, + 476, + 119 + ], + "score": 1.0, + "content": "Prompt Learning with Optimal Transport for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 212, + 116, + 399, + 139 + ], + "spans": [ + { + "bbox": [ + 212, + 116, + 399, + 139 + ], + "score": 1.0, + "content": "Vision-Language Models", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 259, + 179, + 354, + 223 + ], + "lines": [ + { + "bbox": [ + 258, + 178, + 356, + 191 + ], + "spans": [ + { + "bbox": [ + 258, + 178, + 356, + 191 + ], + "score": 1.0, + "content": "Anonymous Author(s)", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 283, + 189, + 328, + 202 + ], + "spans": [ + { + "bbox": [ + 283, + 189, + 328, + 202 + ], + "score": 1.0, + "content": "Affiliation", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 286, + 200, + 324, + 213 + ], + "spans": [ + { + "bbox": [ + 286, + 200, + 324, + 213 + ], + "score": 1.0, + "content": "Address", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 290, + 213, + 320, + 222 + ], + "spans": [ + { + "bbox": [ + 290, + 213, + 320, + 222 + ], + "score": 1.0, + "content": "email", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + } + ], + "index": 3.5, + "bbox_fs": [ + 258, + 178, + 356, + 222 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 252, + 328, + 265 + ], + "lines": [ + { + "bbox": [ + 281, + 251, + 331, + 267 + ], + "spans": [ + { + "bbox": [ + 281, + 251, + 331, + 267 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "index", + "bbox": [ + 91, + 276, + 469, + 430 + ], + "lines": [ + { + "bbox": [ + 93, + 276, + 470, + 289 + ], + "spans": [ + { + "bbox": [ + 93, + 280, + 99, + 287 + ], + "score": 1.0, + "content": "1", + "type": "text" + }, + { + "bbox": [ + 141, + 276, + 470, + 289 + ], + "score": 1.0, + "content": "With the increasing attention to large vision-language models such as CLIP, there", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 287, + 471, + 300 + ], + "spans": [ + { + "bbox": [ + 93, + 290, + 99, + 298 + ], + "score": 1.0, + "content": "2", + "type": "text" + }, + { + "bbox": [ + 141, + 287, + 471, + 300 + ], + "score": 1.0, + "content": "has been a significant amount of effort dedicated to building efficient prompts.", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 297, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 93, + 301, + 99, + 309 + ], + "score": 1.0, + "content": "3", + "type": "text" + }, + { + "bbox": [ + 141, + 297, + 470, + 312 + ], + "score": 1.0, + "content": "Unlike conventional methods of only learning one single prompt, we propose", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 309, + 470, + 322 + ], + "spans": [ + { + "bbox": [ + 92, + 312, + 99, + 321 + ], + "score": 1.0, + "content": "4", + "type": "text" + }, + { + "bbox": [ + 141, + 309, + 470, + 322 + ], + "score": 1.0, + "content": "to learn multiple comprehensive prompts to describe diverse characteristics of", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 320, + 470, + 333 + ], + "spans": [ + { + "bbox": [ + 93, + 323, + 99, + 331 + ], + "score": 1.0, + "content": "5", + "type": "text" + }, + { + "bbox": [ + 141, + 320, + 470, + 333 + ], + "score": 1.0, + "content": "categories such as intrinsic attributes or extrinsic contexts. However, directly", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 331, + 470, + 344 + ], + "spans": [ + { + "bbox": [ + 93, + 334, + 99, + 342 + ], + "score": 1.0, + "content": "6", + "type": "text" + }, + { + "bbox": [ + 141, + 331, + 470, + 344 + ], + "score": 1.0, + "content": "matching each prompt to the same visual feature is problematic, as it pushes the", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 91, + 342, + 469, + 355 + ], + "spans": [ + { + "bbox": [ + 91, + 344, + 99, + 353 + ], + "score": 1.0, + "content": "7", + "type": "text" + }, + { + "bbox": [ + 141, + 342, + 469, + 355 + ], + "score": 1.0, + "content": "prompts to converge to one point. To solve this problem, we propose to apply", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 354, + 470, + 366 + ], + "spans": [ + { + "bbox": [ + 92, + 356, + 99, + 364 + ], + "score": 1.0, + "content": "8", + "type": "text" + }, + { + "bbox": [ + 142, + 354, + 470, + 366 + ], + "score": 1.0, + "content": "optimal transport to match the vision and text modalities. Specifically, we first", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 363, + 470, + 376 + ], + "spans": [ + { + "bbox": [ + 92, + 366, + 100, + 376 + ], + "score": 1.0, + "content": "9", + "type": "text" + }, + { + "bbox": [ + 141, + 363, + 470, + 376 + ], + "score": 1.0, + "content": "model images and the categories with visual and textual feature sets. Then, we", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 374, + 470, + 388 + ], + "spans": [ + { + "bbox": [ + 90, + 376, + 100, + 387 + ], + "score": 1.0, + "content": "10", + "type": "text" + }, + { + "bbox": [ + 141, + 374, + 470, + 388 + ], + "score": 1.0, + "content": "apply a two-stage optimization strategy to learn the prompts. In the inner loop, we", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 385, + 470, + 398 + ], + "spans": [ + { + "bbox": [ + 90, + 388, + 99, + 397 + ], + "score": 1.0, + "content": "11", + "type": "text" + }, + { + "bbox": [ + 141, + 385, + 470, + 398 + ], + "score": 1.0, + "content": "optimize the optimal transport distance to align visual features and prompts by the", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 396, + 470, + 409 + ], + "spans": [ + { + "bbox": [ + 89, + 398, + 100, + 409 + ], + "score": 1.0, + "content": "12", + "type": "text" + }, + { + "bbox": [ + 141, + 396, + 470, + 409 + ], + "score": 1.0, + "content": "Sinkhorn algorithm, while in the outer loop, we learn the prompts by this distance", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 407, + 470, + 420 + ], + "spans": [ + { + "bbox": [ + 90, + 410, + 99, + 419 + ], + "score": 1.0, + "content": "13", + "type": "text" + }, + { + "bbox": [ + 141, + 407, + 470, + 420 + ], + "score": 1.0, + "content": "from the supervised data. Extensive experiments are conducted on the few-shot", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 419, + 467, + 431 + ], + "spans": [ + { + "bbox": [ + 89, + 421, + 100, + 430 + ], + "score": 1.0, + "content": "14", + "type": "text" + }, + { + "bbox": [ + 141, + 419, + 467, + 431 + ], + "score": 1.0, + "content": "recognition task and the improvement demonstrates the superiority of our method.", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + } + ], + "index": 13.5, + "bbox_fs": [ + 89, + 276, + 471, + 431 + ] + }, + { + "type": "title", + "bbox": [ + 91, + 444, + 191, + 457 + ], + "lines": [ + { + "bbox": [ + 87, + 443, + 192, + 459 + ], + "spans": [ + { + "bbox": [ + 87, + 443, + 192, + 459 + ], + "score": 1.0, + "content": "15 1 Introduction", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "index", + "bbox": [ + 90, + 463, + 507, + 518 + ], + "lines": [ + { + "bbox": [ + 90, + 464, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 90, + 465, + 100, + 474 + ], + "score": 1.0, + "content": "16", + "type": "text" + }, + { + "bbox": [ + 106, + 464, + 506, + 475 + ], + "score": 1.0, + "content": "In the past few years, large-scale vision-language pre-trained (VLP) models, such as CLIP [39],", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 473, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 90, + 477, + 100, + 486 + ], + "score": 1.0, + "content": "17", + "type": "text" + }, + { + "bbox": [ + 105, + 473, + 507, + 488 + ], + "score": 1.0, + "content": "ALIGN [17], and BLIP [23] have achieved remarkable success in open-world visual concept learning.", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 484, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 90, + 488, + 99, + 496 + ], + "score": 1.0, + "content": "18", + "type": "text" + }, + { + "bbox": [ + 105, + 484, + 506, + 498 + ], + "score": 1.0, + "content": "These methods have brought new light but also pose a new question: how to efficiently adapt the", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 89, + 498, + 100, + 507 + ], + "score": 1.0, + "content": "19", + "type": "text" + }, + { + "bbox": [ + 105, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "knowledge from pretraining to the downstream tasks since these models are typical of massive sizes", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 507, + 308, + 519 + ], + "spans": [ + { + "bbox": [ + 89, + 509, + 100, + 518 + ], + "score": 1.0, + "content": "20", + "type": "text" + }, + { + "bbox": [ + 105, + 507, + 308, + 519 + ], + "score": 1.0, + "content": "which are not feasible for normal users to re-train.", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 522, + 240, + 534 + ], + "spans": [ + { + "bbox": [ + 88, + 524, + 100, + 534 + ], + "score": 1.0, + "content": "21", + "type": "text" + }, + { + "bbox": [ + 105, + 522, + 240, + 534 + ], + "score": 1.0, + "content": "One of the conventional", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 534, + 240, + 546 + ], + "spans": [ + { + "bbox": [ + 89, + 536, + 100, + 546 + ], + "score": 1.0, + "content": "22", + "type": "text" + }, + { + "bbox": [ + 104, + 534, + 240, + 546 + ], + "score": 1.0, + "content": "paradigms of utilizing pretrained", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 543, + 242, + 558 + ], + "spans": [ + { + "bbox": [ + 88, + 547, + 100, + 557 + ], + "score": 1.0, + "content": "23", + "type": "text" + }, + { + "bbox": [ + 104, + 543, + 242, + 558 + ], + "score": 1.0, + "content": "knowledge is “pre-training,", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 556, + 240, + 567 + ], + "spans": [ + { + "bbox": [ + 88, + 558, + 100, + 567 + ], + "score": 1.0, + "content": "24", + "type": "text" + }, + { + "bbox": [ + 105, + 556, + 240, + 567 + ], + "score": 1.0, + "content": "fine-tuning”, which fixes the", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 567, + 240, + 578 + ], + "spans": [ + { + "bbox": [ + 88, + 569, + 100, + 578 + ], + "score": 1.0, + "content": "25", + "type": "text" + }, + { + "bbox": [ + 106, + 567, + 240, + 578 + ], + "score": 1.0, + "content": "architecture of the pre-trained", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 578, + 240, + 589 + ], + "spans": [ + { + "bbox": [ + 88, + 579, + 100, + 589 + ], + "score": 1.0, + "content": "26", + "type": "text" + }, + { + "bbox": [ + 105, + 578, + 240, + 589 + ], + "score": 1.0, + "content": "neural network and tunes its", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 589, + 240, + 600 + ], + "spans": [ + { + "bbox": [ + 88, + 590, + 100, + 600 + ], + "score": 1.0, + "content": "27", + "type": "text" + }, + { + "bbox": [ + 105, + 589, + 240, + 600 + ], + "score": 1.0, + "content": "parameters using task-specific", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 599, + 241, + 612 + ], + "spans": [ + { + "bbox": [ + 88, + 601, + 100, + 611 + ], + "score": 1.0, + "content": "28", + "type": "text" + }, + { + "bbox": [ + 105, + 599, + 241, + 612 + ], + "score": 1.0, + "content": "objective functions. Beyond", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 610, + 242, + 623 + ], + "spans": [ + { + "bbox": [ + 88, + 612, + 100, + 622 + ], + "score": 1.0, + "content": "29", + "type": "text" + }, + { + "bbox": [ + 105, + 610, + 242, + 623 + ], + "score": 1.0, + "content": "fine-tuning the parameters,", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 621, + 241, + 633 + ], + "spans": [ + { + "bbox": [ + 89, + 624, + 99, + 632 + ], + "score": 1.0, + "content": "30", + "type": "text" + }, + { + "bbox": [ + 104, + 621, + 241, + 633 + ], + "score": 1.0, + "content": "many recent methods [63, 64]", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 631, + 241, + 645 + ], + "spans": [ + { + "bbox": [ + 88, + 634, + 100, + 644 + ], + "score": 1.0, + "content": "31", + "type": "text" + }, + { + "bbox": [ + 104, + 631, + 241, + 645 + ], + "score": 1.0, + "content": "introduce the concept of prompt", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 93, + 646, + 100, + 654 + ], + "score": 1.0, + "content": "32", + "type": "text" + }, + { + "bbox": [ + 105, + 643, + 505, + 656 + ], + "score": 1.0, + "content": "learning from the field of NLP to the vision domain and achieve striking performance gain for the", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 653, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 93, + 658, + 99, + 665 + ], + "score": 1.0, + "content": "33", + "type": "text" + }, + { + "bbox": [ + 104, + 653, + 505, + 668 + ], + "score": 1.0, + "content": "few-shot visual classification. They fix the model parameters and instead learn suitable prompts", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 93, + 668, + 99, + 675 + ], + "score": 1.0, + "content": "34", + "type": "text" + }, + { + "bbox": [ + 104, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "by turning a template sentence into a set of learnable vectors. Then, these prompts are learned by", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 93, + 676, + 464, + 689 + ], + "spans": [ + { + "bbox": [ + 93, + 680, + 99, + 687 + ], + "score": 1.0, + "content": "35", + "type": "text" + }, + { + "bbox": [ + 105, + 676, + 464, + 689 + ], + "score": 1.0, + "content": "minimizing the distance between the visual features and prompt-based language features.", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 692, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 89, + 694, + 100, + 703 + ], + "score": 1.0, + "content": "36", + "type": "text" + }, + { + "bbox": [ + 105, + 692, + 505, + 705 + ], + "score": 1.0, + "content": "Despite significant improvements over manual prompts, learning only a sentence is intuitively", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 703, + 505, + 716 + ], + "spans": [ + { + "bbox": [ + 89, + 705, + 101, + 715 + ], + "score": 1.0, + "content": "37", + "type": "text" + }, + { + "bbox": [ + 104, + 703, + 505, + 716 + ], + "score": 1.0, + "content": "insufficient to represent a class. One class can be described by many intrinsic characteristics and", + "type": "text" + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 73, + 504, + 84 + ], + "spans": [ + { + "bbox": [ + 89, + 75, + 100, + 84 + ], + "score": 1.0, + "content": "38", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 73, + 504, + 83 + ], + "score": 1.0, + "content": "even extrinsic context relations. Thus, for one object, we may have multiple prompt candidates", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 89, + 86, + 100, + 95 + ], + "score": 1.0, + "content": "39", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "which focus on different attributes. As shown in Figure 1, we can describe the class “Brambling” in", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 89, + 96, + 100, + 106 + ], + "score": 1.0, + "content": "40", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "different views: such as the color of the wing, the color of the crown and eyes, the shape and color of", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 89, + 108, + 100, + 117 + ], + "score": 1.0, + "content": "41", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "the tail, and even the living environment information. It motivates us to learn multiple prompts to", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 117, + 382, + 128 + ], + "spans": [ + { + "bbox": [ + 89, + 119, + 100, + 128 + ], + "score": 1.0, + "content": "42", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 117, + 382, + 128 + ], + "score": 1.0, + "content": "comprehensively represent the class and thus facilitate classification.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 131, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 89, + 134, + 100, + 144 + ], + "score": 1.0, + "content": "43", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 131, + 506, + 146 + ], + "score": 1.0, + "content": "The most natural solution is to directly learn multiple prompts by respectively matching each prompt", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 89, + 146, + 100, + 155 + ], + "score": 1.0, + "content": "44", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "with the visual features. However, it is the same as matching the mean of prompt features and the", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 154, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 89, + 156, + 100, + 166 + ], + "score": 1.0, + "content": "45", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 154, + 506, + 168 + ], + "score": 1.0, + "content": "visual features. This solution is problematic since all prompts are encouraged to be closer to one single", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 88, + 167, + 100, + 177 + ], + "score": 1.0, + "content": "46", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "point and thus tend to learn the same characteristics. It contradicts our purpose to learn comprehensive", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 89, + 178, + 100, + 189 + ], + "score": 1.0, + "content": "47", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "prompts. To solve this problem, we tested adding some constraints to push away the prompt from", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 186, + 507, + 201 + ], + "spans": [ + { + "bbox": [ + 89, + 190, + 100, + 199 + ], + "score": 1.0, + "content": "48", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 186, + 507, + 201 + ], + "score": 1.0, + "content": "each other, but found that this solution still fails to learn representative and comprehensive prompts.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 197, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 89, + 200, + 100, + 210 + ], + "score": 1.0, + "content": "49", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 197, + 506, + 211 + ], + "score": 1.0, + "content": "This solution treats the visual representation as one single point, and such a unified view of visual", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 89, + 212, + 100, + 221 + ], + "score": 1.0, + "content": "50", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "features ignores the fact that different prompts may only focus on one or a subset of characteristics.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 225, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 89, + 227, + 99, + 237 + ], + "score": 1.0, + "content": "51", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "score": 1.0, + "content": "To address this problem, in this paper, we propose Prompt Learning with Optimal Transport (PLOT),", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 236, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 89, + 239, + 99, + 248 + ], + "score": 1.0, + "content": "52", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 236, + 506, + 248 + ], + "score": 1.0, + "content": "which applies optimal transport (OT) to align the local visual features and multiple textual prompts.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 247, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 89, + 249, + 100, + 259 + ], + "score": 1.0, + "content": "53", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 247, + 506, + 260 + ], + "score": 1.0, + "content": "Optimal transport can calculate the distance between two distributions under the form of multiple", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 259, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 89, + 261, + 100, + 270 + ], + "score": 1.0, + "content": "54", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 259, + 505, + 270 + ], + "score": 1.0, + "content": "sampling. In our prompt learning framework, we formulate local visual features and multiple prompts", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 269, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 89, + 271, + 99, + 281 + ], + "score": 1.0, + "content": "55", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 269, + 506, + 281 + ], + "score": 1.0, + "content": "as the samplings of two discrete distributions and use OT to encourage fine-grained cross-modal", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 280, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 88, + 282, + 100, + 291 + ], + "score": 1.0, + "content": "56", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 280, + 506, + 292 + ], + "score": 1.0, + "content": "matching. Specifically, to obtain the local visual features with different semantic clues, we extract all", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 290, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 88, + 293, + 99, + 302 + ], + "score": 1.0, + "content": "57", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 290, + 506, + 304 + ], + "score": 1.0, + "content": "feature maps as the visual representation instead of the single global representation. Fortunately, we", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 302, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 88, + 304, + 100, + 313 + ], + "score": 1.0, + "content": "58", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 302, + 506, + 314 + ], + "score": 1.0, + "content": "can easily obtain the visual feature maps from the visual encoder of CLIP by using all outputs of the", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 89, + 315, + 99, + 324 + ], + "score": 1.0, + "content": "59", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "multi-head self-attention layer [42]. Then the problem comes down to how to calculate the distance", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 324, + 209, + 335 + ], + "spans": [ + { + "bbox": [ + 89, + 326, + 100, + 335 + ], + "score": 1.0, + "content": "60", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 324, + 209, + 335 + ], + "score": 1.0, + "content": "between two feature sets.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 89, + 342, + 99, + 351 + ], + "score": 1.0, + "content": "61", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "We solve this problem by introducing the optimal transport theory [51] and formulate the feature sets", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 350, + 507, + 364 + ], + "spans": [ + { + "bbox": [ + 89, + 353, + 100, + 362 + ], + "score": 1.0, + "content": "62", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 350, + 507, + 364 + ], + "score": 1.0, + "content": "as a discrete probability distribution where each feature has an equal probability value. Furthermore,", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 362, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 89, + 364, + 100, + 373 + ], + "score": 1.0, + "content": "63", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "to reduce the computational cost and avoid the extra model parameters, we learn the prompts with", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 372, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 88, + 374, + 101, + 385 + ], + "score": 1.0, + "content": "64", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 372, + 506, + 385 + ], + "score": 1.0, + "content": "a two-stage optimization strategy. At the first stage in the inner loop, we fix both visual and text", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 89, + 386, + 100, + 395 + ], + "score": 1.0, + "content": "65", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "features and optimize the optimal transport problem by a fast Sinkhorn distances algorithm [6]. Then,", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 88, + 396, + 101, + 407 + ], + "score": 1.0, + "content": "66", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "in the outer loop, we fix all parameters of optimal transport and back-propagate the gradient to learn", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 89, + 407, + 100, + 417 + ], + "score": 1.0, + "content": "67", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "the prompts with different characteristics. Compared with conventional distance (such as Euclidean", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 416, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 88, + 418, + 101, + 428 + ], + "score": 1.0, + "content": "68", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 416, + 506, + 429 + ], + "score": 1.0, + "content": "distance of mean features), optimal transport can align different visual features for each local prompt,", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 427, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 89, + 429, + 100, + 439 + ], + "score": 1.0, + "content": "69", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 427, + 506, + 440 + ], + "score": 1.0, + "content": "which is more robust to the visual misalignment and tolerates well feature shift [44]. It is because OT", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 439, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 89, + 440, + 100, + 450 + ], + "score": 1.0, + "content": "70", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 439, + 505, + 450 + ], + "score": 1.0, + "content": "learns an adaptive transport plan to align features, which achieves fine-grained matching across two", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 89, + 451, + 99, + 460 + ], + "score": 1.0, + "content": "71", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "modalities. We conduct experiments on 11 datasets following the standard setting of CLIP [39] and", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 89, + 462, + 100, + 471 + ], + "score": 1.0, + "content": "72", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "CoOp [63] to evaluate our method. These experiments span the visual classification of generic objects,", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 89, + 473, + 100, + 482 + ], + "score": 1.0, + "content": "73", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "scenes, actions, fine-grained categories, and so on. The significant result improvement demonstrates", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 481, + 409, + 495 + ], + "spans": [ + { + "bbox": [ + 89, + 484, + 100, + 493 + ], + "score": 1.0, + "content": "74", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 481, + 409, + 495 + ], + "score": 1.0, + "content": "that PLOT can effectively learn representative and comprehensive prompts.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_start_line": true + } + ], + "index": 24, + "bbox_fs": [ + 89, + 464, + 507, + 519 + ] + }, + { + "type": "index", + "bbox": [ + 90, + 524, + 240, + 643 + ], + "lines": [], + "index": 32, + "bbox_fs": [ + 88, + 522, + 242, + 645 + ], + "lines_deleted": true + }, + { + "type": "image", + "bbox": [ + 248, + 525, + 503, + 608 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 248, + 525, + 503, + 608 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 248, + 525, + 503, + 608 + ], + "spans": [ + { + "bbox": [ + 248, + 525, + 503, + 608 + ], + "score": 0.97, + "type": "image", + "image_path": "c57e9c1381d6db8b3dade63c3241e4fb4e845de8801b03d88426df4af12cc486.jpg" + } + ] + } + ], + "index": 39, + "virtual_lines": [ + { + "bbox": [ + 248, + 525, + 503, + 552.6666666666666 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 248, + 552.6666666666666, + 503, + 580.3333333333333 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 248, + 580.3333333333333, + 503, + 607.9999999999999 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 246, + 614, + 505, + 636 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 246, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 246, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "Figure 1: The motivation that one category can be complementar-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 246, + 623, + 498, + 637 + ], + "spans": [ + { + "bbox": [ + 246, + 623, + 498, + 637 + ], + "score": 1.0, + "content": "ily described in different views (An example of “Brambling”).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + } + ], + "index": 40.25 + }, + { + "type": "index", + "bbox": [ + 94, + 644, + 506, + 687 + ], + "lines": [], + "index": 44.5, + "bbox_fs": [ + 93, + 643, + 505, + 689 + ], + "lines_deleted": true + }, + { + "type": "index", + "bbox": [ + 89, + 693, + 506, + 714 + ], + "lines": [], + "index": 47.5, + "bbox_fs": [ + 89, + 692, + 505, + 716 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 90, + 73, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 89, + 73, + 504, + 84 + ], + "spans": [ + { + "bbox": [ + 89, + 75, + 100, + 84 + ], + "score": 1.0, + "content": "38", + "type": "text" + }, + { + "bbox": [ + 106, + 73, + 504, + 83 + ], + "score": 1.0, + "content": "even extrinsic context relations. Thus, for one object, we may have multiple prompt candidates", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 89, + 86, + 100, + 95 + ], + "score": 1.0, + "content": "39", + "type": "text" + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "which focus on different attributes. As shown in Figure 1, we can describe the class “Brambling” in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 89, + 96, + 100, + 106 + ], + "score": 1.0, + "content": "40", + "type": "text" + }, + { + "bbox": [ + 106, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "different views: such as the color of the wing, the color of the crown and eyes, the shape and color of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 89, + 108, + 100, + 117 + ], + "score": 1.0, + "content": "41", + "type": "text" + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "the tail, and even the living environment information. It motivates us to learn multiple prompts to", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 117, + 382, + 128 + ], + "spans": [ + { + "bbox": [ + 89, + 119, + 100, + 128 + ], + "score": 1.0, + "content": "42", + "type": "text" + }, + { + "bbox": [ + 105, + 117, + 382, + 128 + ], + "score": 1.0, + "content": "comprehensively represent the class and thus facilitate classification.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 90, + 132, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 89, + 131, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 89, + 134, + 100, + 144 + ], + "score": 1.0, + "content": "43", + "type": "text" + }, + { + "bbox": [ + 104, + 131, + 506, + 146 + ], + "score": 1.0, + "content": "The most natural solution is to directly learn multiple prompts by respectively matching each prompt", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 89, + 146, + 100, + 155 + ], + "score": 1.0, + "content": "44", + "type": "text" + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "with the visual features. However, it is the same as matching the mean of prompt features and the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 154, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 89, + 156, + 100, + 166 + ], + "score": 1.0, + "content": "45", + "type": "text" + }, + { + "bbox": [ + 104, + 154, + 506, + 168 + ], + "score": 1.0, + "content": "visual features. This solution is problematic since all prompts are encouraged to be closer to one single", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 88, + 167, + 100, + 177 + ], + "score": 1.0, + "content": "46", + "type": "text" + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "point and thus tend to learn the same characteristics. It contradicts our purpose to learn comprehensive", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 89, + 178, + 100, + 189 + ], + "score": 1.0, + "content": "47", + "type": "text" + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "prompts. To solve this problem, we tested adding some constraints to push away the prompt from", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 186, + 507, + 201 + ], + "spans": [ + { + "bbox": [ + 89, + 190, + 100, + 199 + ], + "score": 1.0, + "content": "48", + "type": "text" + }, + { + "bbox": [ + 104, + 186, + 507, + 201 + ], + "score": 1.0, + "content": "each other, but found that this solution still fails to learn representative and comprehensive prompts.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 197, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 89, + 200, + 100, + 210 + ], + "score": 1.0, + "content": "49", + "type": "text" + }, + { + "bbox": [ + 105, + 197, + 506, + 211 + ], + "score": 1.0, + "content": "This solution treats the visual representation as one single point, and such a unified view of visual", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 89, + 212, + 100, + 221 + ], + "score": 1.0, + "content": "50", + "type": "text" + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "features ignores the fact that different prompts may only focus on one or a subset of characteristics.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 90, + 226, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 89, + 225, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 89, + 227, + 99, + 237 + ], + "score": 1.0, + "content": "51", + "type": "text" + }, + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "score": 1.0, + "content": "To address this problem, in this paper, we propose Prompt Learning with Optimal Transport (PLOT),", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 89, + 236, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 89, + 239, + 99, + 248 + ], + "score": 1.0, + "content": "52", + "type": "text" + }, + { + "bbox": [ + 105, + 236, + 506, + 248 + ], + "score": 1.0, + "content": "which applies optimal transport (OT) to align the local visual features and multiple textual prompts.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 89, + 247, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 89, + 249, + 100, + 259 + ], + "score": 1.0, + "content": "53", + "type": "text" + }, + { + "bbox": [ + 105, + 247, + 506, + 260 + ], + "score": 1.0, + "content": "Optimal transport can calculate the distance between two distributions under the form of multiple", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 89, + 259, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 89, + 261, + 100, + 270 + ], + "score": 1.0, + "content": "54", + "type": "text" + }, + { + "bbox": [ + 106, + 259, + 505, + 270 + ], + "score": 1.0, + "content": "sampling. In our prompt learning framework, we formulate local visual features and multiple prompts", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 89, + 269, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 89, + 271, + 99, + 281 + ], + "score": 1.0, + "content": "55", + "type": "text" + }, + { + "bbox": [ + 104, + 269, + 506, + 281 + ], + "score": 1.0, + "content": "as the samplings of two discrete distributions and use OT to encourage fine-grained cross-modal", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 88, + 280, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 88, + 282, + 100, + 291 + ], + "score": 1.0, + "content": "56", + "type": "text" + }, + { + "bbox": [ + 105, + 280, + 506, + 292 + ], + "score": 1.0, + "content": "matching. Specifically, to obtain the local visual features with different semantic clues, we extract all", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 88, + 290, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 88, + 293, + 99, + 302 + ], + "score": 1.0, + "content": "57", + "type": "text" + }, + { + "bbox": [ + 105, + 290, + 506, + 304 + ], + "score": 1.0, + "content": "feature maps as the visual representation instead of the single global representation. Fortunately, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 302, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 88, + 304, + 100, + 313 + ], + "score": 1.0, + "content": "58", + "type": "text" + }, + { + "bbox": [ + 105, + 302, + 506, + 314 + ], + "score": 1.0, + "content": "can easily obtain the visual feature maps from the visual encoder of CLIP by using all outputs of the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 89, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 89, + 315, + 99, + 324 + ], + "score": 1.0, + "content": "59", + "type": "text" + }, + { + "bbox": [ + 105, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "multi-head self-attention layer [42]. Then the problem comes down to how to calculate the distance", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 89, + 324, + 209, + 335 + ], + "spans": [ + { + "bbox": [ + 89, + 326, + 100, + 335 + ], + "score": 1.0, + "content": "60", + "type": "text" + }, + { + "bbox": [ + 105, + 324, + 209, + 335 + ], + "score": 1.0, + "content": "between two feature sets.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 90, + 340, + 505, + 493 + ], + "lines": [ + { + "bbox": [ + 89, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 89, + 342, + 99, + 351 + ], + "score": 1.0, + "content": "61", + "type": "text" + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "We solve this problem by introducing the optimal transport theory [51] and formulate the feature sets", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 89, + 350, + 507, + 364 + ], + "spans": [ + { + "bbox": [ + 89, + 353, + 100, + 362 + ], + "score": 1.0, + "content": "62", + "type": "text" + }, + { + "bbox": [ + 104, + 350, + 507, + 364 + ], + "score": 1.0, + "content": "as a discrete probability distribution where each feature has an equal probability value. Furthermore,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 89, + 362, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 89, + 364, + 100, + 373 + ], + "score": 1.0, + "content": "63", + "type": "text" + }, + { + "bbox": [ + 105, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "to reduce the computational cost and avoid the extra model parameters, we learn the prompts with", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 88, + 372, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 88, + 374, + 101, + 385 + ], + "score": 1.0, + "content": "64", + "type": "text" + }, + { + "bbox": [ + 104, + 372, + 506, + 385 + ], + "score": 1.0, + "content": "a two-stage optimization strategy. At the first stage in the inner loop, we fix both visual and text", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 89, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 89, + 386, + 100, + 395 + ], + "score": 1.0, + "content": "65", + "type": "text" + }, + { + "bbox": [ + 105, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "features and optimize the optimal transport problem by a fast Sinkhorn distances algorithm [6]. Then,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 88, + 395, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 88, + 396, + 101, + 407 + ], + "score": 1.0, + "content": "66", + "type": "text" + }, + { + "bbox": [ + 105, + 395, + 505, + 407 + ], + "score": 1.0, + "content": "in the outer loop, we fix all parameters of optimal transport and back-propagate the gradient to learn", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 89, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 89, + 407, + 100, + 417 + ], + "score": 1.0, + "content": "67", + "type": "text" + }, + { + "bbox": [ + 105, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "the prompts with different characteristics. Compared with conventional distance (such as Euclidean", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 88, + 416, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 88, + 418, + 101, + 428 + ], + "score": 1.0, + "content": "68", + "type": "text" + }, + { + "bbox": [ + 106, + 416, + 506, + 429 + ], + "score": 1.0, + "content": "distance of mean features), optimal transport can align different visual features for each local prompt,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 89, + 427, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 89, + 429, + 100, + 439 + ], + "score": 1.0, + "content": "69", + "type": "text" + }, + { + "bbox": [ + 106, + 427, + 506, + 440 + ], + "score": 1.0, + "content": "which is more robust to the visual misalignment and tolerates well feature shift [44]. It is because OT", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 89, + 439, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 89, + 440, + 100, + 450 + ], + "score": 1.0, + "content": "70", + "type": "text" + }, + { + "bbox": [ + 106, + 439, + 505, + 450 + ], + "score": 1.0, + "content": "learns an adaptive transport plan to align features, which achieves fine-grained matching across two", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 89, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 89, + 451, + 99, + 460 + ], + "score": 1.0, + "content": "71", + "type": "text" + }, + { + "bbox": [ + 105, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "modalities. We conduct experiments on 11 datasets following the standard setting of CLIP [39] and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 89, + 459, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 89, + 462, + 100, + 471 + ], + "score": 1.0, + "content": "72", + "type": "text" + }, + { + "bbox": [ + 105, + 459, + 506, + 473 + ], + "score": 1.0, + "content": "CoOp [63] to evaluate our method. These experiments span the visual classification of generic objects,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 89, + 471, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 89, + 473, + 100, + 482 + ], + "score": 1.0, + "content": "73", + "type": "text" + }, + { + "bbox": [ + 105, + 471, + 505, + 483 + ], + "score": 1.0, + "content": "scenes, actions, fine-grained categories, and so on. The significant result improvement demonstrates", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 89, + 481, + 409, + 495 + ], + "spans": [ + { + "bbox": [ + 89, + 484, + 100, + 493 + ], + "score": 1.0, + "content": "74", + "type": "text" + }, + { + "bbox": [ + 104, + 481, + 409, + 495 + ], + "score": 1.0, + "content": "that PLOT can effectively learn representative and comprehensive prompts.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 93, + 508, + 196, + 521 + ], + "lines": [ + { + "bbox": [ + 89, + 507, + 198, + 523 + ], + "spans": [ + { + "bbox": [ + 89, + 507, + 198, + 523 + ], + "score": 1.0, + "content": "75 2 Related Work", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 105, + 531, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "score": 1.0, + "content": "Optimal Transport The Optimal Transport [30] is initially introduced to solve the problem of how", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 542, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 553 + ], + "score": 1.0, + "content": "to reduce the cost when moving several items simultaneously. Recently, OT theory has drawn wide", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 552, + 504, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 504, + 565 + ], + "score": 1.0, + "content": "attention in the machine learning and computer vision community by comparing distributions readily", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "available to them under the form of feature sets [37]. Due to the brilliant property of distribution", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "matching, OT has been applied in many theoretic and application tasks including generative models [1,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 584, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 506, + 598 + ], + "score": 1.0, + "content": "45, 60], structural matching [4, 57, 61, 56] (e.g. sequence matching [4] and graph matching [56]),", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "and other distribution-based tasks (such as clustering [22], distribution estimation [2], and causal", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "discovery [50]). In this paper, we use OT to align the features of vision and language modalities", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 618, + 428, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 428, + 631 + ], + "score": 1.0, + "content": "which represents the data structure by learning an adaptive transport plan [44].", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 90, + 634, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 89, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 89, + 637, + 100, + 646 + ], + "score": 1.0, + "content": "85", + "type": "text" + }, + { + "bbox": [ + 106, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "Vision-Language Pre-trained Models Vision-Language Pre-trained (VLP) models aim to explore", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 89, + 645, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 89, + 648, + 100, + 657 + ], + "score": 1.0, + "content": "86", + "type": "text" + }, + { + "bbox": [ + 106, + 645, + 506, + 659 + ], + "score": 1.0, + "content": "the semantic correspondence between the vision and language modalities through large-scale pre-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 89, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 89, + 658, + 100, + 668 + ], + "score": 1.0, + "content": "87", + "type": "text" + }, + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "training. Recently, VLP models have achieved an exciting performance improvement in the zero-shot", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 89, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 89, + 669, + 100, + 679 + ], + "score": 1.0, + "content": "88", + "type": "text" + }, + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "and few-shot visual recognition [39, 10, 63, 64, 59], which shows the great potential to promote", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 88, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 88, + 680, + 101, + 691 + ], + "score": 1.0, + "content": "89", + "type": "text" + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "open-world visual understanding with the help of language. One key part of learning VLP models is", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 89, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 89, + 691, + 100, + 700 + ], + "score": 1.0, + "content": "90", + "type": "text" + }, + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "the self-supervised learning objective on two modalities. The popular VLP objectives can be divided", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 89, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 89, + 702, + 100, + 711 + ], + "score": 1.0, + "content": "91", + "type": "text" + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "into reconstruction [25, 15, 8, 20], contrastive matching [39, 17, 16], or the combination of both", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 89, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 89, + 713, + 100, + 722 + ], + "score": 1.0, + "content": "92", + "type": "text" + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "two [24, 54, 19]. Besides, recent progress in the field of VLP also benefits a lot from large-scale", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 50.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "index", + "bbox": [ + 90, + 73, + 505, + 127 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 89, + 73, + 506, + 128 + ], + "lines_deleted": true + }, + { + "type": "index", + "bbox": [ + 90, + 132, + 505, + 221 + ], + "lines": [], + "index": 8.5, + "bbox_fs": [ + 88, + 131, + 507, + 222 + ], + "lines_deleted": true + }, + { + "type": "index", + "bbox": [ + 90, + 226, + 505, + 335 + ], + "lines": [], + "index": 17.5, + "bbox_fs": [ + 88, + 225, + 506, + 335 + ], + "lines_deleted": true + }, + { + "type": "index", + "bbox": [ + 90, + 340, + 505, + 493 + ], + "lines": [], + "index": 29.5, + "bbox_fs": [ + 88, + 339, + 507, + 495 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 93, + 508, + 196, + 521 + ], + "lines": [ + { + "bbox": [ + 89, + 507, + 198, + 523 + ], + "spans": [ + { + "bbox": [ + 89, + 507, + 198, + 523 + ], + "score": 1.0, + "content": "75 2 Related Work", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 105, + 531, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 542 + ], + "score": 1.0, + "content": "Optimal Transport The Optimal Transport [30] is initially introduced to solve the problem of how", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 542, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 553 + ], + "score": 1.0, + "content": "to reduce the cost when moving several items simultaneously. Recently, OT theory has drawn wide", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 552, + 504, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 504, + 565 + ], + "score": 1.0, + "content": "attention in the machine learning and computer vision community by comparing distributions readily", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "available to them under the form of feature sets [37]. Due to the brilliant property of distribution", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "matching, OT has been applied in many theoretic and application tasks including generative models [1,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 584, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 506, + 598 + ], + "score": 1.0, + "content": "45, 60], structural matching [4, 57, 61, 56] (e.g. sequence matching [4] and graph matching [56]),", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "and other distribution-based tasks (such as clustering [22], distribution estimation [2], and causal", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 620 + ], + "score": 1.0, + "content": "discovery [50]). In this paper, we use OT to align the features of vision and language modalities", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 618, + 428, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 428, + 631 + ], + "score": 1.0, + "content": "which represents the data structure by learning an adaptive transport plan [44].", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 531, + 506, + 631 + ] + }, + { + "type": "index", + "bbox": [ + 90, + 634, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 89, + 635, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 89, + 637, + 100, + 646 + ], + "score": 1.0, + "content": "85", + "type": "text" + }, + { + "bbox": [ + 106, + 635, + 505, + 647 + ], + "score": 1.0, + "content": "Vision-Language Pre-trained Models Vision-Language Pre-trained (VLP) models aim to explore", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 645, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 89, + 648, + 100, + 657 + ], + "score": 1.0, + "content": "86", + "type": "text" + }, + { + "bbox": [ + 106, + 645, + 506, + 659 + ], + "score": 1.0, + "content": "the semantic correspondence between the vision and language modalities through large-scale pre-", + "type": "text" + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 89, + 658, + 100, + 668 + ], + "score": 1.0, + "content": "87", + "type": "text" + }, + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "training. Recently, VLP models have achieved an exciting performance improvement in the zero-shot", + "type": "text" + } + ], + "index": 49, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 89, + 669, + 100, + 679 + ], + "score": 1.0, + "content": "88", + "type": "text" + }, + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "and few-shot visual recognition [39, 10, 63, 64, 59], which shows the great potential to promote", + "type": "text" + } + ], + "index": 50, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 88, + 680, + 101, + 691 + ], + "score": 1.0, + "content": "89", + "type": "text" + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "open-world visual understanding with the help of language. One key part of learning VLP models is", + "type": "text" + } + ], + "index": 51, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 89, + 691, + 100, + 700 + ], + "score": 1.0, + "content": "90", + "type": "text" + }, + { + "bbox": [ + 106, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "the self-supervised learning objective on two modalities. The popular VLP objectives can be divided", + "type": "text" + } + ], + "index": 52, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 89, + 702, + 100, + 711 + ], + "score": 1.0, + "content": "91", + "type": "text" + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "into reconstruction [25, 15, 8, 20], contrastive matching [39, 17, 16], or the combination of both", + "type": "text" + } + ], + "index": 53, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 89, + 713, + 100, + 722 + ], + "score": 1.0, + "content": "92", + "type": "text" + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "two [24, 54, 19]. Besides, recent progress in the field of VLP also benefits a lot from large-scale", + "type": "text" + } + ], + "index": 54, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 89, + 75, + 100, + 84 + ], + "score": 1.0, + "content": "93", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "pair-wised datasets. For example, CLIP [39] applies 400 million image-text pairs for contrastive", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 89, + 86, + 100, + 96 + ], + "score": 1.0, + "content": "94", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "learning, while ALIGN even exploits 1.8 billion data pairs. Beyond recognition, these VLP models", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 89, + 96, + 100, + 106 + ], + "score": 1.0, + "content": "95", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "also show great potential for other downstream applications, such as dense prediction [42, 62], image", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 105, + 342, + 119 + ], + "spans": [ + { + "bbox": [ + 88, + 107, + 100, + 117 + ], + "score": 1.0, + "content": "96", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 105, + 342, + 119 + ], + "score": 1.0, + "content": "generation [31, 41, 35], and action understanding [53, 48].", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 89, + 124, + 100, + 133 + ], + "score": 1.0, + "content": "97", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "Prompt Learning Prompt learning is introduced from the field of NLP to efficiently adapt the large", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 89, + 131, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 89, + 135, + 100, + 144 + ], + "score": 1.0, + "content": "98", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 131, + 505, + 147 + ], + "score": 1.0, + "content": "language model to downstream tasks. Different from the conventional “pre-training, fine-tuning”", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 142, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 88, + 146, + 100, + 155 + ], + "score": 1.0, + "content": "99", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 142, + 506, + 158 + ], + "score": 1.0, + "content": "paradigm which initializes the pre-trained model and tunes the parameters of the network using", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 155, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 86, + 156, + 100, + 167 + ], + "score": 1.0, + "content": "100", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 155, + 506, + 168 + ], + "score": 1.0, + "content": "downstream task-specific objective functions, prompt learning applies textual prompt to reformulate", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 86, + 168, + 99, + 176 + ], + "score": 1.0, + "content": "101", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "the downstream tasks as the original pretrained task [27, 36]. By the prompt, the domain shift between", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 86, + 178, + 100, + 189 + ], + "score": 1.0, + "content": "102", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "pretrained task and downstream application is reduced and thus the pretrained knowledge can be", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 86, + 190, + 99, + 199 + ], + "score": 1.0, + "content": "103", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "easier adapted to downstream tasks. The concept of prompt learning [36, 40, 38] begins from the", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 86, + 200, + 100, + 210 + ], + "score": 1.0, + "content": "104", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "success of GPT [40] series. Early prompt learning methods (such as Petroni et al. [36] and Pörner et", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 210, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 86, + 212, + 100, + 221 + ], + "score": 1.0, + "content": "105", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 210, + 506, + 222 + ], + "score": 1.0, + "content": "al. [38]) always manually create templates based on human prior knowledge. Furthermore, some", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 86, + 222, + 100, + 232 + ], + "score": 1.0, + "content": "106", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "mining-based methods [18] and gradient-based methods [46] are proposed to automatically search for", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 231, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 86, + 234, + 99, + 243 + ], + "score": 1.0, + "content": "107", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 231, + 506, + 244 + ], + "score": 1.0, + "content": "appropriate templates. Beyond search in the discrete space, some methods [26, 49, 28] remove the", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 241, + 507, + 254 + ], + "spans": [ + { + "bbox": [ + 86, + 244, + 100, + 254 + ], + "score": 1.0, + "content": "108", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 241, + 507, + 254 + ], + "score": 1.0, + "content": "constraint that the prompts are “words” and instead learn prompts in the continuous embedding space.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 254, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 86, + 255, + 100, + 265 + ], + "score": 1.0, + "content": "109", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 106, + 254, + 506, + 265 + ], + "score": 1.0, + "content": "Recently, CoOp [63] and its extended version [64] introduce prompt learning into open-world visual", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 85, + 266, + 100, + 276 + ], + "score": 1.0, + "content": "110", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "understanding to adapt the knowledge from the large-scale visual-language pretrained models and", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 273, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 85, + 276, + 100, + 286 + ], + "score": 1.0, + "content": "111", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 273, + 477, + 288 + ], + "score": 1.0, + "content": "achieve great performance improvement on the few-shot visual recognition. Compared with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 477, + 275, + 502, + 286 + ], + "score": 0.52, + "content": "\\mathrm { C o O p }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 503, + 273, + 506, + 288 + ], + "score": 1.0, + "content": ",", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 285, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 86, + 288, + 100, + 298 + ], + "score": 1.0, + "content": "112", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 105, + 285, + 507, + 299 + ], + "score": 1.0, + "content": "our PLOT method further improves prompt learning by introducing the optimal transport distance to", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 295, + 434, + 311 + ], + "spans": [ + { + "bbox": [ + 86, + 299, + 100, + 308 + ], + "score": 1.0, + "content": "113", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 104, + 295, + 434, + 311 + ], + "score": 1.0, + "content": "learn multiple local prompts and achieves fine-grained vision-language matching.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + } + ], + "index": 50.5, + "bbox_fs": [ + 88, + 635, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 90, + 72, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 89, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 89, + 75, + 100, + 84 + ], + "score": 1.0, + "content": "93", + "type": "text" + }, + { + "bbox": [ + 105, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "pair-wised datasets. For example, CLIP [39] applies 400 million image-text pairs for contrastive", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 89, + 86, + 100, + 96 + ], + "score": 1.0, + "content": "94", + "type": "text" + }, + { + "bbox": [ + 105, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "learning, while ALIGN even exploits 1.8 billion data pairs. Beyond recognition, these VLP models", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 89, + 96, + 100, + 106 + ], + "score": 1.0, + "content": "95", + "type": "text" + }, + { + "bbox": [ + 105, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "also show great potential for other downstream applications, such as dense prediction [42, 62], image", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 105, + 342, + 119 + ], + "spans": [ + { + "bbox": [ + 88, + 107, + 100, + 117 + ], + "score": 1.0, + "content": "96", + "type": "text" + }, + { + "bbox": [ + 105, + 105, + 342, + 119 + ], + "score": 1.0, + "content": "generation [31, 41, 35], and action understanding [53, 48].", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 87, + 122, + 506, + 308 + ], + "lines": [ + { + "bbox": [ + 89, + 121, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 89, + 124, + 100, + 133 + ], + "score": 1.0, + "content": "97", + "type": "text" + }, + { + "bbox": [ + 105, + 121, + 506, + 135 + ], + "score": 1.0, + "content": "Prompt Learning Prompt learning is introduced from the field of NLP to efficiently adapt the large", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 131, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 89, + 135, + 100, + 144 + ], + "score": 1.0, + "content": "98", + "type": "text" + }, + { + "bbox": [ + 104, + 131, + 505, + 147 + ], + "score": 1.0, + "content": "language model to downstream tasks. Different from the conventional “pre-training, fine-tuning”", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 142, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 88, + 146, + 100, + 155 + ], + "score": 1.0, + "content": "99", + "type": "text" + }, + { + "bbox": [ + 104, + 142, + 506, + 158 + ], + "score": 1.0, + "content": "paradigm which initializes the pre-trained model and tunes the parameters of the network using", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 155, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 86, + 156, + 100, + 167 + ], + "score": 1.0, + "content": "100", + "type": "text" + }, + { + "bbox": [ + 105, + 155, + 506, + 168 + ], + "score": 1.0, + "content": "downstream task-specific objective functions, prompt learning applies textual prompt to reformulate", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 86, + 168, + 99, + 176 + ], + "score": 1.0, + "content": "101", + "type": "text" + }, + { + "bbox": [ + 105, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "the downstream tasks as the original pretrained task [27, 36]. By the prompt, the domain shift between", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 86, + 178, + 100, + 189 + ], + "score": 1.0, + "content": "102", + "type": "text" + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "pretrained task and downstream application is reduced and thus the pretrained knowledge can be", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 86, + 190, + 99, + 199 + ], + "score": 1.0, + "content": "103", + "type": "text" + }, + { + "bbox": [ + 105, + 187, + 506, + 200 + ], + "score": 1.0, + "content": "easier adapted to downstream tasks. The concept of prompt learning [36, 40, 38] begins from the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 86, + 200, + 100, + 210 + ], + "score": 1.0, + "content": "104", + "type": "text" + }, + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "success of GPT [40] series. Early prompt learning methods (such as Petroni et al. [36] and Pörner et", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 210, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 86, + 212, + 100, + 221 + ], + "score": 1.0, + "content": "105", + "type": "text" + }, + { + "bbox": [ + 106, + 210, + 506, + 222 + ], + "score": 1.0, + "content": "al. [38]) always manually create templates based on human prior knowledge. Furthermore, some", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 86, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 86, + 222, + 100, + 232 + ], + "score": 1.0, + "content": "106", + "type": "text" + }, + { + "bbox": [ + 105, + 220, + 506, + 232 + ], + "score": 1.0, + "content": "mining-based methods [18] and gradient-based methods [46] are proposed to automatically search for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 86, + 231, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 86, + 234, + 99, + 243 + ], + "score": 1.0, + "content": "107", + "type": "text" + }, + { + "bbox": [ + 105, + 231, + 506, + 244 + ], + "score": 1.0, + "content": "appropriate templates. Beyond search in the discrete space, some methods [26, 49, 28] remove the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 86, + 241, + 507, + 254 + ], + "spans": [ + { + "bbox": [ + 86, + 244, + 100, + 254 + ], + "score": 1.0, + "content": "108", + "type": "text" + }, + { + "bbox": [ + 105, + 241, + 507, + 254 + ], + "score": 1.0, + "content": "constraint that the prompts are “words” and instead learn prompts in the continuous embedding space.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 86, + 254, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 86, + 255, + 100, + 265 + ], + "score": 1.0, + "content": "109", + "type": "text" + }, + { + "bbox": [ + 106, + 254, + 506, + 265 + ], + "score": 1.0, + "content": "Recently, CoOp [63] and its extended version [64] introduce prompt learning into open-world visual", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 85, + 263, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 85, + 266, + 100, + 276 + ], + "score": 1.0, + "content": "110", + "type": "text" + }, + { + "bbox": [ + 105, + 263, + 506, + 276 + ], + "score": 1.0, + "content": "understanding to adapt the knowledge from the large-scale visual-language pretrained models and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 85, + 273, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 85, + 276, + 100, + 286 + ], + "score": 1.0, + "content": "111", + "type": "text" + }, + { + "bbox": [ + 105, + 273, + 477, + 288 + ], + "score": 1.0, + "content": "achieve great performance improvement on the few-shot visual recognition. Compared with", + "type": "text" + }, + { + "bbox": [ + 477, + 275, + 502, + 286 + ], + "score": 0.52, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 273, + 506, + 288 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 86, + 285, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 86, + 288, + 100, + 298 + ], + "score": 1.0, + "content": "112", + "type": "text" + }, + { + "bbox": [ + 105, + 285, + 507, + 299 + ], + "score": 1.0, + "content": "our PLOT method further improves prompt learning by introducing the optimal transport distance to", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 86, + 295, + 434, + 311 + ], + "spans": [ + { + "bbox": [ + 86, + 299, + 100, + 308 + ], + "score": 1.0, + "content": "113", + "type": "text" + }, + { + "bbox": [ + 104, + 295, + 434, + 311 + ], + "score": 1.0, + "content": "learn multiple local prompts and achieves fine-grained vision-language matching.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 95, + 319, + 176, + 333 + ], + "lines": [ + { + "bbox": [ + 90, + 316, + 178, + 336 + ], + "spans": [ + { + "bbox": [ + 90, + 316, + 178, + 336 + ], + "score": 1.0, + "content": "14 3 Approach", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 105, + 339, + 503, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 341, + 351 + ], + "score": 1.0, + "content": "In this section we will first revisit the baseline method", + "type": "text" + }, + { + "bbox": [ + 342, + 339, + 384, + 351 + ], + "score": 0.27, + "content": "\\mathrm { C o O p } \\ 3 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 339, + 505, + 351 + ], + "score": 1.0, + "content": ", review the preliminaries of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 350, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 505, + 362 + ], + "score": 1.0, + "content": "optimal transport 3.2, and then introduce our proposed PLOT 3.3 to show how we can learn multiple", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 361, + 206, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 206, + 374 + ], + "score": 1.0, + "content": "comprehensive prompts.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 93, + 380, + 209, + 392 + ], + "lines": [ + { + "bbox": [ + 89, + 379, + 210, + 395 + ], + "spans": [ + { + "bbox": [ + 89, + 379, + 210, + 395 + ], + "score": 1.0, + "content": "118 3.1 A Revisit of CoOp", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 87, + 396, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 86, + 397, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 86, + 399, + 99, + 408 + ], + "score": 1.0, + "content": "119", + "type": "text" + }, + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "score": 1.0, + "content": "CoOp [63] is one of the pioneering methods to learn the prompts for using vision language pretrained", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 86, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 86, + 410, + 99, + 419 + ], + "score": 1.0, + "content": "120", + "type": "text" + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "knowledge (such as CLIP [39]) for downstream open-world visual recognition. Different from CLIP", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 86, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 86, + 420, + 99, + 429 + ], + "score": 1.0, + "content": "121", + "type": "text" + }, + { + "bbox": [ + 106, + 419, + 294, + 430 + ], + "score": 1.0, + "content": "which manually designs the prompt templates,", + "type": "text" + }, + { + "bbox": [ + 294, + 418, + 320, + 430 + ], + "score": 0.42, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 419, + 505, + 430 + ], + "score": 1.0, + "content": "sets a part of context words in the template as", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 86, + 429, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 86, + 431, + 100, + 441 + ], + "score": 1.0, + "content": "122", + "type": "text" + }, + { + "bbox": [ + 105, + 429, + 506, + 441 + ], + "score": 1.0, + "content": "continuous learnable parameters which can be learned from the few-shot data. Then the classification", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 86, + 440, + 468, + 453 + ], + "spans": [ + { + "bbox": [ + 86, + 442, + 99, + 451 + ], + "score": 1.0, + "content": "123", + "type": "text" + }, + { + "bbox": [ + 105, + 440, + 468, + 453 + ], + "score": 1.0, + "content": "weights can be represented by the distance between the learned prompt and visual feature.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 86, + 455, + 505, + 523 + ], + "lines": [ + { + "bbox": [ + 86, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 86, + 457, + 100, + 468 + ], + "score": 1.0, + "content": "124", + "type": "text" + }, + { + "bbox": [ + 105, + 455, + 225, + 468 + ], + "score": 1.0, + "content": "Specifically, given an image", + "type": "text" + }, + { + "bbox": [ + 225, + 459, + 232, + 466 + ], + "score": 0.72, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 455, + 302, + 468 + ], + "score": 1.0, + "content": ", a visual feature", + "type": "text" + }, + { + "bbox": [ + 303, + 456, + 345, + 468 + ], + "score": 0.92, + "content": "\\pmb { f } = f ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 455, + 485, + 468 + ], + "score": 1.0, + "content": "is obtained by the visual encoder", + "type": "text" + }, + { + "bbox": [ + 485, + 457, + 492, + 468 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 86, + 466, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 86, + 469, + 100, + 479 + ], + "score": 1.0, + "content": "125", + "type": "text" + }, + { + "bbox": [ + 105, + 466, + 314, + 481 + ], + "score": 1.0, + "content": "CLIP. Then, the textual prompt can be formulated as", + "type": "text" + }, + { + "bbox": [ + 315, + 468, + 453, + 479 + ], + "score": 0.89, + "content": "\\pmb { t } _ { k } = \\{ \\pmb { v e c } _ { 1 } , \\pmb { v e c } _ { 2 } , \\dots , \\pmb { v e c } _ { L } , \\pmb { c } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 466, + 483, + 481 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 484, + 469, + 495, + 479 + ], + "score": 0.87, + "content": "c _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 466, + 506, + 481 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 86, + 477, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 86, + 480, + 100, + 490 + ], + "score": 1.0, + "content": "126", + "type": "text" + }, + { + "bbox": [ + 105, + 477, + 263, + 491 + ], + "score": 1.0, + "content": "the word embedding of the class name,", + "type": "text" + }, + { + "bbox": [ + 263, + 478, + 309, + 491 + ], + "score": 0.93, + "content": "\\{ v e c _ { l } | _ { l = 1 } ^ { L } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 477, + 506, + 491 + ], + "score": 1.0, + "content": "are learnable vectors with the same dimension as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 86, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 86, + 491, + 99, + 500 + ], + "score": 1.0, + "content": "127", + "type": "text" + }, + { + "bbox": [ + 105, + 488, + 241, + 501 + ], + "score": 1.0, + "content": "the original word embedding and", + "type": "text" + }, + { + "bbox": [ + 241, + 489, + 249, + 499 + ], + "score": 0.64, + "content": "\\mathrm { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 488, + 427, + 501 + ], + "score": 1.0, + "content": "is the length of context words. With prompt", + "type": "text" + }, + { + "bbox": [ + 428, + 490, + 438, + 500 + ], + "score": 0.87, + "content": "\\mathbf { \\Delta } _ { t _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "as the input, the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 86, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 86, + 502, + 100, + 512 + ], + "score": 1.0, + "content": "128", + "type": "text" + }, + { + "bbox": [ + 105, + 500, + 156, + 513 + ], + "score": 1.0, + "content": "text encoder", + "type": "text" + }, + { + "bbox": [ + 157, + 502, + 163, + 512 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 500, + 279, + 513 + ], + "score": 1.0, + "content": "outputs the textual feature as", + "type": "text" + }, + { + "bbox": [ + 279, + 500, + 326, + 512 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathbf { g } _ { k } = g ( \\mathbf { \\hat { t } } _ { k } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 500, + 505, + 513 + ], + "score": 1.0, + "content": ". The final prediction probability is computed", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 86, + 511, + 242, + 523 + ], + "spans": [ + { + "bbox": [ + 86, + 513, + 100, + 522 + ], + "score": 1.0, + "content": "129", + "type": "text" + }, + { + "bbox": [ + 105, + 511, + 242, + 523 + ], + "score": 1.0, + "content": "by the matching score as follows:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 528, + 393, + 558 + ], + "lines": [ + { + "bbox": [ + 217, + 528, + 393, + 558 + ], + "spans": [ + { + "bbox": [ + 217, + 528, + 393, + 558 + ], + "score": 0.94, + "content": "p ( y = k | \\pmb { x } ) = \\frac { e x p ( \\sin ( \\pmb { f } , \\pmb { g } _ { k } ) / \\tau ) } { \\sum _ { k ^ { \\prime } = 1 } ^ { K } e x p ( \\sin ( \\pmb { f } , \\pmb { g } _ { k ^ { \\prime } } ) / \\tau ) } ,", + "type": "interline_equation", + "image_path": "5390e40a4870304fef575eadc9a89e1c5c6a24f464e8c8244af2af5b74d16de2.jpg" + } + ] + } + ], + "index": 37.5, + "virtual_lines": [ + { + "bbox": [ + 217, + 528, + 393, + 543.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 217, + 543.0, + 393, + 558.0 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "text", + "bbox": [ + 98, + 563, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 562, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 133, + 577 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 563, + 166, + 576 + ], + "score": 0.9, + "content": "\\sin ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 562, + 390, + 577 + ], + "score": 1.0, + "content": "denotes a metric function such as cosine similarity, and", + "type": "text" + }, + { + "bbox": [ + 390, + 566, + 397, + 573 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 562, + 506, + 577 + ], + "score": 1.0, + "content": "stands for the temperature", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 573, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 315, + 589 + ], + "score": 1.0, + "content": "of Softmax. Then we can optimize the parameters of", + "type": "text" + }, + { + "bbox": [ + 315, + 574, + 362, + 587 + ], + "score": 0.94, + "content": "\\{ v e c _ { l } | _ { l = 1 } ^ { L } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 573, + 506, + 589 + ], + "score": 1.0, + "content": "with the cross-entropy loss between", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 585, + 254, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 254, + 598 + ], + "score": 1.0, + "content": "the prediction and the labeled target.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "title", + "bbox": [ + 106, + 605, + 212, + 617 + ], + "lines": [ + { + "bbox": [ + 104, + 604, + 213, + 620 + ], + "spans": [ + { + "bbox": [ + 104, + 604, + 213, + 620 + ], + "score": 1.0, + "content": "3.2 Optimal Transport", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 93, + 621, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 92, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 92, + 624, + 99, + 631 + ], + "score": 1.0, + "content": "34", + "type": "text" + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "Optimal transport (OT) distance is a widely used metric for the comparison of distributions. Here, we", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 92, + 632, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 92, + 635, + 99, + 642 + ], + "score": 1.0, + "content": "35", + "type": "text" + }, + { + "bbox": [ + 104, + 632, + 505, + 644 + ], + "score": 1.0, + "content": "only focus on the discrete situation which is more related to our framework. Assuming we have two", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 92, + 643, + 380, + 655 + ], + "spans": [ + { + "bbox": [ + 92, + 646, + 99, + 653 + ], + "score": 1.0, + "content": "36", + "type": "text" + }, + { + "bbox": [ + 103, + 643, + 380, + 655 + ], + "score": 1.0, + "content": "sets of points (features), the discrete distributions are formulated as:", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 660, + 401, + 694 + ], + "lines": [ + { + "bbox": [ + 210, + 660, + 401, + 694 + ], + "spans": [ + { + "bbox": [ + 210, + 660, + 401, + 694 + ], + "score": 0.94, + "content": "U = \\sum _ { m = 1 } ^ { M } u _ { m } \\delta _ { f _ { m } } \\qquad \\mathrm { a n d } \\qquad V = \\sum _ { n = 1 } ^ { N } v _ { n } \\delta _ { g _ { n } } ,", + "type": "interline_equation", + "image_path": "400262e1ffbb02047ef330acecd754adaed327a7b6c8db82178ed2d19b14989b.jpg" + } + ] + } + ], + "index": 46.5, + "virtual_lines": [ + { + "bbox": [ + 210, + 660, + 401, + 677.0 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 210, + 677.0, + 401, + 694.0 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "text", + "bbox": [ + 84, + 699, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 87, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 87, + 702, + 100, + 711 + ], + "score": 1.0, + "content": "137", + "type": "text" + }, + { + "bbox": [ + 103, + 700, + 133, + 712 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 702, + 142, + 710 + ], + "score": 0.76, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 700, + 161, + 712 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 161, + 702, + 168, + 710 + ], + "score": 0.75, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 700, + 390, + 712 + ], + "score": 1.0, + "content": "are the discrete probability vectors that sum to 1, and", + "type": "text" + }, + { + "bbox": [ + 391, + 700, + 402, + 712 + ], + "score": 0.89, + "content": "\\delta _ { f }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "is a Dirac delta function", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 84, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 84, + 711, + 200, + 723 + ], + "score": 1.0, + "content": "138 placed at support point", + "type": "text" + }, + { + "bbox": [ + 200, + 711, + 207, + 722 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "in the embedding space. Then, the total distance of these two distributions", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "index", + "bbox": [ + 90, + 72, + 505, + 117 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 88, + 73, + 505, + 119 + ], + "lines_deleted": true + }, + { + "type": "index", + "bbox": [ + 87, + 122, + 506, + 308 + ], + "lines": [], + "index": 12, + "bbox_fs": [ + 85, + 121, + 507, + 311 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 95, + 319, + 176, + 333 + ], + "lines": [ + { + "bbox": [ + 90, + 316, + 178, + 336 + ], + "spans": [ + { + "bbox": [ + 90, + 316, + 178, + 336 + ], + "score": 1.0, + "content": "14 3 Approach", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 105, + 339, + 503, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 339, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 341, + 351 + ], + "score": 1.0, + "content": "In this section we will first revisit the baseline method", + "type": "text" + }, + { + "bbox": [ + 342, + 339, + 384, + 351 + ], + "score": 0.27, + "content": "\\mathrm { C o O p } \\ 3 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 384, + 339, + 505, + 351 + ], + "score": 1.0, + "content": ", review the preliminaries of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 350, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 505, + 362 + ], + "score": 1.0, + "content": "optimal transport 3.2, and then introduce our proposed PLOT 3.3 to show how we can learn multiple", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 361, + 206, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 206, + 374 + ], + "score": 1.0, + "content": "comprehensive prompts.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 339, + 505, + 374 + ] + }, + { + "type": "title", + "bbox": [ + 93, + 380, + 209, + 392 + ], + "lines": [ + { + "bbox": [ + 89, + 379, + 210, + 395 + ], + "spans": [ + { + "bbox": [ + 89, + 379, + 210, + 395 + ], + "score": 1.0, + "content": "118 3.1 A Revisit of CoOp", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "index", + "bbox": [ + 87, + 396, + 505, + 451 + ], + "lines": [ + { + "bbox": [ + 86, + 397, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 86, + 399, + 99, + 408 + ], + "score": 1.0, + "content": "119", + "type": "text" + }, + { + "bbox": [ + 106, + 397, + 505, + 408 + ], + "score": 1.0, + "content": "CoOp [63] is one of the pioneering methods to learn the prompts for using vision language pretrained", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 86, + 410, + 99, + 419 + ], + "score": 1.0, + "content": "120", + "type": "text" + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "knowledge (such as CLIP [39]) for downstream open-world visual recognition. Different from CLIP", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 86, + 420, + 99, + 429 + ], + "score": 1.0, + "content": "121", + "type": "text" + }, + { + "bbox": [ + 106, + 419, + 294, + 430 + ], + "score": 1.0, + "content": "which manually designs the prompt templates,", + "type": "text" + }, + { + "bbox": [ + 294, + 418, + 320, + 430 + ], + "score": 0.42, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 419, + 505, + 430 + ], + "score": 1.0, + "content": "sets a part of context words in the template as", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 429, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 86, + 431, + 100, + 441 + ], + "score": 1.0, + "content": "122", + "type": "text" + }, + { + "bbox": [ + 105, + 429, + 506, + 441 + ], + "score": 1.0, + "content": "continuous learnable parameters which can be learned from the few-shot data. Then the classification", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 440, + 468, + 453 + ], + "spans": [ + { + "bbox": [ + 86, + 442, + 99, + 451 + ], + "score": 1.0, + "content": "123", + "type": "text" + }, + { + "bbox": [ + 105, + 440, + 468, + 453 + ], + "score": 1.0, + "content": "weights can be represented by the distance between the learned prompt and visual feature.", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 455, + 506, + 468 + ], + "spans": [ + { + "bbox": [ + 86, + 457, + 100, + 468 + ], + "score": 1.0, + "content": "124", + "type": "text" + }, + { + "bbox": [ + 105, + 455, + 225, + 468 + ], + "score": 1.0, + "content": "Specifically, given an image", + "type": "text" + }, + { + "bbox": [ + 225, + 459, + 232, + 466 + ], + "score": 0.72, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 455, + 302, + 468 + ], + "score": 1.0, + "content": ", a visual feature", + "type": "text" + }, + { + "bbox": [ + 303, + 456, + 345, + 468 + ], + "score": 0.92, + "content": "\\pmb { f } = f ( \\pmb { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 455, + 485, + 468 + ], + "score": 1.0, + "content": "is obtained by the visual encoder", + "type": "text" + }, + { + "bbox": [ + 485, + 457, + 492, + 468 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 455, + 506, + 468 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 466, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 86, + 469, + 100, + 479 + ], + "score": 1.0, + "content": "125", + "type": "text" + }, + { + "bbox": [ + 105, + 466, + 314, + 481 + ], + "score": 1.0, + "content": "CLIP. Then, the textual prompt can be formulated as", + "type": "text" + }, + { + "bbox": [ + 315, + 468, + 453, + 479 + ], + "score": 0.89, + "content": "\\pmb { t } _ { k } = \\{ \\pmb { v e c } _ { 1 } , \\pmb { v e c } _ { 2 } , \\dots , \\pmb { v e c } _ { L } , \\pmb { c } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 466, + 483, + 481 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 484, + 469, + 495, + 479 + ], + "score": 0.87, + "content": "c _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 466, + 506, + 481 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 477, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 86, + 480, + 100, + 490 + ], + "score": 1.0, + "content": "126", + "type": "text" + }, + { + "bbox": [ + 105, + 477, + 263, + 491 + ], + "score": 1.0, + "content": "the word embedding of the class name,", + "type": "text" + }, + { + "bbox": [ + 263, + 478, + 309, + 491 + ], + "score": 0.93, + "content": "\\{ v e c _ { l } | _ { l = 1 } ^ { L } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 477, + 506, + 491 + ], + "score": 1.0, + "content": "are learnable vectors with the same dimension as", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 86, + 491, + 99, + 500 + ], + "score": 1.0, + "content": "127", + "type": "text" + }, + { + "bbox": [ + 105, + 488, + 241, + 501 + ], + "score": 1.0, + "content": "the original word embedding and", + "type": "text" + }, + { + "bbox": [ + 241, + 489, + 249, + 499 + ], + "score": 0.64, + "content": "\\mathrm { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 488, + 427, + 501 + ], + "score": 1.0, + "content": "is the length of context words. With prompt", + "type": "text" + }, + { + "bbox": [ + 428, + 490, + 438, + 500 + ], + "score": 0.87, + "content": "\\mathbf { \\Delta } _ { t _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "as the input, the", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 86, + 502, + 100, + 512 + ], + "score": 1.0, + "content": "128", + "type": "text" + }, + { + "bbox": [ + 105, + 500, + 156, + 513 + ], + "score": 1.0, + "content": "text encoder", + "type": "text" + }, + { + "bbox": [ + 157, + 502, + 163, + 512 + ], + "score": 0.8, + "content": "g", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 500, + 279, + 513 + ], + "score": 1.0, + "content": "outputs the textual feature as", + "type": "text" + }, + { + "bbox": [ + 279, + 500, + 326, + 512 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathbf { g } _ { k } = g ( \\mathbf { \\hat { t } } _ { k } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 500, + 505, + 513 + ], + "score": 1.0, + "content": ". The final prediction probability is computed", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 511, + 242, + 523 + ], + "spans": [ + { + "bbox": [ + 86, + 513, + 100, + 522 + ], + "score": 1.0, + "content": "129", + "type": "text" + }, + { + "bbox": [ + 105, + 511, + 242, + 523 + ], + "score": 1.0, + "content": "by the matching score as follows:", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + } + ], + "index": 28, + "bbox_fs": [ + 86, + 397, + 506, + 453 + ] + }, + { + "type": "index", + "bbox": [ + 86, + 455, + 505, + 523 + ], + "lines": [], + "index": 33.5, + "bbox_fs": [ + 86, + 455, + 506, + 523 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 528, + 393, + 558 + ], + "lines": [ + { + "bbox": [ + 217, + 528, + 393, + 558 + ], + "spans": [ + { + "bbox": [ + 217, + 528, + 393, + 558 + ], + "score": 0.94, + "content": "p ( y = k | \\pmb { x } ) = \\frac { e x p ( \\sin ( \\pmb { f } , \\pmb { g } _ { k } ) / \\tau ) } { \\sum _ { k ^ { \\prime } = 1 } ^ { K } e x p ( \\sin ( \\pmb { f } , \\pmb { g } _ { k ^ { \\prime } } ) / \\tau ) } ,", + "type": "interline_equation", + "image_path": "5390e40a4870304fef575eadc9a89e1c5c6a24f464e8c8244af2af5b74d16de2.jpg" + } + ] + } + ], + "index": 37.5, + "virtual_lines": [ + { + "bbox": [ + 217, + 528, + 393, + 543.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 217, + 543.0, + 393, + 558.0 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "text", + "bbox": [ + 98, + 563, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 562, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 133, + 577 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 563, + 166, + 576 + ], + "score": 0.9, + "content": "\\sin ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 562, + 390, + 577 + ], + "score": 1.0, + "content": "denotes a metric function such as cosine similarity, and", + "type": "text" + }, + { + "bbox": [ + 390, + 566, + 397, + 573 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 562, + 506, + 577 + ], + "score": 1.0, + "content": "stands for the temperature", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 573, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 315, + 589 + ], + "score": 1.0, + "content": "of Softmax. Then we can optimize the parameters of", + "type": "text" + }, + { + "bbox": [ + 315, + 574, + 362, + 587 + ], + "score": 0.94, + "content": "\\{ v e c _ { l } | _ { l = 1 } ^ { L } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 573, + 506, + 589 + ], + "score": 1.0, + "content": "with the cross-entropy loss between", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 585, + 254, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 254, + 598 + ], + "score": 1.0, + "content": "the prediction and the labeled target.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 562, + 506, + 598 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 605, + 212, + 617 + ], + "lines": [ + { + "bbox": [ + 104, + 604, + 213, + 620 + ], + "spans": [ + { + "bbox": [ + 104, + 604, + 213, + 620 + ], + "score": 1.0, + "content": "3.2 Optimal Transport", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "index", + "bbox": [ + 93, + 621, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 92, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 92, + 624, + 99, + 631 + ], + "score": 1.0, + "content": "34", + "type": "text" + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "Optimal transport (OT) distance is a widely used metric for the comparison of distributions. Here, we", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 632, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 92, + 635, + 99, + 642 + ], + "score": 1.0, + "content": "35", + "type": "text" + }, + { + "bbox": [ + 104, + 632, + 505, + 644 + ], + "score": 1.0, + "content": "only focus on the discrete situation which is more related to our framework. Assuming we have two", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 92, + 643, + 380, + 655 + ], + "spans": [ + { + "bbox": [ + 92, + 646, + 99, + 653 + ], + "score": 1.0, + "content": "36", + "type": "text" + }, + { + "bbox": [ + 103, + 643, + 380, + 655 + ], + "score": 1.0, + "content": "sets of points (features), the discrete distributions are formulated as:", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + } + ], + "index": 44, + "bbox_fs": [ + 92, + 621, + 506, + 655 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 660, + 401, + 694 + ], + "lines": [ + { + "bbox": [ + 210, + 660, + 401, + 694 + ], + "spans": [ + { + "bbox": [ + 210, + 660, + 401, + 694 + ], + "score": 0.94, + "content": "U = \\sum _ { m = 1 } ^ { M } u _ { m } \\delta _ { f _ { m } } \\qquad \\mathrm { a n d } \\qquad V = \\sum _ { n = 1 } ^ { N } v _ { n } \\delta _ { g _ { n } } ,", + "type": "interline_equation", + "image_path": "400262e1ffbb02047ef330acecd754adaed327a7b6c8db82178ed2d19b14989b.jpg" + } + ] + } + ], + "index": 46.5, + "virtual_lines": [ + { + "bbox": [ + 210, + 660, + 401, + 677.0 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 210, + 677.0, + 401, + 694.0 + ], + "spans": [], + "index": 47 + } + ] + }, + { + "type": "index", + "bbox": [ + 84, + 699, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 87, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 87, + 702, + 100, + 711 + ], + "score": 1.0, + "content": "137", + "type": "text" + }, + { + "bbox": [ + 103, + 700, + 133, + 712 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 702, + 142, + 710 + ], + "score": 0.76, + "content": "\\textbf { \\em u }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 700, + 161, + 712 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 161, + 702, + 168, + 710 + ], + "score": 0.75, + "content": "\\textbf { { v } }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 700, + 390, + 712 + ], + "score": 1.0, + "content": "are the discrete probability vectors that sum to 1, and", + "type": "text" + }, + { + "bbox": [ + 391, + 700, + 402, + 712 + ], + "score": 0.89, + "content": "\\delta _ { f }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "is a Dirac delta function", + "type": "text" + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 84, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 84, + 711, + 200, + 723 + ], + "score": 1.0, + "content": "138 placed at support point", + "type": "text" + }, + { + "bbox": [ + 200, + 711, + 207, + 722 + ], + "score": 0.86, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "in the embedding space. Then, the total distance of these two distributions", + "type": "text" + } + ], + "index": 49, + "is_list_start_line": true + } + ], + "index": 48.5, + "bbox_fs": [ + 84, + 700, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 72, + 505, + 194 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 72, + 505, + 194 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 194 + ], + "score": 0.97, + "type": "image", + "image_path": "2d85bb34d9ff604845e43487f68a39419971c3db82cc68daca2f31d8e66755e1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 72, + 505, + 112.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 112.66666666666666, + 505, + 153.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 153.33333333333331, + 505, + 193.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 200, + 505, + 235 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 200, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 506, + 213 + ], + "score": 1.0, + "content": "Figure 2: The framework of PLOT. PLOT first describes each category with multiple prompts and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 212, + 507, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 507, + 225 + ], + "score": 1.0, + "content": "obtains a set of prompt features by text encoder. The image is also encoded as a set of local features.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 222, + 448, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 448, + 235 + ], + "score": 1.0, + "content": "Then the optimal transport is used as the metric between prompts and visual features.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 87, + 244, + 164, + 255 + ], + "lines": [ + { + "bbox": [ + 85, + 243, + 166, + 255 + ], + "spans": [ + { + "bbox": [ + 85, + 243, + 166, + 255 + ], + "score": 1.0, + "content": "139 are written as:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 251, + 375, + 286 + ], + "lines": [ + { + "bbox": [ + 236, + 251, + 375, + 286 + ], + "spans": [ + { + "bbox": [ + 236, + 251, + 375, + 286 + ], + "score": 0.95, + "content": "< T , C > = \\sum _ { m = 1 } ^ { M } \\sum _ { n = 1 } ^ { N } T _ { m , n } { \\cal C } _ { m , n } .", + "type": "interline_equation", + "image_path": "73c6d5d67053831369f730face7a896aa4f2dbf1f716ad4663ef50c2533427e6.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 236, + 251, + 375, + 268.5 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 236, + 268.5, + 375, + 286.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 86, + 286, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 86, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 86, + 288, + 100, + 298 + ], + "score": 1.0, + "content": "140", + "type": "text" + }, + { + "bbox": [ + 103, + 285, + 141, + 299 + ], + "score": 1.0, + "content": "We call", + "type": "text" + }, + { + "bbox": [ + 142, + 287, + 152, + 297 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 285, + 418, + 299 + ], + "score": 1.0, + "content": "the cost matrix in which each point denotes the cost between", + "type": "text" + }, + { + "bbox": [ + 419, + 287, + 433, + 298 + ], + "score": 0.89, + "content": "f _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 285, + 453, + 299 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 453, + 288, + 466, + 298 + ], + "score": 0.85, + "content": "\\scriptstyle { \\pmb { g } } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 285, + 506, + 299 + ], + "score": 1.0, + "content": ", such as", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 296, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 85, + 299, + 100, + 309 + ], + "score": 1.0, + "content": "141", + "type": "text" + }, + { + "bbox": [ + 107, + 297, + 212, + 309 + ], + "score": 0.91, + "content": "C _ { m , n } = 1 - \\sin ( { f _ { m } } , { \\pmb { g } } _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 296, + 256, + 311 + ], + "score": 1.0, + "content": ". While the", + "type": "text" + }, + { + "bbox": [ + 256, + 298, + 266, + 307 + ], + "score": 0.81, + "content": "\\mathbf { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 296, + 506, + 311 + ], + "score": 1.0, + "content": "is called the transport plan, which is learned to minimize the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 87, + 308, + 423, + 320 + ], + "spans": [ + { + "bbox": [ + 87, + 311, + 100, + 319 + ], + "score": 1.0, + "content": "142", + "type": "text" + }, + { + "bbox": [ + 105, + 308, + 423, + 320 + ], + "score": 1.0, + "content": "total distance. The optimization problem of optimal transport is formulated as:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "interline_equation", + "bbox": [ + 221, + 322, + 390, + 358 + ], + "lines": [ + { + "bbox": [ + 221, + 322, + 390, + 358 + ], + "spans": [ + { + "bbox": [ + 221, + 322, + 390, + 358 + ], + "score": 0.67, + "content": "\\begin{array} { l l } { { d _ { O T } ( { \\pmb u } , { \\pmb v } | { \\cal C } ) = \\underset { \\cal T } { \\mathrm { m i n i m i z e } } < { \\cal T } , { \\cal C } > } } \\\\ { { \\mathrm { s u b j e c t ~ t o ~ } } } & { { { \\cal T } { \\bf 1 } = { \\pmb u } , { \\cal T } ^ { T } { \\bf 1 } = { \\pmb v } , { \\cal T } \\geq 0 . } } \\end{array}", + "type": "interline_equation", + "image_path": "492430b29380fae5e25024cb766d92baaf5e617b5cc37bc9abba3343dcfe568e.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 221, + 322, + 390, + 340.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 221, + 340.0, + 390, + 358.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 86, + 365, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 84, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 84, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "143 As directly optimizing the above objective is always time-consuming, we apply the Sinkhorn dis-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 87, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 87, + 379, + 100, + 387 + ], + "score": 1.0, + "content": "144", + "type": "text" + }, + { + "bbox": [ + 104, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "tance [6] to use an entropic constraint for fast optimization. The optimization problem with a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 86, + 388, + 301, + 400 + ], + "spans": [ + { + "bbox": [ + 86, + 390, + 100, + 399 + ], + "score": 1.0, + "content": "145", + "type": "text" + }, + { + "bbox": [ + 104, + 388, + 301, + 400 + ], + "score": 1.0, + "content": "Lagrange multiplier of the entropy constraint is:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 401, + 405, + 437 + ], + "lines": [ + { + "bbox": [ + 205, + 401, + 405, + 437 + ], + "spans": [ + { + "bbox": [ + 205, + 401, + 405, + 437 + ], + "score": 0.69, + "content": "\\begin{array} { r l } & { d _ { O T , \\lambda } ( \\pmb { u } , \\pmb { v } | C ) = \\underset { T } { \\mathrm { m i n i m i z e } } < T , C > - \\lambda h ( \\pmb { T } ) } \\\\ & { } \\\\ & { \\mathrm { s u b j e c t ~ t o } \\quad T \\mathbf { 1 } = \\pmb { u } , T ^ { T } \\mathbf { 1 } = \\pmb { v } , } \\end{array}", + "type": "interline_equation", + "image_path": "3c81eb9ce7b52c2923e3d3c8467ed9a16f9ec73e9e8f6ee25037f99bcc4d8a01.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 205, + 401, + 405, + 419.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 205, + 419.0, + 405, + 437.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 91, + 439, + 504, + 461 + ], + "lines": [ + { + "bbox": [ + 87, + 439, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 87, + 439, + 132, + 453 + ], + "score": 1.0, + "content": "146 where", + "type": "text" + }, + { + "bbox": [ + 133, + 440, + 150, + 452 + ], + "score": 0.91, + "content": "h ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 439, + 209, + 453 + ], + "score": 1.0, + "content": "is entropy and", + "type": "text" + }, + { + "bbox": [ + 210, + 440, + 235, + 451 + ], + "score": 0.93, + "content": "\\lambda \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 439, + 505, + 453 + ], + "score": 1.0, + "content": "is a hyper-parameter. Then we can have a fast optimization solution", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 88, + 450, + 204, + 463 + ], + "spans": [ + { + "bbox": [ + 88, + 450, + 204, + 463 + ], + "score": 1.0, + "content": "147 with a few iterations as:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 460, + 379, + 474 + ], + "lines": [ + { + "bbox": [ + 229, + 460, + 379, + 474 + ], + "spans": [ + { + "bbox": [ + 229, + 460, + 379, + 474 + ], + "score": 0.89, + "content": "\\pmb { T } ^ { * } = \\mathrm { d i a g } ( \\pmb { u } ^ { t } ) e x p ( - \\pmb { C } / \\lambda ) \\mathrm { d i a g } ( \\pmb { v } ^ { t } ) ,", + "type": "interline_equation", + "image_path": "c4fffb8dc5cdbb88c0da5ddf3cd3cc92a4317c99e2c970ad2bd06ea091e11e57.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 229, + 460, + 379, + 474 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 101, + 475, + 506, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 504, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 136, + 488 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 136, + 477, + 142, + 485 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 473, + 320, + 488 + ], + "score": 1.0, + "content": "denotes iteration and in each iteration", + "type": "text" + }, + { + "bbox": [ + 321, + 475, + 451, + 487 + ], + "score": 0.89, + "content": "\\begin{array} { c c l } { { { \\pmb u } ^ { t } } } & { { = } } & { { { \\pmb u } / ( ( e x p ( - { \\pmb C } / \\lambda ) { \\pmb v } ^ { t - 1 } ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 473, + 476, + 488 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 476, + 474, + 504, + 487 + ], + "score": 0.87, + "content": "\\begin{array} { r l } { \\boldsymbol { v } ^ { t } } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 485, + 310, + 499 + ], + "spans": [ + { + "bbox": [ + 107, + 486, + 197, + 499 + ], + "score": 0.92, + "content": "\\pmb { v } / ( ( e x p ( - C / \\lambda ) ^ { T } \\pmb { u } ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 485, + 274, + 499 + ], + "score": 1.0, + "content": ", with the initiation", + "type": "text" + }, + { + "bbox": [ + 275, + 486, + 305, + 497 + ], + "score": 0.89, + "content": "\\mathbf { \\nabla } \\mathbf { v } ^ { 0 } = \\mathbf { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 485, + 310, + 499 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 107, + 505, + 310, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 311, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 311, + 520 + ], + "score": 1.0, + "content": "3.3 Prompt Learning with Optimal Transport", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 103, + 520, + 503, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 519, + 504, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 504, + 532 + ], + "score": 1.0, + "content": "In this subsection, we introduce the details of our PLOT, which learns multiple prompts to describe", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 530, + 393, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 393, + 545 + ], + "score": 1.0, + "content": "different characteristics of the category by minimizing the OT distance.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 86, + 547, + 506, + 652 + ], + "lines": [ + { + "bbox": [ + 86, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 86, + 549, + 100, + 560 + ], + "score": 1.0, + "content": "153", + "type": "text" + }, + { + "bbox": [ + 106, + 547, + 308, + 559 + ], + "score": 1.0, + "content": "Specifically, as shown in Figure 2, given an image", + "type": "text" + }, + { + "bbox": [ + 308, + 550, + 316, + 558 + ], + "score": 0.71, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 547, + 506, + 559 + ], + "score": 1.0, + "content": ", we first feed it to the visual encoder branch of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 86, + 555, + 503, + 576 + ], + "spans": [ + { + "bbox": [ + 86, + 560, + 99, + 570 + ], + "score": 1.0, + "content": "154", + "type": "text" + }, + { + "bbox": [ + 102, + 555, + 280, + 576 + ], + "score": 1.0, + "content": "CLIP. Apart from the global visual feature", + "type": "text" + }, + { + "bbox": [ + 281, + 559, + 288, + 570 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 555, + 458, + 576 + ], + "score": 1.0, + "content": ", we can also obtain a set of local features", + "type": "text" + }, + { + "bbox": [ + 459, + 558, + 503, + 571 + ], + "score": 0.93, + "content": "\\{ f _ { m } | _ { m = 1 } ^ { M } \\}", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 86, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 86, + 572, + 99, + 581 + ], + "score": 1.0, + "content": "155", + "type": "text" + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "The visual encoder has a multi-head attention pooling layer in which the input is the combination of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 86, + 579, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 86, + 582, + 99, + 592 + ], + "score": 1.0, + "content": "156", + "type": "text" + }, + { + "bbox": [ + 105, + 579, + 506, + 593 + ], + "score": 1.0, + "content": "the global feature and a set of local features (feature map) and the output is a tensor with the shape", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 86, + 588, + 507, + 607 + ], + "spans": [ + { + "bbox": [ + 86, + 594, + 100, + 604 + ], + "score": 1.0, + "content": "157", + "type": "text" + }, + { + "bbox": [ + 107, + 591, + 166, + 602 + ], + "score": 0.91, + "content": "\\mathbb { R } ^ { ( \\bar { H } \\times W + 1 ) \\times C }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 588, + 195, + 607 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 195, + 592, + 206, + 602 + ], + "score": 0.81, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 588, + 223, + 607 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 223, + 592, + 235, + 602 + ], + "score": 0.72, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 588, + 399, + 607 + ], + "score": 1.0, + "content": "is the height and width of feature map and", + "type": "text" + }, + { + "bbox": [ + 400, + 592, + 408, + 602 + ], + "score": 0.82, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 588, + 507, + 607 + ], + "score": 1.0, + "content": "is the feature dimension.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 86, + 603, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 86, + 605, + 100, + 615 + ], + "score": 1.0, + "content": "158", + "type": "text" + }, + { + "bbox": [ + 106, + 603, + 211, + 615 + ], + "score": 1.0, + "content": "Therefore, we can obtain", + "type": "text" + }, + { + "bbox": [ + 212, + 604, + 270, + 614 + ], + "score": 0.89, + "content": "M = H \\times W", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 603, + 506, + 615 + ], + "score": 1.0, + "content": "local features and a global feature. At the same time, for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 86, + 611, + 502, + 633 + ], + "spans": [ + { + "bbox": [ + 86, + 618, + 100, + 628 + ], + "score": 1.0, + "content": "159", + "type": "text" + }, + { + "bbox": [ + 103, + 611, + 129, + 633 + ], + "score": 1.0, + "content": "class", + "type": "text" + }, + { + "bbox": [ + 129, + 616, + 136, + 626 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 611, + 209, + 633 + ], + "score": 1.0, + "content": ", we can initialize", + "type": "text" + }, + { + "bbox": [ + 210, + 616, + 219, + 626 + ], + "score": 0.31, + "content": "_ \\mathrm { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 611, + 291, + 633 + ], + "score": 1.0, + "content": "local prompts as", + "type": "text" + }, + { + "bbox": [ + 291, + 615, + 336, + 628 + ], + "score": 0.94, + "content": "\\{ t _ { k , n } | _ { n = 1 } ^ { N } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 611, + 431, + 633 + ], + "score": 1.0, + "content": "with learnable vectors", + "type": "text" + }, + { + "bbox": [ + 432, + 614, + 502, + 630 + ], + "score": 0.94, + "content": "\\{ v e c _ { l , n } | _ { l = 1 , n = 1 } ^ { L , N } \\}", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 86, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 86, + 629, + 100, + 640 + ], + "score": 1.0, + "content": "160", + "type": "text" + }, + { + "bbox": [ + 106, + 628, + 265, + 640 + ], + "score": 1.0, + "content": "where each is the same as the prompt in", + "type": "text" + }, + { + "bbox": [ + 266, + 629, + 290, + 639 + ], + "score": 0.38, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 628, + 505, + 640 + ], + "score": 1.0, + "content": ". With both visual and textual encoders, we can obtain", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 86, + 635, + 491, + 655 + ], + "spans": [ + { + "bbox": [ + 86, + 641, + 99, + 650 + ], + "score": 1.0, + "content": "161", + "type": "text" + }, + { + "bbox": [ + 102, + 635, + 189, + 655 + ], + "score": 1.0, + "content": "local visual features", + "type": "text" + }, + { + "bbox": [ + 189, + 639, + 295, + 651 + ], + "score": 0.91, + "content": "{ \\pmb F } = \\{ { \\pmb f } _ { m } | _ { m = 1 } ^ { M } \\} \\in \\mathbb { R } ^ { M \\times C }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 635, + 379, + 655 + ], + "score": 1.0, + "content": "and prompt features", + "type": "text" + }, + { + "bbox": [ + 380, + 639, + 486, + 651 + ], + "score": 0.91, + "content": "{ G } _ { k } = \\{ \\pmb { g } _ { n } | _ { n = 1 } ^ { N } \\} \\in \\mathbb { R } ^ { N \\times C }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 635, + 491, + 655 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 94, + 655, + 505, + 678 + ], + "lines": [ + { + "bbox": [ + 90, + 653, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 90, + 653, + 286, + 669 + ], + "score": 1.0, + "content": "62 In the inner loop, we learn the transport plan", + "type": "text" + }, + { + "bbox": [ + 287, + 658, + 295, + 665 + ], + "score": 0.86, + "content": "\\mathbf { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 653, + 412, + 669 + ], + "score": 1.0, + "content": "with these fixed support sets", + "type": "text" + }, + { + "bbox": [ + 413, + 655, + 440, + 667 + ], + "score": 0.91, + "content": "{ \\bf \\nabla } F , G _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 653, + 506, + 669 + ], + "score": 1.0, + "content": ", by minimizing", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 90, + 666, + 285, + 678 + ], + "spans": [ + { + "bbox": [ + 90, + 666, + 244, + 678 + ], + "score": 1.0, + "content": "63 the following OT distance to push", + "type": "text" + }, + { + "bbox": [ + 245, + 668, + 259, + 677 + ], + "score": 0.86, + "content": "G _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 666, + 271, + 678 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 271, + 669, + 280, + 676 + ], + "score": 0.86, + "content": "\\pmb { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 666, + 285, + 678 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 680, + 375, + 695 + ], + "lines": [ + { + "bbox": [ + 235, + 680, + 375, + 695 + ], + "spans": [ + { + "bbox": [ + 235, + 680, + 375, + 695 + ], + "score": 0.92, + "content": "d _ { O T } ( k ) = d _ { O T } ( \\boldsymbol { u } , \\boldsymbol { v } | \\mathbf { 1 } - \\mathbf { F } ^ { T } \\mathbf { G } _ { k } ) ,", + "type": "interline_equation", + "image_path": "979aac41062dcd78dbc80f5aef0646091e63fd2eb7cd4a7a67bb45f034353eb9.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 235, + 680, + 375, + 695 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "text", + "bbox": [ + 85, + 699, + 508, + 723 + ], + "lines": [ + { + "bbox": [ + 85, + 699, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 85, + 699, + 132, + 712 + ], + "score": 1.0, + "content": "where 164", + "type": "text" + }, + { + "bbox": [ + 132, + 699, + 201, + 711 + ], + "score": 0.92, + "content": "{ \\pmb { C } } = { \\bf 1 } - { \\pmb { F } } ^ { T } { \\pmb { G } } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 699, + 390, + 712 + ], + "score": 1.0, + "content": "denotes that we use the cosine distance between", + "type": "text" + }, + { + "bbox": [ + 390, + 700, + 400, + 710 + ], + "score": 0.83, + "content": "\\pmb { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 699, + 417, + 712 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 418, + 700, + 432, + 711 + ], + "score": 0.89, + "content": "G _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 699, + 507, + 712 + ], + "score": 1.0, + "content": "as the cost matrix.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 86, + 710, + 498, + 723 + ], + "spans": [ + { + "bbox": [ + 86, + 710, + 304, + 723 + ], + "score": 1.0, + "content": "Then we can obtain the solution of transport plan 165", + "type": "text" + }, + { + "bbox": [ + 305, + 712, + 318, + 721 + ], + "score": 0.85, + "content": "\\mathbf { T } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 710, + 460, + 723 + ], + "score": 1.0, + "content": "as Eq (6) and the final OT distance", + "type": "text" + }, + { + "bbox": [ + 460, + 711, + 494, + 723 + ], + "score": 0.93, + "content": "d o T ( k )", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 710, + 498, + 723 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 72, + 505, + 194 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 72, + 505, + 194 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 194 + ], + "score": 0.97, + "type": "image", + "image_path": "2d85bb34d9ff604845e43487f68a39419971c3db82cc68daca2f31d8e66755e1.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 72, + 505, + 112.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 112.66666666666666, + 505, + 153.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 153.33333333333331, + 505, + 193.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 200, + 505, + 235 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 200, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 506, + 213 + ], + "score": 1.0, + "content": "Figure 2: The framework of PLOT. PLOT first describes each category with multiple prompts and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 212, + 507, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 507, + 225 + ], + "score": 1.0, + "content": "obtains a set of prompt features by text encoder. The image is also encoded as a set of local features.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 222, + 448, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 222, + 448, + 235 + ], + "score": 1.0, + "content": "Then the optimal transport is used as the metric between prompts and visual features.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 87, + 244, + 164, + 255 + ], + "lines": [ + { + "bbox": [ + 85, + 243, + 166, + 255 + ], + "spans": [ + { + "bbox": [ + 85, + 243, + 166, + 255 + ], + "score": 1.0, + "content": "139 are written as:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 85, + 243, + 166, + 255 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 251, + 375, + 286 + ], + "lines": [ + { + "bbox": [ + 236, + 251, + 375, + 286 + ], + "spans": [ + { + "bbox": [ + 236, + 251, + 375, + 286 + ], + "score": 0.95, + "content": "< T , C > = \\sum _ { m = 1 } ^ { M } \\sum _ { n = 1 } ^ { N } T _ { m , n } { \\cal C } _ { m , n } .", + "type": "interline_equation", + "image_path": "73c6d5d67053831369f730face7a896aa4f2dbf1f716ad4663ef50c2533427e6.jpg" + } + ] + } + ], + "index": 7.5, + "virtual_lines": [ + { + "bbox": [ + 236, + 251, + 375, + 268.5 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 236, + 268.5, + 375, + 286.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "index", + "bbox": [ + 86, + 286, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 86, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 86, + 288, + 100, + 298 + ], + "score": 1.0, + "content": "140", + "type": "text" + }, + { + "bbox": [ + 103, + 285, + 141, + 299 + ], + "score": 1.0, + "content": "We call", + "type": "text" + }, + { + "bbox": [ + 142, + 287, + 152, + 297 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 285, + 418, + 299 + ], + "score": 1.0, + "content": "the cost matrix in which each point denotes the cost between", + "type": "text" + }, + { + "bbox": [ + 419, + 287, + 433, + 298 + ], + "score": 0.89, + "content": "f _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 285, + 453, + 299 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 453, + 288, + 466, + 298 + ], + "score": 0.85, + "content": "\\scriptstyle { \\pmb { g } } _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 285, + 506, + 299 + ], + "score": 1.0, + "content": ", such as", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 296, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 85, + 299, + 100, + 309 + ], + "score": 1.0, + "content": "141", + "type": "text" + }, + { + "bbox": [ + 107, + 297, + 212, + 309 + ], + "score": 0.91, + "content": "C _ { m , n } = 1 - \\sin ( { f _ { m } } , { \\pmb { g } } _ { n } )", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 296, + 256, + 311 + ], + "score": 1.0, + "content": ". While the", + "type": "text" + }, + { + "bbox": [ + 256, + 298, + 266, + 307 + ], + "score": 0.81, + "content": "\\mathbf { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 296, + 506, + 311 + ], + "score": 1.0, + "content": "is called the transport plan, which is learned to minimize the", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 308, + 423, + 320 + ], + "spans": [ + { + "bbox": [ + 87, + 311, + 100, + 319 + ], + "score": 1.0, + "content": "142", + "type": "text" + }, + { + "bbox": [ + 105, + 308, + 423, + 320 + ], + "score": 1.0, + "content": "total distance. The optimization problem of optimal transport is formulated as:", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + } + ], + "index": 10, + "bbox_fs": [ + 85, + 285, + 506, + 320 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 221, + 322, + 390, + 358 + ], + "lines": [ + { + "bbox": [ + 221, + 322, + 390, + 358 + ], + "spans": [ + { + "bbox": [ + 221, + 322, + 390, + 358 + ], + "score": 0.67, + "content": "\\begin{array} { l l } { { d _ { O T } ( { \\pmb u } , { \\pmb v } | { \\cal C } ) = \\underset { \\cal T } { \\mathrm { m i n i m i z e } } < { \\cal T } , { \\cal C } > } } \\\\ { { \\mathrm { s u b j e c t ~ t o ~ } } } & { { { \\cal T } { \\bf 1 } = { \\pmb u } , { \\cal T } ^ { T } { \\bf 1 } = { \\pmb v } , { \\cal T } \\geq 0 . } } \\end{array}", + "type": "interline_equation", + "image_path": "492430b29380fae5e25024cb766d92baaf5e617b5cc37bc9abba3343dcfe568e.jpg" + } + ] + } + ], + "index": 12.5, + "virtual_lines": [ + { + "bbox": [ + 221, + 322, + 390, + 340.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 221, + 340.0, + 390, + 358.0 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "index", + "bbox": [ + 86, + 365, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 84, + 365, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 84, + 365, + 506, + 378 + ], + "score": 1.0, + "content": "143 As directly optimizing the above objective is always time-consuming, we apply the Sinkhorn dis-", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 87, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 87, + 379, + 100, + 387 + ], + "score": 1.0, + "content": "144", + "type": "text" + }, + { + "bbox": [ + 104, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "tance [6] to use an entropic constraint for fast optimization. The optimization problem with a", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 388, + 301, + 400 + ], + "spans": [ + { + "bbox": [ + 86, + 390, + 100, + 399 + ], + "score": 1.0, + "content": "145", + "type": "text" + }, + { + "bbox": [ + 104, + 388, + 301, + 400 + ], + "score": 1.0, + "content": "Lagrange multiplier of the entropy constraint is:", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + } + ], + "index": 15, + "bbox_fs": [ + 84, + 365, + 506, + 400 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 401, + 405, + 437 + ], + "lines": [ + { + "bbox": [ + 205, + 401, + 405, + 437 + ], + "spans": [ + { + "bbox": [ + 205, + 401, + 405, + 437 + ], + "score": 0.69, + "content": "\\begin{array} { r l } & { d _ { O T , \\lambda } ( \\pmb { u } , \\pmb { v } | C ) = \\underset { T } { \\mathrm { m i n i m i z e } } < T , C > - \\lambda h ( \\pmb { T } ) } \\\\ & { } \\\\ & { \\mathrm { s u b j e c t ~ t o } \\quad T \\mathbf { 1 } = \\pmb { u } , T ^ { T } \\mathbf { 1 } = \\pmb { v } , } \\end{array}", + "type": "interline_equation", + "image_path": "3c81eb9ce7b52c2923e3d3c8467ed9a16f9ec73e9e8f6ee25037f99bcc4d8a01.jpg" + } + ] + } + ], + "index": 17.5, + "virtual_lines": [ + { + "bbox": [ + 205, + 401, + 405, + 419.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 205, + 419.0, + 405, + 437.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "index", + "bbox": [ + 91, + 439, + 504, + 461 + ], + "lines": [ + { + "bbox": [ + 87, + 439, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 87, + 439, + 132, + 453 + ], + "score": 1.0, + "content": "146 where", + "type": "text" + }, + { + "bbox": [ + 133, + 440, + 150, + 452 + ], + "score": 0.91, + "content": "h ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 439, + 209, + 453 + ], + "score": 1.0, + "content": "is entropy and", + "type": "text" + }, + { + "bbox": [ + 210, + 440, + 235, + 451 + ], + "score": 0.93, + "content": "\\lambda \\geq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 439, + 505, + 453 + ], + "score": 1.0, + "content": "is a hyper-parameter. Then we can have a fast optimization solution", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 88, + 450, + 204, + 463 + ], + "spans": [ + { + "bbox": [ + 88, + 450, + 204, + 463 + ], + "score": 1.0, + "content": "147 with a few iterations as:", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + } + ], + "index": 19.5, + "bbox_fs": [ + 87, + 439, + 505, + 463 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 460, + 379, + 474 + ], + "lines": [ + { + "bbox": [ + 229, + 460, + 379, + 474 + ], + "spans": [ + { + "bbox": [ + 229, + 460, + 379, + 474 + ], + "score": 0.89, + "content": "\\pmb { T } ^ { * } = \\mathrm { d i a g } ( \\pmb { u } ^ { t } ) e x p ( - \\pmb { C } / \\lambda ) \\mathrm { d i a g } ( \\pmb { v } ^ { t } ) ,", + "type": "interline_equation", + "image_path": "c4fffb8dc5cdbb88c0da5ddf3cd3cc92a4317c99e2c970ad2bd06ea091e11e57.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 229, + 460, + 379, + 474 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 101, + 475, + 506, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 504, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 136, + 488 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 136, + 477, + 142, + 485 + ], + "score": 0.71, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 473, + 320, + 488 + ], + "score": 1.0, + "content": "denotes iteration and in each iteration", + "type": "text" + }, + { + "bbox": [ + 321, + 475, + 451, + 487 + ], + "score": 0.89, + "content": "\\begin{array} { c c l } { { { \\pmb u } ^ { t } } } & { { = } } & { { { \\pmb u } / ( ( e x p ( - { \\pmb C } / \\lambda ) { \\pmb v } ^ { t - 1 } ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 473, + 476, + 488 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 476, + 474, + 504, + 487 + ], + "score": 0.87, + "content": "\\begin{array} { r l } { \\boldsymbol { v } ^ { t } } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 485, + 310, + 499 + ], + "spans": [ + { + "bbox": [ + 107, + 486, + 197, + 499 + ], + "score": 0.92, + "content": "\\pmb { v } / ( ( e x p ( - C / \\lambda ) ^ { T } \\pmb { u } ^ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 485, + 274, + 499 + ], + "score": 1.0, + "content": ", with the initiation", + "type": "text" + }, + { + "bbox": [ + 275, + 486, + 305, + 497 + ], + "score": 0.89, + "content": "\\mathbf { \\nabla } \\mathbf { v } ^ { 0 } = \\mathbf { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 485, + 310, + 499 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 473, + 504, + 499 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 505, + 310, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 504, + 311, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 311, + 520 + ], + "score": 1.0, + "content": "3.3 Prompt Learning with Optimal Transport", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 103, + 520, + 503, + 543 + ], + "lines": [ + { + "bbox": [ + 106, + 519, + 504, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 504, + 532 + ], + "score": 1.0, + "content": "In this subsection, we introduce the details of our PLOT, which learns multiple prompts to describe", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 530, + 393, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 393, + 545 + ], + "score": 1.0, + "content": "different characteristics of the category by minimizing the OT distance.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 519, + 504, + 545 + ] + }, + { + "type": "index", + "bbox": [ + 86, + 547, + 506, + 652 + ], + "lines": [ + { + "bbox": [ + 86, + 547, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 86, + 549, + 100, + 560 + ], + "score": 1.0, + "content": "153", + "type": "text" + }, + { + "bbox": [ + 106, + 547, + 308, + 559 + ], + "score": 1.0, + "content": "Specifically, as shown in Figure 2, given an image", + "type": "text" + }, + { + "bbox": [ + 308, + 550, + 316, + 558 + ], + "score": 0.71, + "content": "_ { \\textbf { \\em x } }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 547, + 506, + 559 + ], + "score": 1.0, + "content": ", we first feed it to the visual encoder branch of", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 555, + 503, + 576 + ], + "spans": [ + { + "bbox": [ + 86, + 560, + 99, + 570 + ], + "score": 1.0, + "content": "154", + "type": "text" + }, + { + "bbox": [ + 102, + 555, + 280, + 576 + ], + "score": 1.0, + "content": "CLIP. Apart from the global visual feature", + "type": "text" + }, + { + "bbox": [ + 281, + 559, + 288, + 570 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 555, + 458, + 576 + ], + "score": 1.0, + "content": ", we can also obtain a set of local features", + "type": "text" + }, + { + "bbox": [ + 459, + 558, + 503, + 571 + ], + "score": 0.93, + "content": "\\{ f _ { m } | _ { m = 1 } ^ { M } \\}", + "type": "inline_equation" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 86, + 572, + 99, + 581 + ], + "score": 1.0, + "content": "155", + "type": "text" + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "The visual encoder has a multi-head attention pooling layer in which the input is the combination of", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 579, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 86, + 582, + 99, + 592 + ], + "score": 1.0, + "content": "156", + "type": "text" + }, + { + "bbox": [ + 105, + 579, + 506, + 593 + ], + "score": 1.0, + "content": "the global feature and a set of local features (feature map) and the output is a tensor with the shape", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 588, + 507, + 607 + ], + "spans": [ + { + "bbox": [ + 86, + 594, + 100, + 604 + ], + "score": 1.0, + "content": "157", + "type": "text" + }, + { + "bbox": [ + 107, + 591, + 166, + 602 + ], + "score": 0.91, + "content": "\\mathbb { R } ^ { ( \\bar { H } \\times W + 1 ) \\times C }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 588, + 195, + 607 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 195, + 592, + 206, + 602 + ], + "score": 0.81, + "content": "H", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 588, + 223, + 607 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 223, + 592, + 235, + 602 + ], + "score": 0.72, + "content": "W", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 588, + 399, + 607 + ], + "score": 1.0, + "content": "is the height and width of feature map and", + "type": "text" + }, + { + "bbox": [ + 400, + 592, + 408, + 602 + ], + "score": 0.82, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 588, + 507, + 607 + ], + "score": 1.0, + "content": "is the feature dimension.", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 603, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 86, + 605, + 100, + 615 + ], + "score": 1.0, + "content": "158", + "type": "text" + }, + { + "bbox": [ + 106, + 603, + 211, + 615 + ], + "score": 1.0, + "content": "Therefore, we can obtain", + "type": "text" + }, + { + "bbox": [ + 212, + 604, + 270, + 614 + ], + "score": 0.89, + "content": "M = H \\times W", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 603, + 506, + 615 + ], + "score": 1.0, + "content": "local features and a global feature. At the same time, for", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 611, + 502, + 633 + ], + "spans": [ + { + "bbox": [ + 86, + 618, + 100, + 628 + ], + "score": 1.0, + "content": "159", + "type": "text" + }, + { + "bbox": [ + 103, + 611, + 129, + 633 + ], + "score": 1.0, + "content": "class", + "type": "text" + }, + { + "bbox": [ + 129, + 616, + 136, + 626 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 611, + 209, + 633 + ], + "score": 1.0, + "content": ", we can initialize", + "type": "text" + }, + { + "bbox": [ + 210, + 616, + 219, + 626 + ], + "score": 0.31, + "content": "_ \\mathrm { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 611, + 291, + 633 + ], + "score": 1.0, + "content": "local prompts as", + "type": "text" + }, + { + "bbox": [ + 291, + 615, + 336, + 628 + ], + "score": 0.94, + "content": "\\{ t _ { k , n } | _ { n = 1 } ^ { N } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 611, + 431, + 633 + ], + "score": 1.0, + "content": "with learnable vectors", + "type": "text" + }, + { + "bbox": [ + 432, + 614, + 502, + 630 + ], + "score": 0.94, + "content": "\\{ v e c _ { l , n } | _ { l = 1 , n = 1 } ^ { L , N } \\}", + "type": "inline_equation" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 86, + 629, + 100, + 640 + ], + "score": 1.0, + "content": "160", + "type": "text" + }, + { + "bbox": [ + 106, + 628, + 265, + 640 + ], + "score": 1.0, + "content": "where each is the same as the prompt in", + "type": "text" + }, + { + "bbox": [ + 266, + 629, + 290, + 639 + ], + "score": 0.38, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 628, + 505, + 640 + ], + "score": 1.0, + "content": ". With both visual and textual encoders, we can obtain", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 635, + 491, + 655 + ], + "spans": [ + { + "bbox": [ + 86, + 641, + 99, + 650 + ], + "score": 1.0, + "content": "161", + "type": "text" + }, + { + "bbox": [ + 102, + 635, + 189, + 655 + ], + "score": 1.0, + "content": "local visual features", + "type": "text" + }, + { + "bbox": [ + 189, + 639, + 295, + 651 + ], + "score": 0.91, + "content": "{ \\pmb F } = \\{ { \\pmb f } _ { m } | _ { m = 1 } ^ { M } \\} \\in \\mathbb { R } ^ { M \\times C }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 635, + 379, + 655 + ], + "score": 1.0, + "content": "and prompt features", + "type": "text" + }, + { + "bbox": [ + 380, + 639, + 486, + 651 + ], + "score": 0.91, + "content": "{ G } _ { k } = \\{ \\pmb { g } _ { n } | _ { n = 1 } ^ { N } \\} \\in \\mathbb { R } ^ { N \\times C }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 635, + 491, + 655 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 653, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 90, + 653, + 286, + 669 + ], + "score": 1.0, + "content": "62 In the inner loop, we learn the transport plan", + "type": "text" + }, + { + "bbox": [ + 287, + 658, + 295, + 665 + ], + "score": 0.86, + "content": "\\mathbf { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 653, + 412, + 669 + ], + "score": 1.0, + "content": "with these fixed support sets", + "type": "text" + }, + { + "bbox": [ + 413, + 655, + 440, + 667 + ], + "score": 0.91, + "content": "{ \\bf \\nabla } F , G _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 653, + 506, + 669 + ], + "score": 1.0, + "content": ", by minimizing", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 90, + 666, + 285, + 678 + ], + "spans": [ + { + "bbox": [ + 90, + 666, + 244, + 678 + ], + "score": 1.0, + "content": "63 the following OT distance to push", + "type": "text" + }, + { + "bbox": [ + 245, + 668, + 259, + 677 + ], + "score": 0.86, + "content": "G _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 666, + 271, + 678 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 271, + 669, + 280, + 676 + ], + "score": 0.86, + "content": "\\pmb { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 666, + 285, + 678 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + } + ], + "index": 31, + "bbox_fs": [ + 86, + 547, + 507, + 655 + ] + }, + { + "type": "index", + "bbox": [ + 94, + 655, + 505, + 678 + ], + "lines": [], + "index": 36.5, + "bbox_fs": [ + 90, + 653, + 506, + 678 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 680, + 375, + 695 + ], + "lines": [ + { + "bbox": [ + 235, + 680, + 375, + 695 + ], + "spans": [ + { + "bbox": [ + 235, + 680, + 375, + 695 + ], + "score": 0.92, + "content": "d _ { O T } ( k ) = d _ { O T } ( \\boldsymbol { u } , \\boldsymbol { v } | \\mathbf { 1 } - \\mathbf { F } ^ { T } \\mathbf { G } _ { k } ) ,", + "type": "interline_equation", + "image_path": "979aac41062dcd78dbc80f5aef0646091e63fd2eb7cd4a7a67bb45f034353eb9.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 235, + 680, + 375, + 695 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "list", + "bbox": [ + 85, + 699, + 508, + 723 + ], + "lines": [ + { + "bbox": [ + 85, + 699, + 507, + 712 + ], + "spans": [ + { + "bbox": [ + 85, + 699, + 132, + 712 + ], + "score": 1.0, + "content": "where 164", + "type": "text" + }, + { + "bbox": [ + 132, + 699, + 201, + 711 + ], + "score": 0.92, + "content": "{ \\pmb { C } } = { \\bf 1 } - { \\pmb { F } } ^ { T } { \\pmb { G } } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 699, + 390, + 712 + ], + "score": 1.0, + "content": "denotes that we use the cosine distance between", + "type": "text" + }, + { + "bbox": [ + 390, + 700, + 400, + 710 + ], + "score": 0.83, + "content": "\\pmb { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 699, + 417, + 712 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 418, + 700, + 432, + 711 + ], + "score": 0.89, + "content": "G _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 699, + 507, + 712 + ], + "score": 1.0, + "content": "as the cost matrix.", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 86, + 710, + 498, + 723 + ], + "spans": [ + { + "bbox": [ + 86, + 710, + 304, + 723 + ], + "score": 1.0, + "content": "Then we can obtain the solution of transport plan 165", + "type": "text" + }, + { + "bbox": [ + 305, + 712, + 318, + 721 + ], + "score": 0.85, + "content": "\\mathbf { T } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 710, + 460, + 723 + ], + "score": 1.0, + "content": "as Eq (6) and the final OT distance", + "type": "text" + }, + { + "bbox": [ + 460, + 711, + 494, + 723 + ], + "score": 0.93, + "content": "d o T ( k )", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 710, + 498, + 723 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 39.5, + "bbox_fs": [ + 85, + 699, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 83, + 72, + 460, + 85 + ], + "lines": [ + { + "bbox": [ + 84, + 70, + 462, + 87 + ], + "spans": [ + { + "bbox": [ + 84, + 70, + 234, + 87 + ], + "score": 1.0, + "content": "166 Given the OT distance between", + "type": "text" + }, + { + "bbox": [ + 234, + 73, + 249, + 84 + ], + "score": 0.89, + "content": "G _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 70, + 267, + 87 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 267, + 73, + 276, + 83 + ], + "score": 0.83, + "content": "\\pmb { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 70, + 462, + 87 + ], + "score": 1.0, + "content": ", we reformulate the prediction probability as:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 90, + 403, + 119 + ], + "lines": [ + { + "bbox": [ + 207, + 90, + 403, + 119 + ], + "spans": [ + { + "bbox": [ + 207, + 90, + 403, + 119 + ], + "score": 0.93, + "content": "p _ { o t } ( y = k | x ) = \\frac { e x p ( ( 1 - d _ { O T } ( k ) ) / \\tau ) } { \\sum _ { k ^ { \\prime } = 1 } ^ { K } e x p ( ( 1 - d _ { O T } ( k ^ { \\prime } ) ) / \\tau ) } .", + "type": "interline_equation", + "image_path": "2da49d79f6d0e704321219551d37442169c247852e80bd9c7cf485cb76499532.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 90, + 403, + 104.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 207, + 104.5, + 403, + 119.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 86, + 124, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 84, + 123, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 84, + 123, + 287, + 138 + ], + "score": 1.0, + "content": "In the outer loop, we fix the transport plan 167", + "type": "text" + }, + { + "bbox": [ + 288, + 124, + 301, + 134 + ], + "score": 0.86, + "content": "\\mathbf { T } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 123, + 505, + 138 + ], + "score": 1.0, + "content": "and apply the cross entropy loss to optimize the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 133, + 195, + 153 + ], + "spans": [ + { + "bbox": [ + 82, + 133, + 107, + 153 + ], + "score": 1.0, + "content": "168", + "type": "text" + }, + { + "bbox": [ + 107, + 135, + 178, + 151 + ], + "score": 0.93, + "content": "\\{ v e c _ { l , n } | _ { l = 1 , n = 1 } ^ { L , N } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 133, + 195, + 153 + ], + "score": 1.0, + "content": "as:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 149, + 390, + 184 + ], + "lines": [ + { + "bbox": [ + 220, + 149, + 390, + 184 + ], + "spans": [ + { + "bbox": [ + 220, + 149, + 390, + 184 + ], + "score": 0.95, + "content": "L _ { C E } = - \\frac { 1 } { | \\mathcal { X } | } \\sum _ { \\pmb { x } \\in \\mathcal { X } } \\sum _ { k = 1 } ^ { K } y _ { \\pmb { x } , k } p _ { o t } ( y = k | \\pmb { x } ) ,", + "type": "interline_equation", + "image_path": "861e1e5366c6a6a7ffe47567844bcac19211b7a9747078254574f9e8cb9d827c.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 220, + 149, + 390, + 166.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 220, + 166.5, + 390, + 184.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 94, + 186, + 504, + 198 + ], + "lines": [ + { + "bbox": [ + 90, + 184, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 90, + 184, + 132, + 200 + ], + "score": 1.0, + "content": "69 where", + "type": "text" + }, + { + "bbox": [ + 133, + 188, + 145, + 198 + ], + "score": 0.83, + "content": "\\scriptstyle { \\mathbf { } } _ { \\mathbf { } } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf ", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 184, + 505, + 200 + ], + "score": 1.0, + "content": "is a one-hot label vector. The detail algorithm can be found in the supplementary materials.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 105, + 203, + 505, + 280 + ], + "lines": [ + { + "bbox": [ + 105, + 203, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 215 + ], + "score": 1.0, + "content": "Though the optimization strategy of the optimal transport and prompts is two-stage, the whole", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 506, + 226 + ], + "score": 1.0, + "content": "training flow is end-to-end. It is because that the transport plan is computed using a small number", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 225, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 505, + 236 + ], + "score": 1.0, + "content": "of matrix multiplications as one forward module of the neural network. The gradients of these", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "matrix multiplications are taped for backpropagation for end-to-end optimization, which makes the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "whole system fully differentiable (including the iterative algorithm) and easy to implement using an", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "autograd library like PyTorch. In the experiments, we found that it is natural and relatively easy to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 268, + 212, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 212, + 282 + ], + "score": 1.0, + "content": "this optimization strategy.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 289, + 207, + 300 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 208, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 208, + 303 + ], + "score": 1.0, + "content": "3.4 Inference strategy", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 108, + 304, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "In the inference, given one query image and the learned prompts, we first obtain the a visual feature", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 165, + 326 + ], + "score": 1.0, + "content": "set containing", + "type": "text" + }, + { + "bbox": [ + 166, + 317, + 222, + 325 + ], + "score": 0.92, + "content": "M = H \\times W", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 313, + 401, + 326 + ], + "score": 1.0, + "content": "vectors and a prompt feature set containing", + "type": "text" + }, + { + "bbox": [ + 401, + 317, + 430, + 325 + ], + "score": 0.92, + "content": "N \\times C", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "vectors. Then, we", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 491, + 337 + ], + "score": 1.0, + "content": "calculate the distance between the visual feature set and the prompt feature set of each class by", + "type": "text" + }, + { + "bbox": [ + 491, + 328, + 505, + 335 + ], + "score": 0.41, + "content": "O I ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 335, + 500, + 350 + ], + "spans": [ + { + "bbox": [ + 104, + 335, + 500, + 350 + ], + "score": 1.0, + "content": "as (6). After obtaining the OT distance for each class, we sort the distance and classify the image.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 353, + 191, + 367 + ], + "lines": [ + { + "bbox": [ + 104, + 352, + 193, + 370 + ], + "spans": [ + { + "bbox": [ + 104, + 352, + 193, + 370 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 373, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 387 + ], + "score": 1.0, + "content": "Extensive experiments are conducted to evaluate our method, including comparison with CoOp,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 384, + 502, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 502, + 397 + ], + "score": 1.0, + "content": "ablation studies, parameter analysis extensibility analysis, computing cost analysis and visualization.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 106, + 404, + 166, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 403, + 168, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 168, + 416 + ], + "score": 1.0, + "content": "4.1 Datasets", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 420, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "We followed the experimental settings in the CoOp [63] for the few-shot learning evaluation. The", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "experiments are conducted on the 11 visual recognition datasets, including Caltech101 [9], DTD [5],", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 441, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 506, + 453 + ], + "score": 1.0, + "content": "EuroSAT [12], FGVCAircraft [29], Flowers102 [32], Food101 [3], ImageNet [7], OxfordPets [33],", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "score": 1.0, + "content": "StanfordCars [21], SUN397 [55], and UCF101 [47]. These datasets span visual classification of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 463, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 506, + 475 + ], + "score": 1.0, + "content": "generic objects, scenes, actions, fine-grained categories, and so on, which constitutes a comprehensive", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "evaluation of our method. All experiments adopted the few-shot evaluation protocol used in CLIP [39]", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "and CoOp [63], where we respectively choose 1, 2, 4, 8, and 16 shots for model training and use the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "original test set for evaluation. Besides, we also evaluated the robustness of our method with domain", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 173, + 520 + ], + "score": 1.0, + "content": "shift. Following", + "type": "text" + }, + { + "bbox": [ + 174, + 507, + 199, + 518 + ], + "score": 0.28, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 506, + 506, + 520 + ], + "score": 1.0, + "content": ", we used the ImageNet as the source domain and evaluate our method with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "score": 1.0, + "content": "ImageNet-based robustness evaluation datasets including ImageNetV2 [43], ImageNet-Sketch [52],", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "ImageNet-A [14], and ImageNet-R [13]. A detailed introduction of each dataset can be found in the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 540, + 209, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 209, + 553 + ], + "score": 1.0, + "content": "supplementary materials.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 105, + 559, + 229, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 230, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 230, + 572 + ], + "score": 1.0, + "content": "4.2 Implementation details", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 105, + 574, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "We chose CoOp [63] as our main competitor to evaluate our method. Compared with CoOp which", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "only learns a global prompt for one class, our PLOT method learns multiple local prompts and applies", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "the OT distance for better fine-grained alignment. Besides, we also reported the performance of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 607, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 505, + 620 + ], + "score": 1.0, + "content": "training a linear classifier with the CLIP [39] features. It is also a widely-used strategy to adapt the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "pretrained knowledge for the downstream task [46]. We reproduced the performance of CoOp and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 629, + 320, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 320, + 641 + ], + "score": 1.0, + "content": "the CLIP linear probe with the released official code.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 102, + 645, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 658 + ], + "score": 1.0, + "content": "The original CoOp method has different versions with different class token positions and parameter", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 656, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 669 + ], + "score": 1.0, + "content": "initialization strategies. We applied the default model that fixes the class token positions in the end due", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "to the limited performance gap between two different ways of positioning the class token. Besides, we", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 677, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 692 + ], + "score": 1.0, + "content": "used the random parameter initialization strategy but not the class-specific context version. Following", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "the widely used setting in [63, 64, 10, 58], we also chose RN50 [11] as the backbone network of the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "visual branch and set the length of learnable context tokens as 16. All the code of our method is based", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 119, + 723 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 711, + 145, + 722 + ], + "score": 0.55, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 711, + 505, + 723 + ], + "score": 1.0, + "content": ", which adopted the SGD optimizer with 0.002 initial learning rate, CosineAnnealingLR", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 83, + 72, + 460, + 85 + ], + "lines": [ + { + "bbox": [ + 84, + 70, + 462, + 87 + ], + "spans": [ + { + "bbox": [ + 84, + 70, + 234, + 87 + ], + "score": 1.0, + "content": "166 Given the OT distance between", + "type": "text" + }, + { + "bbox": [ + 234, + 73, + 249, + 84 + ], + "score": 0.89, + "content": "G _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 70, + 267, + 87 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 267, + 73, + 276, + 83 + ], + "score": 0.83, + "content": "\\pmb { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 70, + 462, + 87 + ], + "score": 1.0, + "content": ", we reformulate the prediction probability as:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 84, + 70, + 462, + 87 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 207, + 90, + 403, + 119 + ], + "lines": [ + { + "bbox": [ + 207, + 90, + 403, + 119 + ], + "spans": [ + { + "bbox": [ + 207, + 90, + 403, + 119 + ], + "score": 0.93, + "content": "p _ { o t } ( y = k | x ) = \\frac { e x p ( ( 1 - d _ { O T } ( k ) ) / \\tau ) } { \\sum _ { k ^ { \\prime } = 1 } ^ { K } e x p ( ( 1 - d _ { O T } ( k ^ { \\prime } ) ) / \\tau ) } .", + "type": "interline_equation", + "image_path": "2da49d79f6d0e704321219551d37442169c247852e80bd9c7cf485cb76499532.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 207, + 90, + 403, + 104.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 207, + 104.5, + 403, + 119.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 86, + 124, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 84, + 123, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 84, + 123, + 287, + 138 + ], + "score": 1.0, + "content": "In the outer loop, we fix the transport plan 167", + "type": "text" + }, + { + "bbox": [ + 288, + 124, + 301, + 134 + ], + "score": 0.86, + "content": "\\mathbf { T } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 123, + 505, + 138 + ], + "score": 1.0, + "content": "and apply the cross entropy loss to optimize the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 133, + 195, + 153 + ], + "spans": [ + { + "bbox": [ + 82, + 133, + 107, + 153 + ], + "score": 1.0, + "content": "168", + "type": "text" + }, + { + "bbox": [ + 107, + 135, + 178, + 151 + ], + "score": 0.93, + "content": "\\{ v e c _ { l , n } | _ { l = 1 , n = 1 } ^ { L , N } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 133, + 195, + 153 + ], + "score": 1.0, + "content": "as:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 82, + 123, + 505, + 153 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 149, + 390, + 184 + ], + "lines": [ + { + "bbox": [ + 220, + 149, + 390, + 184 + ], + "spans": [ + { + "bbox": [ + 220, + 149, + 390, + 184 + ], + "score": 0.95, + "content": "L _ { C E } = - \\frac { 1 } { | \\mathcal { X } | } \\sum _ { \\pmb { x } \\in \\mathcal { X } } \\sum _ { k = 1 } ^ { K } y _ { \\pmb { x } , k } p _ { o t } ( y = k | \\pmb { x } ) ,", + "type": "interline_equation", + "image_path": "861e1e5366c6a6a7ffe47567844bcac19211b7a9747078254574f9e8cb9d827c.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 220, + 149, + 390, + 166.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 220, + 166.5, + 390, + 184.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 94, + 186, + 504, + 198 + ], + "lines": [ + { + "bbox": [ + 90, + 184, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 90, + 184, + 132, + 200 + ], + "score": 1.0, + "content": "69 where", + "type": "text" + }, + { + "bbox": [ + 133, + 188, + 145, + 198 + ], + "score": 0.83, + "content": "\\scriptstyle { \\mathbf { } } _ { \\mathbf { } } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf { } \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\Psi \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf \\mathbf ", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 184, + 505, + 200 + ], + "score": 1.0, + "content": "is a one-hot label vector. The detail algorithm can be found in the supplementary materials.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 90, + 184, + 505, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 203, + 505, + 280 + ], + "lines": [ + { + "bbox": [ + 105, + 203, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 215 + ], + "score": 1.0, + "content": "Though the optimization strategy of the optimal transport and prompts is two-stage, the whole", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 506, + 226 + ], + "score": 1.0, + "content": "training flow is end-to-end. It is because that the transport plan is computed using a small number", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 225, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 505, + 236 + ], + "score": 1.0, + "content": "of matrix multiplications as one forward module of the neural network. The gradients of these", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "matrix multiplications are taped for backpropagation for end-to-end optimization, which makes the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "whole system fully differentiable (including the iterative algorithm) and easy to implement using an", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 257, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 505, + 270 + ], + "score": 1.0, + "content": "autograd library like PyTorch. In the experiments, we found that it is natural and relatively easy to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 268, + 212, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 212, + 282 + ], + "score": 1.0, + "content": "this optimization strategy.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 203, + 506, + 282 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 289, + 207, + 300 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 208, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 208, + 303 + ], + "score": 1.0, + "content": "3.4 Inference strategy", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 108, + 304, + 505, + 348 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "In the inference, given one query image and the learned prompts, we first obtain the a visual feature", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 165, + 326 + ], + "score": 1.0, + "content": "set containing", + "type": "text" + }, + { + "bbox": [ + 166, + 317, + 222, + 325 + ], + "score": 0.92, + "content": "M = H \\times W", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 313, + 401, + 326 + ], + "score": 1.0, + "content": "vectors and a prompt feature set containing", + "type": "text" + }, + { + "bbox": [ + 401, + 317, + 430, + 325 + ], + "score": 0.92, + "content": "N \\times C", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "vectors. Then, we", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 325, + 505, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 491, + 337 + ], + "score": 1.0, + "content": "calculate the distance between the visual feature set and the prompt feature set of each class by", + "type": "text" + }, + { + "bbox": [ + 491, + 328, + 505, + 335 + ], + "score": 0.41, + "content": "O I ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 335, + 500, + 350 + ], + "spans": [ + { + "bbox": [ + 104, + 335, + 500, + 350 + ], + "score": 1.0, + "content": "as (6). After obtaining the OT distance for each class, we sort the distance and classify the image.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 302, + 506, + 350 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 353, + 191, + 367 + ], + "lines": [ + { + "bbox": [ + 104, + 352, + 193, + 370 + ], + "spans": [ + { + "bbox": [ + 104, + 352, + 193, + 370 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 373, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 372, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 506, + 387 + ], + "score": 1.0, + "content": "Extensive experiments are conducted to evaluate our method, including comparison with CoOp,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 384, + 502, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 502, + 397 + ], + "score": 1.0, + "content": "ablation studies, parameter analysis extensibility analysis, computing cost analysis and visualization.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 372, + 506, + 397 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 404, + 166, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 403, + 168, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 168, + 416 + ], + "score": 1.0, + "content": "4.1 Datasets", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 420, + 505, + 550 + ], + "lines": [ + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "We followed the experimental settings in the CoOp [63] for the few-shot learning evaluation. The", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "experiments are conducted on the 11 visual recognition datasets, including Caltech101 [9], DTD [5],", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 441, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 506, + 453 + ], + "score": 1.0, + "content": "EuroSAT [12], FGVCAircraft [29], Flowers102 [32], Food101 [3], ImageNet [7], OxfordPets [33],", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "score": 1.0, + "content": "StanfordCars [21], SUN397 [55], and UCF101 [47]. These datasets span visual classification of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 463, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 506, + 475 + ], + "score": 1.0, + "content": "generic objects, scenes, actions, fine-grained categories, and so on, which constitutes a comprehensive", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "evaluation of our method. All experiments adopted the few-shot evaluation protocol used in CLIP [39]", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 497 + ], + "score": 1.0, + "content": "and CoOp [63], where we respectively choose 1, 2, 4, 8, and 16 shots for model training and use the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 508 + ], + "score": 1.0, + "content": "original test set for evaluation. Besides, we also evaluated the robustness of our method with domain", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 173, + 520 + ], + "score": 1.0, + "content": "shift. Following", + "type": "text" + }, + { + "bbox": [ + 174, + 507, + 199, + 518 + ], + "score": 0.28, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 506, + 506, + 520 + ], + "score": 1.0, + "content": ", we used the ImageNet as the source domain and evaluate our method with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 531 + ], + "score": 1.0, + "content": "ImageNet-based robustness evaluation datasets including ImageNetV2 [43], ImageNet-Sketch [52],", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "ImageNet-A [14], and ImageNet-R [13]. A detailed introduction of each dataset can be found in the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 540, + 209, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 209, + 553 + ], + "score": 1.0, + "content": "supplementary materials.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 419, + 506, + 553 + ] + }, + { + "type": "title", + "bbox": [ + 105, + 559, + 229, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 558, + 230, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 230, + 572 + ], + "score": 1.0, + "content": "4.2 Implementation details", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 105, + 574, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "We chose CoOp [63] as our main competitor to evaluate our method. Compared with CoOp which", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "only learns a global prompt for one class, our PLOT method learns multiple local prompts and applies", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "the OT distance for better fine-grained alignment. Besides, we also reported the performance of", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 607, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 505, + 620 + ], + "score": 1.0, + "content": "training a linear classifier with the CLIP [39] features. It is also a widely-used strategy to adapt the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "pretrained knowledge for the downstream task [46]. We reproduced the performance of CoOp and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 629, + 320, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 320, + 641 + ], + "score": 1.0, + "content": "the CLIP linear probe with the released official code.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 574, + 506, + 641 + ] + }, + { + "type": "text", + "bbox": [ + 102, + 645, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 644, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 658 + ], + "score": 1.0, + "content": "The original CoOp method has different versions with different class token positions and parameter", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 656, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 669 + ], + "score": 1.0, + "content": "initialization strategies. We applied the default model that fixes the class token positions in the end due", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "to the limited performance gap between two different ways of positioning the class token. Besides, we", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 677, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 692 + ], + "score": 1.0, + "content": "used the random parameter initialization strategy but not the class-specific context version. Following", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "the widely used setting in [63, 64, 10, 58], we also chose RN50 [11] as the backbone network of the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "visual branch and set the length of learnable context tokens as 16. All the code of our method is based", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 119, + 723 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 119, + 711, + 145, + 722 + ], + "score": 0.55, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 711, + 505, + 723 + ], + "score": 1.0, + "content": ", which adopted the SGD optimizer with 0.002 initial learning rate, CosineAnnealingLR", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 644, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 70, + 464, + 304 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 145, + 70, + 464, + 304 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 145, + 70, + 464, + 304 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 464, + 304 + ], + "score": 0.973, + "type": "image", + "image_path": "b84a44759120ceaf7528f443ecdfaf977c2ef9f4d71ac238fa7ff5b4bd56af7f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 145, + 70, + 464, + 148.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 145, + 148.0, + 464, + 226.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 145, + 226.0, + 464, + 304.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 311, + 506, + 353 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 310, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 464, + 324 + ], + "score": 1.0, + "content": "Figure 3: The few-shot learning results on 11 datasets. We compare our PLOT with CoOp,", + "type": "text" + }, + { + "bbox": [ + 465, + 311, + 500, + 322 + ], + "score": 0.25, + "content": "C o C o O p", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 310, + 505, + 324 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "and the Linear Probe method and observe the consistent and significant performance improvement on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 332, + 419, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 419, + 346 + ], + "score": 1.0, + "content": "most datasets. (The average accuracy on all datasets is shown on the left top.)", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 177, + 342, + 433, + 355 + ], + "spans": [ + { + "bbox": [ + 177, + 342, + 284, + 355 + ], + "score": 1.0, + "content": "Table 1: Comparison with", + "type": "text" + }, + { + "bbox": [ + 285, + 343, + 310, + 354 + ], + "score": 0.63, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 342, + 433, + 355 + ], + "score": 1.0, + "content": "on robustness to domain shift.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "table", + "bbox": [ + 146, + 360, + 462, + 418 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 146, + 360, + 462, + 418 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 360, + 462, + 418 + ], + "spans": [ + { + "bbox": [ + 146, + 360, + 462, + 418 + ], + "score": 0.97, + "html": "
MethodSourceTarget
ImageNet-V2-Sketch-A-R
CLIP +CoOp61.9154.2632.4721.7854.21
CLIP + PLOT (N =4)63.0155.1133.0021.8655.61
", + "type": "table", + "image_path": "ad9a968b12ebbcd690b8921b975defd61d74f50a3f4280bcb20005ccca264689.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 146, + 360, + 462, + 379.3333333333333 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 146, + 379.3333333333333, + 462, + 398.66666666666663 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 146, + 398.66666666666663, + 462, + 417.99999999999994 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 90, + 424, + 507, + 446 + ], + "lines": [ + { + "bbox": [ + 86, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 86, + 423, + 506, + 438 + ], + "score": 1.0, + "content": "212 schedule, and a warmup trick with 1e-5 learning rate. Besides, we also followed the epoch strategy to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 435, + 243, + 447 + ], + "spans": [ + { + "bbox": [ + 86, + 435, + 243, + 447 + ], + "score": 1.0, + "content": "213 train more epochs for more shots.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 105, + 451, + 505, + 517 + ], + "lines": [ + { + "bbox": [ + 106, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 146, + 464 + ], + "score": 1.0, + "content": "We apply", + "type": "text" + }, + { + "bbox": [ + 146, + 452, + 175, + 462 + ], + "score": 0.89, + "content": "N = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 451, + 315, + 464 + ], + "score": 1.0, + "content": "prompts for each category and use", + "type": "text" + }, + { + "bbox": [ + 315, + 452, + 363, + 462 + ], + "score": 0.91, + "content": "M = 7 \\times 7", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "due to the feature map size. We set", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 362, + 475 + ], + "score": 1.0, + "content": "the hyper-parameters in the Sinkhorn distances algorithm [6] as", + "type": "text" + }, + { + "bbox": [ + 363, + 462, + 396, + 473 + ], + "score": 0.91, + "content": "\\lambda = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "for all the datasets. We set", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 474, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 485 + ], + "score": 1.0, + "content": "the maximum iteration number of the inner loop as 100 and will early stop the iteration when the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 230, + 497 + ], + "score": 1.0, + "content": "average absolute update value", + "type": "text" + }, + { + "bbox": [ + 231, + 484, + 270, + 495 + ], + "score": 0.89, + "content": "\\Lambda < 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 484, + 422, + 497 + ], + "score": 1.0, + "content": ". We initialize all values in the vector", + "type": "text" + }, + { + "bbox": [ + 423, + 486, + 429, + 495 + ], + "score": 0.7, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 484, + 447, + 497 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 447, + 486, + 455, + 496 + ], + "score": 0.82, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 484, + 466, + 497 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 467, + 484, + 487, + 496 + ], + "score": 0.85, + "content": "1 / N", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 127, + 507 + ], + "score": 0.84, + "content": "1 / M", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "respectively. All models are conducted on the Pytorch [34] 1.7.1 and trained on 4 NVIDIA", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 504, + 507, + 520 + ], + "spans": [ + { + "bbox": [ + 104, + 504, + 507, + 520 + ], + "score": 1.0, + "content": "A100 GPUs. We repeated the experiments three times with different seeds and reported the average.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 526, + 235, + 538 + ], + "lines": [ + { + "bbox": [ + 104, + 523, + 236, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 523, + 236, + 542 + ], + "score": 1.0, + "content": "4.3 Comparison With CoOp", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 103, + 541, + 503, + 564 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "In this subsection, we compare our PLOT with the baseline CoOp on the few-shot recognition and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 552, + 223, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 223, + 565 + ], + "score": 1.0, + "content": "domain generalization tasks.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 105, + 569, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "score": 1.0, + "content": "Few-Shot Learning We summarized the experimental results in Figure 3 where the red line denotes", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 578, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 593 + ], + "score": 1.0, + "content": "our PLOT method, the blue one denotes CoOp, the purple line denotes CoCoOp, and the green one", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "score": 1.0, + "content": "is the CLIP linear probe. The detailed accuracy can be found in the supplementary materials. We", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 601, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 331, + 614 + ], + "score": 1.0, + "content": "observed that both prompt learning methods (PLOT and", + "type": "text" + }, + { + "bbox": [ + 331, + 602, + 358, + 613 + ], + "score": 0.28, + "content": "\\mathrm { C o O p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 601, + 506, + 614 + ], + "score": 1.0, + "content": ") outperform the linear probe method", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 407, + 625 + ], + "score": 1.0, + "content": "by a large margin. Besides, PLOT can further improve the performance of", + "type": "text" + }, + { + "bbox": [ + 408, + 613, + 434, + 624 + ], + "score": 0.38, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "and CoCoOp on", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 624, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 636 + ], + "score": 1.0, + "content": "most of the datasets. Taking the average accuracy (at the left top) as the example, Plot respectively", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 633, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 104, + 633, + 135, + 648 + ], + "score": 1.0, + "content": "gained", + "type": "text" + }, + { + "bbox": [ + 135, + 634, + 162, + 645 + ], + "score": 0.82, + "content": "3 . 0 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 633, + 165, + 648 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 165, + 634, + 192, + 645 + ], + "score": 0.81, + "content": "3 . 4 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 633, + 195, + 648 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 196, + 634, + 223, + 646 + ], + "score": 0.86, + "content": "2 . 1 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 633, + 226, + 648 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 226, + 634, + 253, + 645 + ], + "score": 0.87, + "content": "1 . 3 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 633, + 256, + 648 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 257, + 634, + 284, + 645 + ], + "score": 0.86, + "content": "\\bar { 0 . 6 1 \\% }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 633, + 379, + 648 + ], + "score": 1.0, + "content": "performance boost over", + "type": "text" + }, + { + "bbox": [ + 379, + 635, + 404, + 646 + ], + "score": 0.35, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 633, + 506, + 648 + ], + "score": 1.0, + "content": "at 1, 2, 4, 8, 16 shots. We", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 644, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 659 + ], + "score": 1.0, + "content": "found the performance gap will reduce when shots increase. It is not surprising since both CoOp", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "and PLOT focus on utilizing the pre-trained knowledge, and the effect of pre-training diminishes", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 465, + 680 + ], + "score": 1.0, + "content": "given more training data. Among all datasets, PLOT achieves a larger improvement over", + "type": "text" + }, + { + "bbox": [ + 466, + 667, + 492, + 678 + ], + "score": 0.41, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "o n", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "the FOOD101 and DTD datasets and achieves comparable performance only on the StanfordCars", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 394, + 701 + ], + "score": 1.0, + "content": "datasets. For the FGVCAircraft dataset in which the CoOp only obtains", + "type": "text" + }, + { + "bbox": [ + 395, + 689, + 422, + 699 + ], + "score": 0.88, + "content": "7 . 7 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "accuracy, our PLOT", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 218, + 713 + ], + "score": 1.0, + "content": "can achieve an accuracy of", + "type": "text" + }, + { + "bbox": [ + 218, + 700, + 250, + 711 + ], + "score": 0.89, + "content": "1 7 . 7 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 700, + 506, + 713 + ], + "score": 1.0, + "content": ", twice as high as that of the CoOp. Note that we don’t use the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 709, + 507, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 507, + 725 + ], + "score": 1.0, + "content": "class-specific context, thus the performance on the fine-grained classification datasets is lower, e.g.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 27.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 70, + 464, + 304 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 145, + 70, + 464, + 304 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 145, + 70, + 464, + 304 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 464, + 304 + ], + "score": 0.973, + "type": "image", + "image_path": "b84a44759120ceaf7528f443ecdfaf977c2ef9f4d71ac238fa7ff5b4bd56af7f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 145, + 70, + 464, + 148.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 145, + 148.0, + 464, + 226.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 145, + 226.0, + 464, + 304.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 311, + 506, + 353 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 310, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 464, + 324 + ], + "score": 1.0, + "content": "Figure 3: The few-shot learning results on 11 datasets. We compare our PLOT with CoOp,", + "type": "text" + }, + { + "bbox": [ + 465, + 311, + 500, + 322 + ], + "score": 0.25, + "content": "C o C o O p", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 310, + 505, + 324 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "and the Linear Probe method and observe the consistent and significant performance improvement on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 332, + 419, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 419, + 346 + ], + "score": 1.0, + "content": "most datasets. (The average accuracy on all datasets is shown on the left top.)", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 177, + 342, + 433, + 355 + ], + "spans": [ + { + "bbox": [ + 177, + 342, + 284, + 355 + ], + "score": 1.0, + "content": "Table 1: Comparison with", + "type": "text" + }, + { + "bbox": [ + 285, + 343, + 310, + 354 + ], + "score": 0.63, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 342, + 433, + 355 + ], + "score": 1.0, + "content": "on robustness to domain shift.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "table", + "bbox": [ + 146, + 360, + 462, + 418 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 146, + 360, + 462, + 418 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 360, + 462, + 418 + ], + "spans": [ + { + "bbox": [ + 146, + 360, + 462, + 418 + ], + "score": 0.97, + "html": "
MethodSourceTarget
ImageNet-V2-Sketch-A-R
CLIP +CoOp61.9154.2632.4721.7854.21
CLIP + PLOT (N =4)63.0155.1133.0021.8655.61
", + "type": "table", + "image_path": "ad9a968b12ebbcd690b8921b975defd61d74f50a3f4280bcb20005ccca264689.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 146, + 360, + 462, + 379.3333333333333 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 146, + 379.3333333333333, + 462, + 398.66666666666663 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 146, + 398.66666666666663, + 462, + 417.99999999999994 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 8 + }, + { + "type": "index", + "bbox": [ + 90, + 424, + 507, + 446 + ], + "lines": [ + { + "bbox": [ + 86, + 423, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 86, + 423, + 506, + 438 + ], + "score": 1.0, + "content": "212 schedule, and a warmup trick with 1e-5 learning rate. Besides, we also followed the epoch strategy to", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 435, + 243, + 447 + ], + "spans": [ + { + "bbox": [ + 86, + 435, + 243, + 447 + ], + "score": 1.0, + "content": "213 train more epochs for more shots.", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + } + ], + "index": 10.5, + "bbox_fs": [ + 86, + 423, + 506, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 451, + 505, + 517 + ], + "lines": [ + { + "bbox": [ + 106, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 146, + 464 + ], + "score": 1.0, + "content": "We apply", + "type": "text" + }, + { + "bbox": [ + 146, + 452, + 175, + 462 + ], + "score": 0.89, + "content": "N = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 451, + 315, + 464 + ], + "score": 1.0, + "content": "prompts for each category and use", + "type": "text" + }, + { + "bbox": [ + 315, + 452, + 363, + 462 + ], + "score": 0.91, + "content": "M = 7 \\times 7", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "due to the feature map size. We set", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 362, + 475 + ], + "score": 1.0, + "content": "the hyper-parameters in the Sinkhorn distances algorithm [6] as", + "type": "text" + }, + { + "bbox": [ + 363, + 462, + 396, + 473 + ], + "score": 0.91, + "content": "\\lambda = 0 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "for all the datasets. We set", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 474, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 485 + ], + "score": 1.0, + "content": "the maximum iteration number of the inner loop as 100 and will early stop the iteration when the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 230, + 497 + ], + "score": 1.0, + "content": "average absolute update value", + "type": "text" + }, + { + "bbox": [ + 231, + 484, + 270, + 495 + ], + "score": 0.89, + "content": "\\Lambda < 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 484, + 422, + 497 + ], + "score": 1.0, + "content": ". We initialize all values in the vector", + "type": "text" + }, + { + "bbox": [ + 423, + 486, + 429, + 495 + ], + "score": 0.7, + "content": "v", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 484, + 447, + 497 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 447, + 486, + 455, + 496 + ], + "score": 0.82, + "content": "\\mu", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 484, + 466, + 497 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 467, + 484, + 487, + 496 + ], + "score": 0.85, + "content": "1 / N", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 127, + 507 + ], + "score": 0.84, + "content": "1 / M", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "respectively. All models are conducted on the Pytorch [34] 1.7.1 and trained on 4 NVIDIA", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 504, + 507, + 520 + ], + "spans": [ + { + "bbox": [ + 104, + 504, + 507, + 520 + ], + "score": 1.0, + "content": "A100 GPUs. We repeated the experiments three times with different seeds and reported the average.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 451, + 507, + 520 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 526, + 235, + 538 + ], + "lines": [ + { + "bbox": [ + 104, + 523, + 236, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 523, + 236, + 542 + ], + "score": 1.0, + "content": "4.3 Comparison With CoOp", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 103, + 541, + 503, + 564 + ], + "lines": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "In this subsection, we compare our PLOT with the baseline CoOp on the few-shot recognition and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 552, + 223, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 223, + 565 + ], + "score": 1.0, + "content": "domain generalization tasks.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 540, + 505, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 569, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "score": 1.0, + "content": "Few-Shot Learning We summarized the experimental results in Figure 3 where the red line denotes", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 578, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 593 + ], + "score": 1.0, + "content": "our PLOT method, the blue one denotes CoOp, the purple line denotes CoCoOp, and the green one", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "score": 1.0, + "content": "is the CLIP linear probe. The detailed accuracy can be found in the supplementary materials. We", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 601, + 506, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 331, + 614 + ], + "score": 1.0, + "content": "observed that both prompt learning methods (PLOT and", + "type": "text" + }, + { + "bbox": [ + 331, + 602, + 358, + 613 + ], + "score": 0.28, + "content": "\\mathrm { C o O p } )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 601, + 506, + 614 + ], + "score": 1.0, + "content": ") outperform the linear probe method", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 407, + 625 + ], + "score": 1.0, + "content": "by a large margin. Besides, PLOT can further improve the performance of", + "type": "text" + }, + { + "bbox": [ + 408, + 613, + 434, + 624 + ], + "score": 0.38, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "and CoCoOp on", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 624, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 636 + ], + "score": 1.0, + "content": "most of the datasets. Taking the average accuracy (at the left top) as the example, Plot respectively", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 633, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 104, + 633, + 135, + 648 + ], + "score": 1.0, + "content": "gained", + "type": "text" + }, + { + "bbox": [ + 135, + 634, + 162, + 645 + ], + "score": 0.82, + "content": "3 . 0 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 633, + 165, + 648 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 165, + 634, + 192, + 645 + ], + "score": 0.81, + "content": "3 . 4 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 633, + 195, + 648 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 196, + 634, + 223, + 646 + ], + "score": 0.86, + "content": "2 . 1 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 633, + 226, + 648 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 226, + 634, + 253, + 645 + ], + "score": 0.87, + "content": "1 . 3 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 633, + 256, + 648 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 257, + 634, + 284, + 645 + ], + "score": 0.86, + "content": "\\bar { 0 . 6 1 \\% }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 633, + 379, + 648 + ], + "score": 1.0, + "content": "performance boost over", + "type": "text" + }, + { + "bbox": [ + 379, + 635, + 404, + 646 + ], + "score": 0.35, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 633, + 506, + 648 + ], + "score": 1.0, + "content": "at 1, 2, 4, 8, 16 shots. We", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 644, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 659 + ], + "score": 1.0, + "content": "found the performance gap will reduce when shots increase. It is not surprising since both CoOp", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "and PLOT focus on utilizing the pre-trained knowledge, and the effect of pre-training diminishes", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 465, + 680 + ], + "score": 1.0, + "content": "given more training data. Among all datasets, PLOT achieves a larger improvement over", + "type": "text" + }, + { + "bbox": [ + 466, + 667, + 492, + 678 + ], + "score": 0.41, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "o n", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 690 + ], + "score": 1.0, + "content": "the FOOD101 and DTD datasets and achieves comparable performance only on the StanfordCars", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 394, + 701 + ], + "score": 1.0, + "content": "datasets. For the FGVCAircraft dataset in which the CoOp only obtains", + "type": "text" + }, + { + "bbox": [ + 395, + 689, + 422, + 699 + ], + "score": 0.88, + "content": "7 . 7 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 422, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "accuracy, our PLOT", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 218, + 713 + ], + "score": 1.0, + "content": "can achieve an accuracy of", + "type": "text" + }, + { + "bbox": [ + 218, + 700, + 250, + 711 + ], + "score": 0.89, + "content": "1 7 . 7 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 700, + 506, + 713 + ], + "score": 1.0, + "content": ", twice as high as that of the CoOp. Note that we don’t use the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 709, + 507, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 507, + 725 + ], + "score": 1.0, + "content": "class-specific context, thus the performance on the fine-grained classification datasets is lower, e.g.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 27.5, + "bbox_fs": [ + 104, + 568, + 507, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 116, + 501, + 329 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 70, + 505, + 114 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 69, + 506, + 83 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 438, + 83 + ], + "score": 1.0, + "content": "Table 2: Ablation studies on few-shot recognition. PLOT is our defined model with", + "type": "text" + }, + { + "bbox": [ + 438, + 73, + 466, + 80 + ], + "score": 0.87, + "content": "N = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 69, + 470, + 83 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 471, + 73, + 494, + 82 + ], + "score": 0.34, + "content": "C o O p", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 69, + 506, + 83 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 81, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 93 + ], + "score": 1.0, + "content": "the baseline method, M denotes that we respectively match the global visual feature and multiple", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 91, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 91, + 506, + 105 + ], + "score": 1.0, + "content": "textual prompts, V denotes that we apply a constraint to add the variance of prompts, M indicates", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 102, + 365, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 365, + 116 + ], + "score": 1.0, + "content": "using the visual feature map instead of the global visual feature.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 109, + 116, + 501, + 329 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 116, + 501, + 329 + ], + "spans": [ + { + "bbox": [ + 109, + 116, + 501, + 329 + ], + "score": 0.977, + "html": "
DatasetSettings1 shot2 shots4 shots8 shots16 shots
Caltech101PLOT CoOp89.83±0.339 87.51 ± 1.0290.67 ±0.21 87.84 ±1.1090.80±0.20 89.52 ±0.8091.54 ± 0.33 90.28±0.4292.24±0.38 91.99 ± 0.31
G88.13 ±0.3686.98 ± 1.2588.45 ± 0.7990.16 ± 0.2290.72 ±0.18
G+V88.28 ± 0.4387.72 ± 1.25
M88.45 ± 0.3089.82 ±0.2092.00 ± 0.13
M+V69.78 ± 1.75 66.11 ± 8.2971.57 ± 1.5977.18 ± 2.1681.77 ± 0.4786.21±0.20
DTDPLOT71.45 ± 3.9879.30 ± 3.9686.96 ± 0.7889.80 ±0.17
46.55 ± 2.6251.24±1.9556.03±0.4361.70 ± 0.3565.60 ± 0.82
CoOp43.62 ±1.9645.35 ± 0.3153.94 ±1.3759.69 ± 0.1362.51±0.25
G45.12 ± 1.6948.39 ± 2.0854.75 ± 0.4860.15 ± 0.7063.59 ±0.76
G+V M45.90 ± 2.0048.50 ± 0.9953.96 ± 0.4859.69 ± 1.0163.51 ± 0.66
M+V13.18 ± 4.57 12.61 ± 5.9312.25 ± 3.86 15.11 ± 1.8113.00 ± 4.73 20.35 ± 1.3320.76 ± 5.4226.99 ±1.98
FOOD101PLOT77.74 ± 0.4744.13 ± 2.3956.85 ± 0.54
CoOp77.70±0.0277.21 ±0.4375.31 ± 0.3077.09 ±0.18
G74.25 ±1.5272.61 ± 1.3373.49 ± 2.0371.58 ± 0.7974.48 ± 0.15
G+V74.63 ± 0.1170.15 ±0.4970.41 ± 0.4670.72 ± 0.9873.68 ±0.46
M74.83 ± 0.3170.09 ± 0.8570.86 ± 0.22 46.86 ±1.3970.80 ± 0.6873.93 ± 0.35
M+V52.02 ± 4.86 46.52 ± 1.1546.12 ±1.46 45.95 ± 2.6653.57 ± 0.8353.43 ± 0.88 62.95 ± 0.3761.28 ± 0.23 67.63 ± 1.11
", + "type": "table", + "image_path": "8e2f3b54076261a0efde38b51ee912af6f609a379b1c2d0c1238c2a4cee06d6e.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 109, + 116, + 501, + 187.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 109, + 187.0, + 501, + 258.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 258.0, + 501, + 329.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 109, + 349, + 501, + 501 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 194, + 336, + 415, + 348 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 192, + 334, + 415, + 349 + ], + "spans": [ + { + "bbox": [ + 192, + 334, + 415, + 349 + ], + "score": 1.0, + "content": "Table 3: Parameter analysis for the number of prompts", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "table_body", + "bbox": [ + 109, + 349, + 501, + 501 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 349, + 501, + 501 + ], + "spans": [ + { + "bbox": [ + 109, + 349, + 501, + 501 + ], + "score": 0.976, + "html": "
DatasetSettings1 shot2 shots4 shots8 shots16 shots
Caltech101N=188.47 ± 1.1589.19 ± 0.3989.70 ±0.3890.45 ± 0.2491.56 ± 0.14
N=288.86 ± 0.5189.60 ±0.1090.60 ± 0.1791.25 ±0.6591.89 ± 0.36
N=489.83 ± 0.3390.67 ± 0.2190.80±0.2091.54 ± 0.3392.24±0.38
N=889.74± 0.3090.18 ± 0.4691.02 ± 0.1891.28 ±0.2892.04 ±0.29
DTDN=143.91 ± 0.6548.21 ± 2.2053.69 ± 1.1058.90 ±0.1962.85 ± 0.74
N=245.59 ± 2.4648.06 ±1.9255.58 ± 1.7161.56± 0.1764.60 ±0.92
N=446.55 ± 2.6251.24 ± 1.9556.03 ± 0.4361.70 ± 0.3565.60 ± 0.82
N=846.89 ±1.9451.87 ± 2.0654.45 ± 0.4862.20 ±0.5665.25± 0.38
FOOD101N=175.96 ± 0.4876.12 ± 0.5977.11 ± 0.4176.56 ± 0.6977.43 ± 0.80
N=277.12 ± 0.4976.89 ± 0.2376.16 ± 0.5275.23 ± 0.6976.81± 0.50
N=477.74±0.4777.70±0.0277.21 ± 0.4375.31 ± 0.3077.09 ±0.18
N=878.05 ± 0.1578.19±0.0778.12 ±0.1776.63 ±0.2277.48 ± 0.12
", + "type": "table", + "image_path": "175945c0adcf1cea5984dc8234b713556bd59f73786a8530a581c7d73fc19c48.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 109, + 349, + 501, + 399.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 109, + 399.6666666666667, + 501, + 450.33333333333337 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 109, + 450.33333333333337, + 501, + 501.00000000000006 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 8.0 + }, + { + "type": "text", + "bbox": [ + 106, + 509, + 505, + 576 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 209, + 522 + ], + "score": 1.0, + "content": "the performance of both", + "type": "text" + }, + { + "bbox": [ + 209, + 510, + 235, + 521 + ], + "score": 0.29, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 509, + 506, + 522 + ], + "score": 1.0, + "content": "and PLOT without class-specific context is lower than the linear", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 519, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 104, + 519, + 506, + 533 + ], + "score": 1.0, + "content": "probing on FGVCAircraft. All these performance comparisons can serve as experimental evidence to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 529, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 104, + 529, + 506, + 545 + ], + "score": 1.0, + "content": "demonstrate that multiple local prompts and optimal transport distance facilitate the prompt learning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "of vision-language models. On StanfordCar, learning multiple prompts didn’t significantly improve", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 552, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 566 + ], + "score": 1.0, + "content": "the performance over a single prompt. It may be because the discriminative characters in this dataset", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 563, + 501, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 501, + 577 + ], + "score": 1.0, + "content": "coincide with each other, such that one global prompt and one global visual feature can work well.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 580, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 592 + ], + "score": 1.0, + "content": "Domain generalization The robustness also plays a critical role in model applications since the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 592, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 505, + 603 + ], + "score": 1.0, + "content": "real-world environment may have large domain shifts with the training data. Therefore, we conducted", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 603, + 445, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 445, + 615 + ], + "score": 1.0, + "content": "a robustness evaluation to investigate the transferability of models learned by PLOT.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 618, + 505, + 674 + ], + "lines": [ + { + "bbox": [ + 106, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "Table 1 summarizes the results of our PLOT method and CoOp on four ImageNet-based robustness", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 630, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 506, + 642 + ], + "score": 1.0, + "content": "evaluation datasets. For both methods, we trained the models on ImageNet with 16 shots per class.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 284, + 654 + ], + "score": 1.0, + "content": "For PLOT, we set the number of prompts as", + "type": "text" + }, + { + "bbox": [ + 284, + 641, + 313, + 650 + ], + "score": 0.89, + "content": "N = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 639, + 505, + 654 + ], + "score": 1.0, + "content": ". We can observe that PLOT outperforms CoOp", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 651, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 505, + 664 + ], + "score": 1.0, + "content": "consistently on both source and target domains. These experimental results demonstrate that the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 662, + 507, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 507, + 676 + ], + "score": 1.0, + "content": "performance improvement of our learning multiple prompts doesn’t rely on single-domain overfitting.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 101, + 684, + 284, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 684, + 284, + 696 + ], + "spans": [ + { + "bbox": [ + 106, + 684, + 284, + 696 + ], + "score": 1.0, + "content": "4.4 Ablation Studies and More Analysis", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 90, + 700, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 86, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 86, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "252 In this subsection, we conducted the ablation studies to investigate the effectiveness of different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 86, + 711, + 333, + 723 + ], + "spans": [ + { + "bbox": [ + 86, + 711, + 333, + 723 + ], + "score": 1.0, + "content": "253 components, in order to answer the following questions.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 84, + 502, + 101, + 734 + ], + "lines": [ + { + "bbox": [ + 85, + 511, + 99, + 522 + ], + "spans": [ + { + "bbox": [ + 85, + 511, + 99, + 522 + ], + "score": 1.0, + "content": "237", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 523, + 100, + 533 + ], + "spans": [ + { + "bbox": [ + 85, + 523, + 100, + 533 + ], + "score": 1.0, + "content": "238", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 533, + 100, + 545 + ], + "spans": [ + { + "bbox": [ + 85, + 533, + 100, + 545 + ], + "score": 1.0, + "content": "239", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 544, + 99, + 556 + ], + "spans": [ + { + "bbox": [ + 85, + 544, + 99, + 556 + ], + "score": 1.0, + "content": "240", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 555, + 99, + 567 + ], + "spans": [ + { + "bbox": [ + 85, + 555, + 99, + 567 + ], + "score": 1.0, + "content": "241", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 566, + 100, + 579 + ], + "spans": [ + { + "bbox": [ + 85, + 566, + 100, + 579 + ], + "score": 1.0, + "content": "242", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 582, + 100, + 594 + ], + "spans": [ + { + "bbox": [ + 85, + 582, + 100, + 594 + ], + "score": 1.0, + "content": "243", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 593, + 100, + 605 + ], + "spans": [ + { + "bbox": [ + 85, + 593, + 100, + 605 + ], + "score": 1.0, + "content": "244", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 604, + 100, + 617 + ], + "spans": [ + { + "bbox": [ + 84, + 604, + 100, + 617 + ], + "score": 1.0, + "content": "245", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 620, + 100, + 632 + ], + "spans": [ + { + "bbox": [ + 85, + 620, + 100, + 632 + ], + "score": 1.0, + "content": "246", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 632, + 100, + 642 + ], + "spans": [ + { + "bbox": [ + 85, + 632, + 100, + 642 + ], + "score": 1.0, + "content": "247", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 642, + 100, + 654 + ], + "spans": [ + { + "bbox": [ + 85, + 642, + 100, + 654 + ], + "score": 1.0, + "content": "248", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 653, + 100, + 665 + ], + "spans": [ + { + "bbox": [ + 86, + 653, + 100, + 665 + ], + "score": 1.0, + "content": "249", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 664, + 100, + 677 + ], + "spans": [ + { + "bbox": [ + 85, + 664, + 100, + 677 + ], + "score": 1.0, + "content": "250", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 685, + 100, + 700 + ], + "spans": [ + { + "bbox": [ + 84, + 685, + 100, + 700 + ], + "score": 1.0, + "content": "251", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 116, + 501, + 329 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 70, + 505, + 114 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 69, + 506, + 83 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 438, + 83 + ], + "score": 1.0, + "content": "Table 2: Ablation studies on few-shot recognition. PLOT is our defined model with", + "type": "text" + }, + { + "bbox": [ + 438, + 73, + 466, + 80 + ], + "score": 0.87, + "content": "N = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 69, + 470, + 83 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 471, + 73, + 494, + 82 + ], + "score": 0.34, + "content": "C o O p", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 69, + 506, + 83 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 81, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 93 + ], + "score": 1.0, + "content": "the baseline method, M denotes that we respectively match the global visual feature and multiple", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 91, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 91, + 506, + 105 + ], + "score": 1.0, + "content": "textual prompts, V denotes that we apply a constraint to add the variance of prompts, M indicates", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 102, + 365, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 365, + 116 + ], + "score": 1.0, + "content": "using the visual feature map instead of the global visual feature.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 109, + 116, + 501, + 329 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 116, + 501, + 329 + ], + "spans": [ + { + "bbox": [ + 109, + 116, + 501, + 329 + ], + "score": 0.977, + "html": "
DatasetSettings1 shot2 shots4 shots8 shots16 shots
Caltech101PLOT CoOp89.83±0.339 87.51 ± 1.0290.67 ±0.21 87.84 ±1.1090.80±0.20 89.52 ±0.8091.54 ± 0.33 90.28±0.4292.24±0.38 91.99 ± 0.31
G88.13 ±0.3686.98 ± 1.2588.45 ± 0.7990.16 ± 0.2290.72 ±0.18
G+V88.28 ± 0.4387.72 ± 1.25
M88.45 ± 0.3089.82 ±0.2092.00 ± 0.13
M+V69.78 ± 1.75 66.11 ± 8.2971.57 ± 1.5977.18 ± 2.1681.77 ± 0.4786.21±0.20
DTDPLOT71.45 ± 3.9879.30 ± 3.9686.96 ± 0.7889.80 ±0.17
46.55 ± 2.6251.24±1.9556.03±0.4361.70 ± 0.3565.60 ± 0.82
CoOp43.62 ±1.9645.35 ± 0.3153.94 ±1.3759.69 ± 0.1362.51±0.25
G45.12 ± 1.6948.39 ± 2.0854.75 ± 0.4860.15 ± 0.7063.59 ±0.76
G+V M45.90 ± 2.0048.50 ± 0.9953.96 ± 0.4859.69 ± 1.0163.51 ± 0.66
M+V13.18 ± 4.57 12.61 ± 5.9312.25 ± 3.86 15.11 ± 1.8113.00 ± 4.73 20.35 ± 1.3320.76 ± 5.4226.99 ±1.98
FOOD101PLOT77.74 ± 0.4744.13 ± 2.3956.85 ± 0.54
CoOp77.70±0.0277.21 ±0.4375.31 ± 0.3077.09 ±0.18
G74.25 ±1.5272.61 ± 1.3373.49 ± 2.0371.58 ± 0.7974.48 ± 0.15
G+V74.63 ± 0.1170.15 ±0.4970.41 ± 0.4670.72 ± 0.9873.68 ±0.46
M74.83 ± 0.3170.09 ± 0.8570.86 ± 0.22 46.86 ±1.3970.80 ± 0.6873.93 ± 0.35
M+V52.02 ± 4.86 46.52 ± 1.1546.12 ±1.46 45.95 ± 2.6653.57 ± 0.8353.43 ± 0.88 62.95 ± 0.3761.28 ± 0.23 67.63 ± 1.11
", + "type": "table", + "image_path": "8e2f3b54076261a0efde38b51ee912af6f609a379b1c2d0c1238c2a4cee06d6e.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 109, + 116, + 501, + 187.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 109, + 187.0, + 501, + 258.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 258.0, + 501, + 329.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 109, + 349, + 501, + 501 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 194, + 336, + 415, + 348 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 192, + 334, + 415, + 349 + ], + "spans": [ + { + "bbox": [ + 192, + 334, + 415, + 349 + ], + "score": 1.0, + "content": "Table 3: Parameter analysis for the number of prompts", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "table_body", + "bbox": [ + 109, + 349, + 501, + 501 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 349, + 501, + 501 + ], + "spans": [ + { + "bbox": [ + 109, + 349, + 501, + 501 + ], + "score": 0.976, + "html": "
DatasetSettings1 shot2 shots4 shots8 shots16 shots
Caltech101N=188.47 ± 1.1589.19 ± 0.3989.70 ±0.3890.45 ± 0.2491.56 ± 0.14
N=288.86 ± 0.5189.60 ±0.1090.60 ± 0.1791.25 ±0.6591.89 ± 0.36
N=489.83 ± 0.3390.67 ± 0.2190.80±0.2091.54 ± 0.3392.24±0.38
N=889.74± 0.3090.18 ± 0.4691.02 ± 0.1891.28 ±0.2892.04 ±0.29
DTDN=143.91 ± 0.6548.21 ± 2.2053.69 ± 1.1058.90 ±0.1962.85 ± 0.74
N=245.59 ± 2.4648.06 ±1.9255.58 ± 1.7161.56± 0.1764.60 ±0.92
N=446.55 ± 2.6251.24 ± 1.9556.03 ± 0.4361.70 ± 0.3565.60 ± 0.82
N=846.89 ±1.9451.87 ± 2.0654.45 ± 0.4862.20 ±0.5665.25± 0.38
FOOD101N=175.96 ± 0.4876.12 ± 0.5977.11 ± 0.4176.56 ± 0.6977.43 ± 0.80
N=277.12 ± 0.4976.89 ± 0.2376.16 ± 0.5275.23 ± 0.6976.81± 0.50
N=477.74±0.4777.70±0.0277.21 ± 0.4375.31 ± 0.3077.09 ±0.18
N=878.05 ± 0.1578.19±0.0778.12 ±0.1776.63 ±0.2277.48 ± 0.12
", + "type": "table", + "image_path": "175945c0adcf1cea5984dc8234b713556bd59f73786a8530a581c7d73fc19c48.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 109, + 349, + 501, + 399.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 109, + 399.6666666666667, + 501, + 450.33333333333337 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 109, + 450.33333333333337, + 501, + 501.00000000000006 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 8.0 + }, + { + "type": "text", + "bbox": [ + 106, + 509, + 505, + 576 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 209, + 522 + ], + "score": 1.0, + "content": "the performance of both", + "type": "text" + }, + { + "bbox": [ + 209, + 510, + 235, + 521 + ], + "score": 0.29, + "content": "\\mathrm { C o O p }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 509, + 506, + 522 + ], + "score": 1.0, + "content": "and PLOT without class-specific context is lower than the linear", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 519, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 104, + 519, + 506, + 533 + ], + "score": 1.0, + "content": "probing on FGVCAircraft. All these performance comparisons can serve as experimental evidence to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 529, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 104, + 529, + 506, + 545 + ], + "score": 1.0, + "content": "demonstrate that multiple local prompts and optimal transport distance facilitate the prompt learning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "of vision-language models. On StanfordCar, learning multiple prompts didn’t significantly improve", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 552, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 506, + 566 + ], + "score": 1.0, + "content": "the performance over a single prompt. It may be because the discriminative characters in this dataset", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 563, + 501, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 501, + 577 + ], + "score": 1.0, + "content": "coincide with each other, such that one global prompt and one global visual feature can work well.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5, + "bbox_fs": [ + 104, + 509, + 506, + 577 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 580, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 592 + ], + "score": 1.0, + "content": "Domain generalization The robustness also plays a critical role in model applications since the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 592, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 505, + 603 + ], + "score": 1.0, + "content": "real-world environment may have large domain shifts with the training data. Therefore, we conducted", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 603, + 445, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 445, + 615 + ], + "score": 1.0, + "content": "a robustness evaluation to investigate the transferability of models learned by PLOT.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 581, + 505, + 615 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 618, + 505, + 674 + ], + "lines": [ + { + "bbox": [ + 106, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "Table 1 summarizes the results of our PLOT method and CoOp on four ImageNet-based robustness", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 630, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 506, + 642 + ], + "score": 1.0, + "content": "evaluation datasets. For both methods, we trained the models on ImageNet with 16 shots per class.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 284, + 654 + ], + "score": 1.0, + "content": "For PLOT, we set the number of prompts as", + "type": "text" + }, + { + "bbox": [ + 284, + 641, + 313, + 650 + ], + "score": 0.89, + "content": "N = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 639, + 505, + 654 + ], + "score": 1.0, + "content": ". We can observe that PLOT outperforms CoOp", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 651, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 505, + 664 + ], + "score": 1.0, + "content": "consistently on both source and target domains. These experimental results demonstrate that the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 662, + 507, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 507, + 676 + ], + "score": 1.0, + "content": "performance improvement of our learning multiple prompts doesn’t rely on single-domain overfitting.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 618, + 507, + 676 + ] + }, + { + "type": "title", + "bbox": [ + 101, + 684, + 284, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 684, + 284, + 696 + ], + "spans": [ + { + "bbox": [ + 106, + 684, + 284, + 696 + ], + "score": 1.0, + "content": "4.4 Ablation Studies and More Analysis", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "index", + "bbox": [ + 90, + 700, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 86, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 86, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "252 In this subsection, we conducted the ablation studies to investigate the effectiveness of different", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 711, + 333, + 723 + ], + "spans": [ + { + "bbox": [ + 86, + 711, + 333, + 723 + ], + "score": 1.0, + "content": "253 components, in order to answer the following questions.", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + } + ], + "index": 26.5, + "bbox_fs": [ + 86, + 699, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 87, + 73, + 505, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 86 + ], + "score": 1.0, + "content": "Q: Can we directly learn multiple prompts by respectively matching each prompt with the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 82, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 82, + 506, + 97 + ], + "score": 1.0, + "content": "global visual feature? A: No. As shown in Table 2, we report the performance of directly matching", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "the global visual feature (notated as “G”) and compare it with the baseline CoOp and our PLOT on", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "score": 1.0, + "content": "three datasets including Caltech101, DTD, and FOOD101. We observe that there is no improvement", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "over the baseline on some datasets (such as Caltech101 and FOOD101) if we only directly match", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "prompts and global features. Though “G” obtained the improvement on the DTD dataset, this", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "improvement is still less than that of PLOT. It is because this “G” method is incentivized to learn the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 148, + 507, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 507, + 163 + ], + "score": 1.0, + "content": "indistinguishable prompts, which contradicts our purpose to learn multiple comprehensive prompts.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "We further add some constraints to push away the prompt from each other. For example, we add an", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "objective function to add the distance between every two prompts as a regularization term, which", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 304, + 195 + ], + "score": 1.0, + "content": "is notated as “V”. However, comparing “G” and", + "type": "text" + }, + { + "bbox": [ + 304, + 182, + 334, + 192 + ], + "score": 0.83, + "content": "\\mathrm { ^ { 6 6 } G + V ^ { 5 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 181, + 506, + 195 + ], + "score": 1.0, + "content": ", we do not find significant and consistent", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 193, + 266, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 266, + 205 + ], + "score": 1.0, + "content": "improvement when using variance loss.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 104, + 209, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "Q: Does the improvement mainly come from using all feature maps? A: No. In PLOT, we apply", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "all feature maps of the visual encoder branch, where each feature is a local embedding at one spatial", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "position. Compared with the global feature, these local features are more informative and contain", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "fine-grained clues. However, we demonstrate that the improvement of PLOT does not only rely on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 252, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 266 + ], + "score": 1.0, + "content": "using all feature maps. On the contrary, directly using the feature map to replace the global feature", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 262, + 506, + 278 + ], + "spans": [ + { + "bbox": [ + 104, + 262, + 506, + 278 + ], + "score": 1.0, + "content": "causes a large performance drop. For example, on all three datasets, directly using the feature map", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 273, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 141, + 288 + ], + "score": 1.0, + "content": "(“M” or", + "type": "text" + }, + { + "bbox": [ + 142, + 275, + 175, + 285 + ], + "score": 0.79, + "content": "\\mathbf { \\hat { \\mu } ^ { 6 } M + V } ^ { 5 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 273, + 238, + 288 + ], + "score": 1.0, + "content": "has an around", + "type": "text" + }, + { + "bbox": [ + 238, + 274, + 258, + 285 + ], + "score": 0.84, + "content": "2 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 273, + 506, + 288 + ], + "score": 1.0, + "content": "1 shot accuracy drop over using the global visual feature. It", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "is not surprising since the original CLIP model is trained by matching the global visual feature and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "language feature. Without using the OT method, the distance between the feature map and multiple", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "textual prompts degenerates to the mean distance of each feature-prompt pair. Besides, when using", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "the feature map, adding the variance loss works well, especially for more shots. For example, the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 329, + 428, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 428, + 342 + ], + "score": 1.0, + "content": "accuracy on 16 shots DTD is improved by a large margin (from 26.99 to 56.85).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "Q: How many prompts are needed? A: 4 prompts are enough One important hyper-parameter", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 356, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 368 + ], + "score": 1.0, + "content": "in PLOT is the number of prompts. To analyze the effect of the number of prompts, we conducted", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "the experiments on three datasets with 1, 2, 4, 8 prompts. The results are summarized in the white", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 377, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 104, + 377, + 506, + 391 + ], + "score": 1.0, + "content": "part of Table 3. We can observe that the performance obviously increases when adding the number", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 289, + 402 + ], + "score": 1.0, + "content": "of prompts from 1 to 4. For example, PLOT", + "type": "text" + }, + { + "bbox": [ + 289, + 390, + 312, + 400 + ], + "score": 0.54, + "content": "\\left( \\mathrm { N } { = } 4 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 389, + 397, + 402 + ], + "score": 1.0, + "content": "respectively obtains", + "type": "text" + }, + { + "bbox": [ + 398, + 389, + 425, + 400 + ], + "score": 0.84, + "content": "1 . 3 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 389, + 428, + 402 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 429, + 389, + 456, + 400 + ], + "score": 0.82, + "content": "2 . 6 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 389, + 477, + 402 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 477, + 389, + 505, + 400 + ], + "score": 0.87, + "content": "1 . 6 8 \\%", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 399, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 272, + 414 + ], + "score": 1.0, + "content": "1-shot accuracy improvement over PLOT", + "type": "text" + }, + { + "bbox": [ + 272, + 401, + 293, + 411 + ], + "score": 0.74, + "content": "\\left( \\mathrm { N } { = } 1 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 399, + 506, + 414 + ], + "score": 1.0, + "content": ") on three datasets. Besides, when we further increase", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "score": 1.0, + "content": "the number of prompts, the improvement is not consistent. To balance the improvement and cost,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 135, + 435 + ], + "score": 1.0, + "content": "we set", + "type": "text" + }, + { + "bbox": [ + 135, + 422, + 165, + 432 + ], + "score": 0.9, + "content": "N = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 422, + 506, + 435 + ], + "score": 1.0, + "content": "as the default configuration of our PLOT model. In the experiments, we tuned this", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 432, + 407, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 407, + 446 + ], + "score": 1.0, + "content": "hyper-parameter on the Caltech101 dataset and applied it to other datasets.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 449, + 505, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 447, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 462 + ], + "score": 1.0, + "content": "Q: Can PLOT benefit zero-shot learning? A: No. CLIP [39] shows that manually designing the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "prompts can still achieve good performance. We obtain 7 prompts by prompt engineering on the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 470, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 344, + 483 + ], + "score": 1.0, + "content": "ImageNet dataset and can further ensemble them to obtain", + "type": "text" + }, + { + "bbox": [ + 344, + 471, + 380, + 482 + ], + "score": 0.88, + "content": "{ \\bf 6 0 . 3 8 \\% }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 470, + 506, + 483 + ], + "score": 1.0, + "content": "top 1 accuracy. In this section,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 482, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 493 + ], + "score": 1.0, + "content": "we replace the cosine distance between the global visual feature and prompt ensemble with the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "score": 1.0, + "content": "OT distance between the feature map and all 7 prompts. However, without any learning, the OT", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 195, + 516 + ], + "score": 1.0, + "content": "distance only obtains", + "type": "text" + }, + { + "bbox": [ + 196, + 504, + 232, + 515 + ], + "score": 0.88, + "content": "{ \\bf 5 8 . 7 8 \\% }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "accuracy. It is a limitation of the PLOT to still need few-shot data", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 515, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 526 + ], + "score": 1.0, + "content": "for optimization, which cannot be directly applied in the zero-shot setting. We argue there are two", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "reasons why the OT distance does not work without learning: 1) prompt engineering selects prompts", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "based on the global feature and cosine distance, instead of OT distance with feature map; 2) all these", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 548, + 424, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 424, + 559 + ], + "score": 1.0, + "content": "selected prompts are closed to the global feature and lack the complementarity.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 101, + 564, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "Q: Can PLOT benefit Adapter-based methods? A: Yes. Adapter-based methods [10, 58] is another", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "research direction of the efficient adaptation of pre-trained vision-language models. Different from", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "the prompt learning that fixes the model parameters and tunes the language prompt, adapter-based", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 595, + 507, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 507, + 610 + ], + "score": 1.0, + "content": "methods [10, 58] allow for fine-tuning a part of the network or adding an extra model for training.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 606, + 507, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 507, + 620 + ], + "score": 1.0, + "content": "Recently, adapter-based methods also achieve good performance on few-shot visual recognition.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 618, + 447, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 447, + 630 + ], + "score": 1.0, + "content": "Therefore, we want to explore whether our PLOT method can benefit them, and how.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + }, + { + "type": "text", + "bbox": [ + 102, + 635, + 505, + 711 + ], + "lines": [ + { + "bbox": [ + 105, + 634, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 397, + 648 + ], + "score": 1.0, + "content": "We apply the Tip-adapter-F [58] as our baseline method, which learns a", + "type": "text" + }, + { + "bbox": [ + 398, + 635, + 505, + 646 + ], + "score": 0.8, + "content": "L i n e a r ( d , N _ { c l s } \\times K _ { s h o t s } )", + "type": "inline_equation" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 419, + 658 + ], + "score": 1.0, + "content": "model to describe one image by the similarity with all training samples, where", + "type": "text" + }, + { + "bbox": [ + 419, + 646, + 426, + 655 + ], + "score": 0.82, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "is the dimension of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 165, + 669 + ], + "score": 1.0, + "content": "visual feature,", + "type": "text" + }, + { + "bbox": [ + 166, + 657, + 185, + 667 + ], + "score": 0.91, + "content": "N _ { c l s }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 656, + 417, + 669 + ], + "score": 1.0, + "content": "is the number of categories (e.g. 1000 in ImageNet), and", + "type": "text" + }, + { + "bbox": [ + 417, + 657, + 447, + 667 + ], + "score": 0.92, + "content": "K _ { s h o t s }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "is the number", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 668, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 679 + ], + "score": 1.0, + "content": "of shots. Then, the final similarity consists of the original distance between the visual feature and", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "prompt ensembling and the new distance calculated by the learned feature and one-hot vector of", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 222, + 702 + ], + "score": 1.0, + "content": "labels (whose dimension is", + "type": "text" + }, + { + "bbox": [ + 223, + 689, + 313, + 701 + ], + "score": 0.9, + "content": "( N _ { c l s } \\times K _ { s h o t s } , N _ { c l s } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 688, + 506, + 702 + ], + "score": 1.0, + "content": ". Please find details in Tip-adapter-F [58]. To", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "introduce PLOT to this framework, we first used the feature map to replace the global feature and", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 52 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 85, + 64, + 100, + 734 + ], + "lines": [ + { + "bbox": [ + 85, + 75, + 100, + 86 + ], + "spans": [ + { + "bbox": [ + 85, + 75, + 100, + 86 + ], + "score": 1.0, + "content": "254", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 85, + 101, + 97 + ], + "spans": [ + { + "bbox": [ + 85, + 85, + 101, + 97 + ], + "score": 1.0, + "content": "255", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 96, + 101, + 108 + ], + "spans": [ + { + "bbox": [ + 85, + 96, + 101, + 108 + ], + "score": 1.0, + "content": "256", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 107, + 101, + 118 + ], + "spans": [ + { + "bbox": [ + 85, + 107, + 101, + 118 + ], + "score": 1.0, + "content": "257", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 118, + 101, + 130 + ], + "spans": [ + { + "bbox": [ + 85, + 118, + 101, + 130 + ], + "score": 1.0, + "content": "258", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 129, + 101, + 140 + ], + "spans": [ + { + "bbox": [ + 85, + 129, + 101, + 140 + ], + "score": 1.0, + "content": "259", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 139, + 101, + 151 + ], + "spans": [ + { + "bbox": [ + 85, + 139, + 101, + 151 + ], + "score": 1.0, + "content": "260", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 150, + 100, + 162 + ], + "spans": [ + { + "bbox": [ + 85, + 150, + 100, + 162 + ], + "score": 1.0, + "content": "261", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 162, + 101, + 173 + ], + "spans": [ + { + "bbox": [ + 85, + 162, + 101, + 173 + ], + "score": 1.0, + "content": "262", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 172, + 101, + 184 + ], + "spans": [ + { + "bbox": [ + 85, + 172, + 101, + 184 + ], + "score": 1.0, + "content": "263", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 183, + 101, + 195 + ], + "spans": [ + { + "bbox": [ + 85, + 183, + 101, + 195 + ], + "score": 1.0, + "content": "264", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 194, + 101, + 208 + ], + "spans": [ + { + "bbox": [ + 84, + 194, + 101, + 208 + ], + "score": 1.0, + "content": "265", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 210, + 101, + 222 + ], + "spans": [ + { + "bbox": [ + 85, + 210, + 101, + 222 + ], + "score": 1.0, + "content": "266", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 222, + 100, + 234 + ], + "spans": [ + { + "bbox": [ + 85, + 222, + 100, + 234 + ], + "score": 1.0, + "content": "267", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 233, + 101, + 244 + ], + "spans": [ + { + "bbox": [ + 85, + 233, + 101, + 244 + ], + "score": 1.0, + "content": "268", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 244, + 101, + 255 + ], + "spans": [ + { + "bbox": [ + 85, + 244, + 101, + 255 + ], + "score": 1.0, + "content": "269", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 255, + 101, + 266 + ], + "spans": [ + { + "bbox": [ + 85, + 255, + 101, + 266 + ], + "score": 1.0, + "content": "270", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 266, + 100, + 277 + ], + "spans": [ + { + "bbox": [ + 85, + 266, + 100, + 277 + ], + "score": 1.0, + "content": "271", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 277, + 100, + 288 + ], + "spans": [ + { + "bbox": [ + 85, + 277, + 100, + 288 + ], + "score": 1.0, + "content": "272", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 287, + 101, + 299 + ], + "spans": [ + { + "bbox": [ + 85, + 287, + 101, + 299 + ], + "score": 1.0, + "content": "273", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 298, + 101, + 310 + ], + "spans": [ + { + "bbox": [ + 85, + 298, + 101, + 310 + ], + "score": 1.0, + "content": "274", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 309, + 101, + 321 + ], + "spans": [ + { + "bbox": [ + 85, + 309, + 101, + 321 + ], + "score": 1.0, + "content": "275", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 320, + 101, + 331 + ], + "spans": [ + { + "bbox": [ + 85, + 320, + 101, + 331 + ], + "score": 1.0, + "content": "276", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 330, + 100, + 343 + ], + "spans": [ + { + "bbox": [ + 85, + 330, + 100, + 343 + ], + "score": 1.0, + "content": "277", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 344, + 102, + 360 + ], + "spans": [ + { + "bbox": [ + 84, + 344, + 102, + 360 + ], + "score": 1.0, + "content": "278", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 358, + 101, + 369 + ], + "spans": [ + { + "bbox": [ + 85, + 358, + 101, + 369 + ], + "score": 1.0, + "content": "279", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 369, + 101, + 380 + ], + "spans": [ + { + "bbox": [ + 85, + 369, + 101, + 380 + ], + "score": 1.0, + "content": "280", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 380, + 100, + 392 + ], + "spans": [ + { + "bbox": [ + 85, + 380, + 100, + 392 + ], + "score": 1.0, + "content": "281", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 391, + 101, + 402 + ], + "spans": [ + { + "bbox": [ + 85, + 391, + 101, + 402 + ], + "score": 1.0, + "content": "282", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 402, + 101, + 414 + ], + "spans": [ + { + "bbox": [ + 85, + 402, + 101, + 414 + ], + "score": 1.0, + "content": "283", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 413, + 101, + 424 + ], + "spans": [ + { + "bbox": [ + 85, + 413, + 101, + 424 + ], + "score": 1.0, + "content": "284", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 424, + 101, + 435 + ], + "spans": [ + { + "bbox": [ + 85, + 424, + 101, + 435 + ], + "score": 1.0, + "content": "285", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 435, + 101, + 447 + ], + "spans": [ + { + "bbox": [ + 85, + 435, + 101, + 447 + ], + "score": 1.0, + "content": "286", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 451, + 100, + 462 + ], + "spans": [ + { + "bbox": [ + 85, + 451, + 100, + 462 + ], + "score": 1.0, + "content": "287", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 462, + 101, + 474 + ], + "spans": [ + { + "bbox": [ + 85, + 462, + 101, + 474 + ], + "score": 1.0, + "content": "288", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 472, + 101, + 484 + ], + "spans": [ + { + "bbox": [ + 85, + 472, + 101, + 484 + ], + "score": 1.0, + "content": "289", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 483, + 101, + 495 + ], + "spans": [ + { + "bbox": [ + 85, + 483, + 101, + 495 + ], + "score": 1.0, + "content": "290", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 495, + 100, + 506 + ], + "spans": [ + { + "bbox": [ + 85, + 495, + 100, + 506 + ], + "score": 1.0, + "content": "291", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 505, + 101, + 517 + ], + "spans": [ + { + "bbox": [ + 85, + 505, + 101, + 517 + ], + "score": 1.0, + "content": "292", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 516, + 101, + 527 + ], + "spans": [ + { + "bbox": [ + 85, + 516, + 101, + 527 + ], + "score": 1.0, + "content": "293", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 527, + 101, + 538 + ], + "spans": [ + { + "bbox": [ + 85, + 527, + 101, + 538 + ], + "score": 1.0, + "content": "294", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 538, + 101, + 550 + ], + "spans": [ + { + "bbox": [ + 85, + 538, + 101, + 550 + ], + "score": 1.0, + "content": "295", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 550, + 101, + 561 + ], + "spans": [ + { + "bbox": [ + 85, + 550, + 101, + 561 + ], + "score": 1.0, + "content": "296", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 565, + 100, + 577 + ], + "spans": [ + { + "bbox": [ + 85, + 565, + 100, + 577 + ], + "score": 1.0, + "content": "297", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 577, + 101, + 588 + ], + "spans": [ + { + "bbox": [ + 85, + 577, + 101, + 588 + ], + "score": 1.0, + "content": "298", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 587, + 101, + 599 + ], + "spans": [ + { + "bbox": [ + 85, + 587, + 101, + 599 + ], + "score": 1.0, + "content": "299", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 598, + 101, + 610 + ], + "spans": [ + { + "bbox": [ + 85, + 598, + 101, + 610 + ], + "score": 1.0, + "content": "300", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 609, + 100, + 621 + ], + "spans": [ + { + "bbox": [ + 85, + 609, + 100, + 621 + ], + "score": 1.0, + "content": "301", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 619, + 101, + 633 + ], + "spans": [ + { + "bbox": [ + 84, + 619, + 101, + 633 + ], + "score": 1.0, + "content": "302", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 637, + 101, + 649 + ], + "spans": [ + { + "bbox": [ + 85, + 637, + 101, + 649 + ], + "score": 1.0, + "content": "303", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 648, + 101, + 659 + ], + "spans": [ + { + "bbox": [ + 85, + 648, + 101, + 659 + ], + "score": 1.0, + "content": "304", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 658, + 101, + 669 + ], + "spans": [ + { + "bbox": [ + 85, + 658, + 101, + 669 + ], + "score": 1.0, + "content": "305", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 669, + 101, + 680 + ], + "spans": [ + { + "bbox": [ + 85, + 669, + 101, + 680 + ], + "score": 1.0, + "content": "306", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 680, + 100, + 691 + ], + "spans": [ + { + "bbox": [ + 85, + 680, + 100, + 691 + ], + "score": 1.0, + "content": "307", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 689, + 101, + 703 + ], + "spans": [ + { + "bbox": [ + 85, + 689, + 101, + 703 + ], + "score": 1.0, + "content": "308", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 702, + 101, + 713 + ], + "spans": [ + { + "bbox": [ + 85, + 702, + 101, + 713 + ], + "score": 1.0, + "content": "309", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 87, + 73, + 505, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 86 + ], + "score": 1.0, + "content": "Q: Can we directly learn multiple prompts by respectively matching each prompt with the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 82, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 82, + 506, + 97 + ], + "score": 1.0, + "content": "global visual feature? A: No. As shown in Table 2, we report the performance of directly matching", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "the global visual feature (notated as “G”) and compare it with the baseline CoOp and our PLOT on", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 506, + 118 + ], + "score": 1.0, + "content": "three datasets including Caltech101, DTD, and FOOD101. We observe that there is no improvement", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 115, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 104, + 115, + 506, + 129 + ], + "score": 1.0, + "content": "over the baseline on some datasets (such as Caltech101 and FOOD101) if we only directly match", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "prompts and global features. Though “G” obtained the improvement on the DTD dataset, this", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "improvement is still less than that of PLOT. It is because this “G” method is incentivized to learn the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 148, + 507, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 507, + 163 + ], + "score": 1.0, + "content": "indistinguishable prompts, which contradicts our purpose to learn multiple comprehensive prompts.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "We further add some constraints to push away the prompt from each other. For example, we add an", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "objective function to add the distance between every two prompts as a regularization term, which", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 181, + 506, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 304, + 195 + ], + "score": 1.0, + "content": "is notated as “V”. However, comparing “G” and", + "type": "text" + }, + { + "bbox": [ + 304, + 182, + 334, + 192 + ], + "score": 0.83, + "content": "\\mathrm { ^ { 6 6 } G + V ^ { 5 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 181, + 506, + 195 + ], + "score": 1.0, + "content": ", we do not find significant and consistent", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 193, + 266, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 266, + 205 + ], + "score": 1.0, + "content": "improvement when using variance loss.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 5.5, + "bbox_fs": [ + 104, + 73, + 507, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 209, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "Q: Does the improvement mainly come from using all feature maps? A: No. In PLOT, we apply", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "all feature maps of the visual encoder branch, where each feature is a local embedding at one spatial", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "position. Compared with the global feature, these local features are more informative and contain", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 506, + 255 + ], + "score": 1.0, + "content": "fine-grained clues. However, we demonstrate that the improvement of PLOT does not only rely on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 252, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 505, + 266 + ], + "score": 1.0, + "content": "using all feature maps. On the contrary, directly using the feature map to replace the global feature", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 262, + 506, + 278 + ], + "spans": [ + { + "bbox": [ + 104, + 262, + 506, + 278 + ], + "score": 1.0, + "content": "causes a large performance drop. For example, on all three datasets, directly using the feature map", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 273, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 273, + 141, + 288 + ], + "score": 1.0, + "content": "(“M” or", + "type": "text" + }, + { + "bbox": [ + 142, + 275, + 175, + 285 + ], + "score": 0.79, + "content": "\\mathbf { \\hat { \\mu } ^ { 6 } M + V } ^ { 5 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 273, + 238, + 288 + ], + "score": 1.0, + "content": "has an around", + "type": "text" + }, + { + "bbox": [ + 238, + 274, + 258, + 285 + ], + "score": 0.84, + "content": "2 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 273, + 506, + 288 + ], + "score": 1.0, + "content": "1 shot accuracy drop over using the global visual feature. It", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "is not surprising since the original CLIP model is trained by matching the global visual feature and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "language feature. Without using the OT method, the distance between the feature map and multiple", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 306, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 321 + ], + "score": 1.0, + "content": "textual prompts degenerates to the mean distance of each feature-prompt pair. Besides, when using", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 106, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "the feature map, adding the variance loss works well, especially for more shots. For example, the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 329, + 428, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 428, + 342 + ], + "score": 1.0, + "content": "accuracy on 16 shots DTD is improved by a large margin (from 26.99 to 56.85).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 209, + 506, + 342 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 444 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "Q: How many prompts are needed? A: 4 prompts are enough One important hyper-parameter", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 356, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 368 + ], + "score": 1.0, + "content": "in PLOT is the number of prompts. To analyze the effect of the number of prompts, we conducted", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "the experiments on three datasets with 1, 2, 4, 8 prompts. The results are summarized in the white", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 377, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 104, + 377, + 506, + 391 + ], + "score": 1.0, + "content": "part of Table 3. We can observe that the performance obviously increases when adding the number", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 289, + 402 + ], + "score": 1.0, + "content": "of prompts from 1 to 4. For example, PLOT", + "type": "text" + }, + { + "bbox": [ + 289, + 390, + 312, + 400 + ], + "score": 0.54, + "content": "\\left( \\mathrm { N } { = } 4 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 389, + 397, + 402 + ], + "score": 1.0, + "content": "respectively obtains", + "type": "text" + }, + { + "bbox": [ + 398, + 389, + 425, + 400 + ], + "score": 0.84, + "content": "1 . 3 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 389, + 428, + 402 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 429, + 389, + 456, + 400 + ], + "score": 0.82, + "content": "2 . 6 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 389, + 477, + 402 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 477, + 389, + 505, + 400 + ], + "score": 0.87, + "content": "1 . 6 8 \\%", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 399, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 272, + 414 + ], + "score": 1.0, + "content": "1-shot accuracy improvement over PLOT", + "type": "text" + }, + { + "bbox": [ + 272, + 401, + 293, + 411 + ], + "score": 0.74, + "content": "\\left( \\mathrm { N } { = } 1 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 399, + 506, + 414 + ], + "score": 1.0, + "content": ") on three datasets. Besides, when we further increase", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "score": 1.0, + "content": "the number of prompts, the improvement is not consistent. To balance the improvement and cost,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 135, + 435 + ], + "score": 1.0, + "content": "we set", + "type": "text" + }, + { + "bbox": [ + 135, + 422, + 165, + 432 + ], + "score": 0.9, + "content": "N = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 422, + 506, + 435 + ], + "score": 1.0, + "content": "as the default configuration of our PLOT model. In the experiments, we tuned this", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 432, + 407, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 407, + 446 + ], + "score": 1.0, + "content": "hyper-parameter on the Caltech101 dataset and applied it to other datasets.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 345, + 506, + 446 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 449, + 505, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 447, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 462 + ], + "score": 1.0, + "content": "Q: Can PLOT benefit zero-shot learning? A: No. CLIP [39] shows that manually designing the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 505, + 473 + ], + "score": 1.0, + "content": "prompts can still achieve good performance. We obtain 7 prompts by prompt engineering on the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 470, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 344, + 483 + ], + "score": 1.0, + "content": "ImageNet dataset and can further ensemble them to obtain", + "type": "text" + }, + { + "bbox": [ + 344, + 471, + 380, + 482 + ], + "score": 0.88, + "content": "{ \\bf 6 0 . 3 8 \\% }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 470, + 506, + 483 + ], + "score": 1.0, + "content": "top 1 accuracy. In this section,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 482, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 505, + 493 + ], + "score": 1.0, + "content": "we replace the cosine distance between the global visual feature and prompt ensemble with the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 506, + 505 + ], + "score": 1.0, + "content": "OT distance between the feature map and all 7 prompts. However, without any learning, the OT", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 502, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 195, + 516 + ], + "score": 1.0, + "content": "distance only obtains", + "type": "text" + }, + { + "bbox": [ + 196, + 504, + 232, + 515 + ], + "score": 0.88, + "content": "{ \\bf 5 8 . 7 8 \\% }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 502, + 505, + 516 + ], + "score": 1.0, + "content": "accuracy. It is a limitation of the PLOT to still need few-shot data", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 515, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 526 + ], + "score": 1.0, + "content": "for optimization, which cannot be directly applied in the zero-shot setting. We argue there are two", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "reasons why the OT distance does not work without learning: 1) prompt engineering selects prompts", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "based on the global feature and cosine distance, instead of OT distance with feature map; 2) all these", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 548, + 424, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 424, + 559 + ], + "score": 1.0, + "content": "selected prompts are closed to the global feature and lack the complementarity.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 447, + 506, + 559 + ] + }, + { + "type": "text", + "bbox": [ + 101, + 564, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "Q: Can PLOT benefit Adapter-based methods? A: Yes. Adapter-based methods [10, 58] is another", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "research direction of the efficient adaptation of pre-trained vision-language models. Different from", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "the prompt learning that fixes the model parameters and tunes the language prompt, adapter-based", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 595, + 507, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 507, + 610 + ], + "score": 1.0, + "content": "methods [10, 58] allow for fine-tuning a part of the network or adding an extra model for training.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 606, + 507, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 507, + 620 + ], + "score": 1.0, + "content": "Recently, adapter-based methods also achieve good performance on few-shot visual recognition.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 618, + 447, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 447, + 630 + ], + "score": 1.0, + "content": "Therefore, we want to explore whether our PLOT method can benefit them, and how.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 563, + 507, + 630 + ] + }, + { + "type": "text", + "bbox": [ + 102, + 635, + 505, + 711 + ], + "lines": [ + { + "bbox": [ + 105, + 634, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 397, + 648 + ], + "score": 1.0, + "content": "We apply the Tip-adapter-F [58] as our baseline method, which learns a", + "type": "text" + }, + { + "bbox": [ + 398, + 635, + 505, + 646 + ], + "score": 0.8, + "content": "L i n e a r ( d , N _ { c l s } \\times K _ { s h o t s } )", + "type": "inline_equation" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 419, + 658 + ], + "score": 1.0, + "content": "model to describe one image by the similarity with all training samples, where", + "type": "text" + }, + { + "bbox": [ + 419, + 646, + 426, + 655 + ], + "score": 0.82, + "content": "d", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "is the dimension of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 656, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 165, + 669 + ], + "score": 1.0, + "content": "visual feature,", + "type": "text" + }, + { + "bbox": [ + 166, + 657, + 185, + 667 + ], + "score": 0.91, + "content": "N _ { c l s }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 656, + 417, + 669 + ], + "score": 1.0, + "content": "is the number of categories (e.g. 1000 in ImageNet), and", + "type": "text" + }, + { + "bbox": [ + 417, + 657, + 447, + 667 + ], + "score": 0.92, + "content": "K _ { s h o t s }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 656, + 506, + 669 + ], + "score": 1.0, + "content": "is the number", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 668, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 679 + ], + "score": 1.0, + "content": "of shots. Then, the final similarity consists of the original distance between the visual feature and", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "prompt ensembling and the new distance calculated by the learned feature and one-hot vector of", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 222, + 702 + ], + "score": 1.0, + "content": "labels (whose dimension is", + "type": "text" + }, + { + "bbox": [ + 223, + 689, + 313, + 701 + ], + "score": 0.9, + "content": "( N _ { c l s } \\times K _ { s h o t s } , N _ { c l s } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 688, + 506, + 702 + ], + "score": 1.0, + "content": ". Please find details in Tip-adapter-F [58]. To", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "introduce PLOT to this framework, we first used the feature map to replace the global feature and", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 52, + "bbox_fs": [ + 105, + 634, + 506, + 712 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 69, + 459, + 260 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 151, + 69, + 459, + 260 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 151, + 69, + 459, + 260 + ], + "spans": [ + { + "bbox": [ + 151, + 69, + 459, + 260 + ], + "score": 0.975, + "type": "image", + "image_path": "08d9b7cffd3a033b270221f27d732b125692e3a6c258dfa9ae21f128c885ad2a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 151, + 69, + 459, + 132.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 151, + 132.66666666666666, + 459, + 196.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 151, + 196.33333333333331, + 459, + 260.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 266, + 504, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 265, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 381, + 279 + ], + "score": 1.0, + "content": "Figure 4: Visualizations. We provide the heatmaps of transport plan", + "type": "text" + }, + { + "bbox": [ + 381, + 266, + 391, + 276 + ], + "score": 0.3, + "content": "_ { \\mathbf { \\delta T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 265, + 506, + 279 + ], + "score": 1.0, + "content": "related to each prompt on 4", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 276, + 473, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 473, + 290 + ], + "score": 1.0, + "content": "categories in ImageNet. Different transport plans focus on different attributes of the object.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 117, + 303, + 492, + 355 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 205, + 290, + 405, + 301 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 204, + 289, + 406, + 302 + ], + "spans": [ + { + "bbox": [ + 204, + 289, + 406, + 302 + ], + "score": 1.0, + "content": "Table 4: Comparison with Adapter-based method.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 117, + 303, + 492, + 355 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 303, + 492, + 355 + ], + "spans": [ + { + "bbox": [ + 117, + 303, + 492, + 355 + ], + "score": 0.971, + "html": "
DatasetMethods1 shot2 shots4 shots8 shots16 shots
ImageNetTip-Adapter-F61.3261.6962.5264.0065.51
Tip-Adapter-F+ OT61.4461.9862.8664.1365.76
Tip-Adapter-F +PLOT62.2764.3163.8965.0466.17
", + "type": "table", + "image_path": "de24cecf6f2b11aa55bbb44a1ece49407beb0e3e244dcc2dd699e21b180385f7.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 117, + 303, + 492, + 320.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 117, + 320.3333333333333, + 492, + 337.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 117, + 337.66666666666663, + 492, + 354.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 88, + 360, + 504, + 404 + ], + "lines": [ + { + "bbox": [ + 85, + 359, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 85, + 361, + 100, + 372 + ], + "score": 1.0, + "content": "310", + "type": "text" + }, + { + "bbox": [ + 105, + 359, + 506, + 372 + ], + "score": 1.0, + "content": "then learned multiple linear models. As a result, with different local features and different linear", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 86, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 86, + 372, + 99, + 382 + ], + "score": 1.0, + "content": "311", + "type": "text" + }, + { + "bbox": [ + 105, + 370, + 205, + 383 + ], + "score": 1.0, + "content": "models, we can obtain a", + "type": "text" + }, + { + "bbox": [ + 206, + 371, + 239, + 381 + ], + "score": 0.92, + "content": "M \\times N", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "distance matrix and apply the Sinkhorn algorithm [6] to calculate", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 380, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 85, + 383, + 100, + 394 + ], + "score": 1.0, + "content": "312", + "type": "text" + }, + { + "bbox": [ + 105, + 380, + 505, + 396 + ], + "score": 1.0, + "content": "the OT distance. Furthermore, we can apply the learned prompts as co-partner of the ensembling", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 85, + 392, + 251, + 406 + ], + "spans": [ + { + "bbox": [ + 85, + 394, + 100, + 404 + ], + "score": 1.0, + "content": "313", + "type": "text" + }, + { + "bbox": [ + 104, + 392, + 251, + 406 + ], + "score": 1.0, + "content": "prompt to refine the final similarity.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 409, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "Table 4 summarizes the few-shot recognition results of the original Tip-Adapter-F method and our", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "adapter-based PLOT methods on ImageNet. From this table, We observe that using the OT distance", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 431, + 504, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 504, + 443 + ], + "score": 1.0, + "content": "can improve the performance of the adapter-based method. Using the learned prompts, we can further", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 442, + 253, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 253, + 455 + ], + "score": 1.0, + "content": "promote the accuracy of all settings.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 104, + 458, + 504, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 483, + 470 + ], + "score": 1.0, + "content": "Q: What is the extra computation time cost of PLOT over CoOp baseline? A: Around", + "type": "text" + }, + { + "bbox": [ + 483, + 457, + 505, + 469 + ], + "score": 0.85, + "content": "\\mathbf { 1 0 \\% }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 469, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 190, + 482 + ], + "score": 1.0, + "content": "inference speed and", + "type": "text" + }, + { + "bbox": [ + 190, + 469, + 205, + 480 + ], + "score": 0.85, + "content": "{ \\bf 5 \\% }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 469, + 506, + 482 + ], + "score": 1.0, + "content": "training time. Despite the performance improvement, the extra computation", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 480, + 495, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 495, + 492 + ], + "score": 1.0, + "content": "cost is still a limitation of PLOT. Please see the detailed analysis in the supplementary materials.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 106, + 504, + 186, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 187, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 187, + 516 + ], + "score": 1.0, + "content": "4.5 Visualization", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 106, + 521, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 420, + 532 + ], + "score": 1.0, + "content": "In this subsection, we provide some visualization examples of the transport plans", + "type": "text" + }, + { + "bbox": [ + 421, + 521, + 430, + 531 + ], + "score": 0.7, + "content": "_ { \\mathbf { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 521, + 505, + 532 + ], + "score": 1.0, + "content": "related to different", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 532, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 143, + 544 + ], + "score": 1.0, + "content": "prompts", + "type": "text" + }, + { + "bbox": [ + 143, + 532, + 167, + 542 + ], + "score": 0.28, + "content": "\\left( \\mathrm { N } { = } 4 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 532, + 506, + 544 + ], + "score": 1.0, + "content": "). We translate each transport plan into colorful heatmaps and resize them into their", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 541, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 104, + 541, + 506, + 556 + ], + "score": 1.0, + "content": "original size and combine them with the raw image. As shown in Figure 4, we provide the heatmaps", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 553, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 506, + 565 + ], + "score": 1.0, + "content": "of 4 categories in ImageNet. We observe that different transport plans highlight different regions of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "the image, which demonstrates that the learned multiple prompts are complementary. For the class", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 574, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 104, + 574, + 506, + 589 + ], + "score": 1.0, + "content": "“Brambling”, the prompts respectively focus on the head, tail, wing, and environment. For “Dog", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 585, + 450, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 450, + 599 + ], + "score": 1.0, + "content": "Sled”, the prompts are related to dogs, the sled, some ties, and the snow environment.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 105, + 612, + 183, + 625 + ], + "lines": [ + { + "bbox": [ + 104, + 609, + 185, + 628 + ], + "spans": [ + { + "bbox": [ + 104, + 609, + 185, + 628 + ], + "score": 1.0, + "content": "5 Conclusion", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 648 + ], + "score": 1.0, + "content": "In this paper, we present a method, named PLOT, to learn multiple comprehensive prompts to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "describe diverse characteristics of one category. To avoid convergence to one point, we propose to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 655, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 670 + ], + "score": 1.0, + "content": "apply the optimal transport to achieve the fine-grained alignment between both vision and language", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "domains. We apply a two-stage optimization strategy where the inner loop fixes the prompts and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "learns the transport plan to calculate the cross-modality distance, and the outer loop uses this distance", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "to optimize the prompt learner. We build our method on the base of CoOp and achieve significant", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "improvement on the few-shot recognition task in various datasets, which demonstrates the advantage", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 711, + 304, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 304, + 723 + ], + "score": 1.0, + "content": "to learn multiple prompts instead of a single one.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 69, + 459, + 260 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 151, + 69, + 459, + 260 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 151, + 69, + 459, + 260 + ], + "spans": [ + { + "bbox": [ + 151, + 69, + 459, + 260 + ], + "score": 0.975, + "type": "image", + "image_path": "08d9b7cffd3a033b270221f27d732b125692e3a6c258dfa9ae21f128c885ad2a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 151, + 69, + 459, + 132.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 151, + 132.66666666666666, + 459, + 196.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 151, + 196.33333333333331, + 459, + 260.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 266, + 504, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 265, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 381, + 279 + ], + "score": 1.0, + "content": "Figure 4: Visualizations. We provide the heatmaps of transport plan", + "type": "text" + }, + { + "bbox": [ + 381, + 266, + 391, + 276 + ], + "score": 0.3, + "content": "_ { \\mathbf { \\delta T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 265, + 506, + 279 + ], + "score": 1.0, + "content": "related to each prompt on 4", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 276, + 473, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 473, + 290 + ], + "score": 1.0, + "content": "categories in ImageNet. Different transport plans focus on different attributes of the object.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 117, + 303, + 492, + 355 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 205, + 290, + 405, + 301 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 204, + 289, + 406, + 302 + ], + "spans": [ + { + "bbox": [ + 204, + 289, + 406, + 302 + ], + "score": 1.0, + "content": "Table 4: Comparison with Adapter-based method.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 117, + 303, + 492, + 355 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 303, + 492, + 355 + ], + "spans": [ + { + "bbox": [ + 117, + 303, + 492, + 355 + ], + "score": 0.971, + "html": "
DatasetMethods1 shot2 shots4 shots8 shots16 shots
ImageNetTip-Adapter-F61.3261.6962.5264.0065.51
Tip-Adapter-F+ OT61.4461.9862.8664.1365.76
Tip-Adapter-F +PLOT62.2764.3163.8965.0466.17
", + "type": "table", + "image_path": "de24cecf6f2b11aa55bbb44a1ece49407beb0e3e244dcc2dd699e21b180385f7.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 117, + 303, + 492, + 320.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 117, + 320.3333333333333, + 492, + 337.66666666666663 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 117, + 337.66666666666663, + 492, + 354.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "index", + "bbox": [ + 88, + 360, + 504, + 404 + ], + "lines": [ + { + "bbox": [ + 85, + 359, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 85, + 361, + 100, + 372 + ], + "score": 1.0, + "content": "310", + "type": "text" + }, + { + "bbox": [ + 105, + 359, + 506, + 372 + ], + "score": 1.0, + "content": "then learned multiple linear models. As a result, with different local features and different linear", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 86, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 86, + 372, + 99, + 382 + ], + "score": 1.0, + "content": "311", + "type": "text" + }, + { + "bbox": [ + 105, + 370, + 205, + 383 + ], + "score": 1.0, + "content": "models, we can obtain a", + "type": "text" + }, + { + "bbox": [ + 206, + 371, + 239, + 381 + ], + "score": 0.92, + "content": "M \\times N", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "distance matrix and apply the Sinkhorn algorithm [6] to calculate", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 380, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 85, + 383, + 100, + 394 + ], + "score": 1.0, + "content": "312", + "type": "text" + }, + { + "bbox": [ + 105, + 380, + 505, + 396 + ], + "score": 1.0, + "content": "the OT distance. Furthermore, we can apply the learned prompts as co-partner of the ensembling", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 85, + 392, + 251, + 406 + ], + "spans": [ + { + "bbox": [ + 85, + 394, + 100, + 404 + ], + "score": 1.0, + "content": "313", + "type": "text" + }, + { + "bbox": [ + 104, + 392, + 251, + 406 + ], + "score": 1.0, + "content": "prompt to refine the final similarity.", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + } + ], + "index": 10.5, + "bbox_fs": [ + 85, + 359, + 506, + 406 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 409, + 505, + 453 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "Table 4 summarizes the few-shot recognition results of the original Tip-Adapter-F method and our", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "adapter-based PLOT methods on ImageNet. From this table, We observe that using the OT distance", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 431, + 504, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 504, + 443 + ], + "score": 1.0, + "content": "can improve the performance of the adapter-based method. Using the learned prompts, we can further", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 442, + 253, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 442, + 253, + 455 + ], + "score": 1.0, + "content": "promote the accuracy of all settings.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 408, + 505, + 455 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 458, + 504, + 492 + ], + "lines": [ + { + "bbox": [ + 106, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 483, + 470 + ], + "score": 1.0, + "content": "Q: What is the extra computation time cost of PLOT over CoOp baseline? A: Around", + "type": "text" + }, + { + "bbox": [ + 483, + 457, + 505, + 469 + ], + "score": 0.85, + "content": "\\mathbf { 1 0 \\% }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 469, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 190, + 482 + ], + "score": 1.0, + "content": "inference speed and", + "type": "text" + }, + { + "bbox": [ + 190, + 469, + 205, + 480 + ], + "score": 0.85, + "content": "{ \\bf 5 \\% }", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 469, + 506, + 482 + ], + "score": 1.0, + "content": "training time. Despite the performance improvement, the extra computation", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 480, + 495, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 495, + 492 + ], + "score": 1.0, + "content": "cost is still a limitation of PLOT. Please see the detailed analysis in the supplementary materials.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 457, + 506, + 492 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 504, + 186, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 187, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 187, + 516 + ], + "score": 1.0, + "content": "4.5 Visualization", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 520, + 505, + 597 + ], + "lines": [ + { + "bbox": [ + 106, + 521, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 521, + 420, + 532 + ], + "score": 1.0, + "content": "In this subsection, we provide some visualization examples of the transport plans", + "type": "text" + }, + { + "bbox": [ + 421, + 521, + 430, + 531 + ], + "score": 0.7, + "content": "_ { \\mathbf { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 521, + 505, + 532 + ], + "score": 1.0, + "content": "related to different", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 532, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 143, + 544 + ], + "score": 1.0, + "content": "prompts", + "type": "text" + }, + { + "bbox": [ + 143, + 532, + 167, + 542 + ], + "score": 0.28, + "content": "\\left( \\mathrm { N } { = } 4 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 532, + 506, + 544 + ], + "score": 1.0, + "content": "). We translate each transport plan into colorful heatmaps and resize them into their", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 541, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 104, + 541, + 506, + 556 + ], + "score": 1.0, + "content": "original size and combine them with the raw image. As shown in Figure 4, we provide the heatmaps", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 553, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 506, + 565 + ], + "score": 1.0, + "content": "of 4 categories in ImageNet. We observe that different transport plans highlight different regions of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 104, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "the image, which demonstrates that the learned multiple prompts are complementary. For the class", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 574, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 104, + 574, + 506, + 589 + ], + "score": 1.0, + "content": "“Brambling”, the prompts respectively focus on the head, tail, wing, and environment. For “Dog", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 585, + 450, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 450, + 599 + ], + "score": 1.0, + "content": "Sled”, the prompts are related to dogs, the sled, some ties, and the snow environment.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24, + "bbox_fs": [ + 104, + 521, + 506, + 599 + ] + }, + { + "type": "title", + "bbox": [ + 105, + 612, + 183, + 625 + ], + "lines": [ + { + "bbox": [ + 104, + 609, + 185, + 628 + ], + "spans": [ + { + "bbox": [ + 104, + 609, + 185, + 628 + ], + "score": 1.0, + "content": "5 Conclusion", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 648 + ], + "score": 1.0, + "content": "In this paper, we present a method, named PLOT, to learn multiple comprehensive prompts to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "describe diverse characteristics of one category. To avoid convergence to one point, we propose to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 655, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 670 + ], + "score": 1.0, + "content": "apply the optimal transport to achieve the fine-grained alignment between both vision and language", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "domains. We apply a two-stage optimization strategy where the inner loop fixes the prompts and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "learns the transport plan to calculate the cross-modality distance, and the outer loop uses this distance", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "to optimize the prompt learner. We build our method on the base of CoOp and achieve significant", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "improvement on the few-shot recognition task in various datasets, which demonstrates the advantage", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 711, + 304, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 304, + 723 + ], + "score": 1.0, + "content": "to learn multiple prompts instead of a single one.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 633, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 89, + 72, + 164, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 165, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 165, + 87 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 101, + 86, + 508, + 723 + ], + "lines": [ + { + "bbox": [ + 110, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 110, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 125, + 101, + 234, + 115 + ], + "spans": [ + { + "bbox": [ + 125, + 101, + 234, + 115 + ], + "score": 1.0, + "content": "ICML, pages 214–223, 2017.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 117, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 110, + 117, + 505, + 128 + ], + "score": 1.0, + "content": "[2] Emmanuel Boissard, Thibaut Le Gouic, and Jean-Michel Loubes. Distribution’s template estimate with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 127, + 321, + 138 + ], + "spans": [ + { + "bbox": [ + 126, + 127, + 321, + 138 + ], + "score": 1.0, + "content": "wasserstein metrics. Bernoulli, 21(2):740–759, 2015.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 141, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 110, + 141, + 505, + 153 + ], + "score": 1.0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative components", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 151, + 323, + 165 + ], + "spans": [ + { + "bbox": [ + 125, + 151, + 323, + 165 + ], + "score": 1.0, + "content": "with random forests. In ECCV, pages 446–461, 2014.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 110, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "[4] Liqun Chen, Yizhe Zhang, Ruiyi Zhang, Chenyang Tao, Zhe Gan, Haichao Zhang, Bai Li, Dinghan Shen,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 125, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "Changyou Chen, and Lawrence Carin. Improving sequence-to-sequence learning via optimal transport.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 124, + 185, + 273, + 198 + ], + "spans": [ + { + "bbox": [ + 124, + 185, + 273, + 198 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1901.06283, 2019.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 198, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 108, + 198, + 506, + 214 + ], + "score": 1.0, + "content": "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 124, + 209, + 328, + 222 + ], + "spans": [ + { + "bbox": [ + 124, + 209, + 328, + 222 + ], + "score": 1.0, + "content": "textures in the wild. In CVPR, pages 3606–3613, 2014.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 110, + 224, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 110, + 224, + 506, + 237 + ], + "score": 1.0, + "content": "[6] Marco Cuturi. Sinkhorn distances: lightspeed computation of optimal transport. In NeurIPS, volume 2,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 124, + 236, + 178, + 246 + ], + "spans": [ + { + "bbox": [ + 124, + 236, + 178, + 246 + ], + "score": 1.0, + "content": "page 4, 2013.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 109, + 249, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 109, + 249, + 506, + 262 + ], + "score": 1.0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 124, + 259, + 304, + 272 + ], + "spans": [ + { + "bbox": [ + 124, + 259, + 304, + 272 + ], + "score": 1.0, + "content": "image database. In CVPR, pages 248–255, 2009.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 108, + 272, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 108, + 272, + 506, + 288 + ], + "score": 1.0, + "content": "[8] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Zicheng", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 284, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 125, + 284, + 506, + 297 + ], + "score": 1.0, + "content": "Liu, Michael Zeng, et al. An empirical study of training end-to-end vision-and-language transformers.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 124, + 294, + 274, + 306 + ], + "spans": [ + { + "bbox": [ + 124, + 294, + 274, + 306 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2111.02387, 2021.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 109, + 307, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 109, + 307, + 506, + 322 + ], + "score": 1.0, + "content": "[9] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples:", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 125, + 318, + 493, + 331 + ], + "spans": [ + { + "bbox": [ + 125, + 318, + 493, + 331 + ], + "score": 1.0, + "content": "An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178–178, 2004.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 332, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 506, + 345 + ], + "score": 1.0, + "content": "[10] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 125, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 125, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 353, + 220, + 365 + ], + "spans": [ + { + "bbox": [ + 126, + 353, + 220, + 365 + ], + "score": 1.0, + "content": "arXiv:2110.04544, 2021.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 125, + 377, + 245, + 390 + ], + "spans": [ + { + "bbox": [ + 125, + 377, + 245, + 390 + ], + "score": 1.0, + "content": "In CVPR, pages 770–778, 2016.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 391, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 104, + 391, + 506, + 405 + ], + "score": 1.0, + "content": "[12] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 124, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 124, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 412, + 365, + 423 + ], + "spans": [ + { + "bbox": [ + 126, + 412, + 365, + 423 + ], + "score": 1.0, + "content": "Earth Observations and Remote Sensing, 12(7):2217–2226, 2019.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "[13] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 126, + 436, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 126, + 436, + 506, + 449 + ], + "score": 1.0, + "content": "Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 125, + 445, + 397, + 458 + ], + "spans": [ + { + "bbox": [ + 125, + 445, + 397, + 458 + ], + "score": 1.0, + "content": "out-of-distribution generalization. arXiv preprint arXiv:2006.16241, 2020.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "score": 1.0, + "content": "[14] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 125, + 470, + 312, + 483 + ], + "spans": [ + { + "bbox": [ + 125, + 470, + 312, + 483 + ], + "score": 1.0, + "content": "examples. arXiv preprint arXiv:1907.07174, 2019.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "[15] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 496, + 401, + 509 + ], + "spans": [ + { + "bbox": [ + 126, + 496, + 401, + 509 + ], + "score": 1.0, + "content": "vision-and-language bert for navigation. In CVPR, pages 1643–1653, 2021.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "[16] Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei Yang, and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 125, + 520, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 125, + 520, + 506, + 533 + ], + "score": 1.0, + "content": "Jason Baldridge. Mural: multimodal, multitask retrieval across languages. arXiv preprint arXiv:2109.05125,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 125, + 528, + 151, + 542 + ], + "spans": [ + { + "bbox": [ + 125, + 528, + 151, + 542 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 542, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 506, + 559 + ], + "score": 1.0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 126, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 125, + 564, + 300, + 577 + ], + "spans": [ + { + "bbox": [ + 125, + 564, + 300, + 577 + ], + "score": 1.0, + "content": "supervision. In ICML, pages 4904–4916, 2021.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "[18] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 126, + 589, + 245, + 601 + ], + "spans": [ + { + "bbox": [ + 126, + 589, + 245, + 601 + ], + "score": 1.0, + "content": "know? TACL, 8:423–438, 2020.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "score": 1.0, + "content": "[19] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 614, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 126, + 614, + 506, + 625 + ], + "score": 1.0, + "content": "Mdetr-modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1780–1790, 2021.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "score": 1.0, + "content": "[20] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 125, + 638, + 325, + 651 + ], + "spans": [ + { + "bbox": [ + 125, + 638, + 325, + 651 + ], + "score": 1.0, + "content": "region supervision. In ICML, pages 5583–5594, 2021.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 652, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 506, + 665 + ], + "score": 1.0, + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 125, + 662, + 306, + 676 + ], + "spans": [ + { + "bbox": [ + 125, + 662, + 306, + 676 + ], + "score": 1.0, + "content": "categorization. In ICCVW, pages 554–561, 2013.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 676, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 691 + ], + "score": 1.0, + "content": "[22] Charlotte Laclau, Ievgen Redko, Basarab Matei, Younes Bennani, and Vincent Brault. Co-clustering", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 125, + 687, + 350, + 700 + ], + "spans": [ + { + "bbox": [ + 125, + 687, + 350, + 700 + ], + "score": 1.0, + "content": "through optimal transport. In ICML, pages 1955–1964, 2017.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 699, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 716 + ], + "score": 1.0, + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 125, + 711, + 484, + 725 + ], + "spans": [ + { + "bbox": [ + 125, + 711, + 484, + 725 + ], + "score": 1.0, + "content": "for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 27 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 85, + 79, + 102, + 730 + ], + "lines": [ + { + "bbox": [ + 84, + 76, + 101, + 86 + ], + "spans": [ + { + "bbox": [ + 84, + 76, + 101, + 86 + ], + "score": 1.0, + "content": "338", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 93, + 101, + 104 + ], + "spans": [ + { + "bbox": [ + 84, + 93, + 101, + 104 + ], + "score": 1.0, + "content": "339", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 102, + 100, + 114 + ], + "spans": [ + { + "bbox": [ + 84, + 102, + 100, + 114 + ], + "score": 1.0, + "content": "340", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 118, + 100, + 128 + ], + "spans": [ + { + "bbox": [ + 84, + 118, + 100, + 128 + ], + "score": 1.0, + "content": "341", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 128, + 100, + 139 + ], + "spans": [ + { + "bbox": [ + 84, + 128, + 100, + 139 + ], + "score": 1.0, + "content": "342", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 142, + 100, + 154 + ], + "spans": [ + { + "bbox": [ + 84, + 142, + 100, + 154 + ], + "score": 1.0, + "content": "343", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 152, + 100, + 164 + ], + "spans": [ + { + "bbox": [ + 84, + 152, + 100, + 164 + ], + "score": 1.0, + "content": "344", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 166, + 100, + 178 + ], + "spans": [ + { + "bbox": [ + 84, + 166, + 100, + 178 + ], + "score": 1.0, + "content": "345", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 177, + 100, + 189 + ], + "spans": [ + { + "bbox": [ + 84, + 177, + 100, + 189 + ], + "score": 1.0, + "content": "346", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 186, + 100, + 199 + ], + "spans": [ + { + "bbox": [ + 84, + 186, + 100, + 199 + ], + "score": 1.0, + "content": "347", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 201, + 100, + 213 + ], + "spans": [ + { + "bbox": [ + 84, + 201, + 100, + 213 + ], + "score": 1.0, + "content": "348", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 210, + 101, + 224 + ], + "spans": [ + { + "bbox": [ + 84, + 210, + 101, + 224 + ], + "score": 1.0, + "content": "349", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 226, + 100, + 237 + ], + "spans": [ + { + "bbox": [ + 84, + 226, + 100, + 237 + ], + "score": 1.0, + "content": "350", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 236, + 100, + 248 + ], + "spans": [ + { + "bbox": [ + 84, + 236, + 100, + 248 + ], + "score": 1.0, + "content": "351", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 250, + 100, + 262 + ], + "spans": [ + { + "bbox": [ + 84, + 250, + 100, + 262 + ], + "score": 1.0, + "content": "352", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 259, + 101, + 274 + ], + "spans": [ + { + "bbox": [ + 84, + 259, + 101, + 274 + ], + "score": 1.0, + "content": "353", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 275, + 101, + 286 + ], + "spans": [ + { + "bbox": [ + 84, + 275, + 101, + 286 + ], + "score": 1.0, + "content": "354", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 284, + 100, + 296 + ], + "spans": [ + { + "bbox": [ + 84, + 284, + 100, + 296 + ], + "score": 1.0, + "content": "355", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 294, + 100, + 307 + ], + "spans": [ + { + "bbox": [ + 84, + 294, + 100, + 307 + ], + "score": 1.0, + "content": "356", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 309, + 100, + 320 + ], + "spans": [ + { + "bbox": [ + 84, + 309, + 100, + 320 + ], + "score": 1.0, + "content": "357", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 320, + 100, + 331 + ], + "spans": [ + { + "bbox": [ + 84, + 320, + 100, + 331 + ], + "score": 1.0, + "content": "358", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 334, + 100, + 344 + ], + "spans": [ + { + "bbox": [ + 84, + 334, + 100, + 344 + ], + "score": 1.0, + "content": "359", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 344, + 100, + 355 + ], + "spans": [ + { + "bbox": [ + 84, + 344, + 100, + 355 + ], + "score": 1.0, + "content": "360", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 354, + 100, + 366 + ], + "spans": [ + { + "bbox": [ + 84, + 354, + 100, + 366 + ], + "score": 1.0, + "content": "361", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 368, + 100, + 380 + ], + "spans": [ + { + "bbox": [ + 84, + 368, + 100, + 380 + ], + "score": 1.0, + "content": "362", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 378, + 100, + 390 + ], + "spans": [ + { + "bbox": [ + 84, + 378, + 100, + 390 + ], + "score": 1.0, + "content": "363", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 393, + 101, + 404 + ], + "spans": [ + { + "bbox": [ + 84, + 393, + 101, + 404 + ], + "score": 1.0, + "content": "364", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 403, + 100, + 414 + ], + "spans": [ + { + "bbox": [ + 84, + 403, + 100, + 414 + ], + "score": 1.0, + "content": "365", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 412, + 102, + 425 + ], + "spans": [ + { + "bbox": [ + 84, + 412, + 102, + 425 + ], + "score": 1.0, + "content": "366", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 427, + 100, + 438 + ], + "spans": [ + { + "bbox": [ + 84, + 427, + 100, + 438 + ], + "score": 1.0, + "content": "367", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 438, + 100, + 449 + ], + "spans": [ + { + "bbox": [ + 84, + 438, + 100, + 449 + ], + "score": 1.0, + "content": "368", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 447, + 100, + 459 + ], + "spans": [ + { + "bbox": [ + 86, + 447, + 100, + 459 + ], + "score": 1.0, + "content": "369", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 462, + 100, + 473 + ], + "spans": [ + { + "bbox": [ + 84, + 462, + 100, + 473 + ], + "score": 1.0, + "content": "370", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 472, + 100, + 483 + ], + "spans": [ + { + "bbox": [ + 84, + 472, + 100, + 483 + ], + "score": 1.0, + "content": "371", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 486, + 101, + 498 + ], + "spans": [ + { + "bbox": [ + 84, + 486, + 101, + 498 + ], + "score": 1.0, + "content": "372", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 496, + 100, + 508 + ], + "spans": [ + { + "bbox": [ + 84, + 496, + 100, + 508 + ], + "score": 1.0, + "content": "373", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 511, + 101, + 522 + ], + "spans": [ + { + "bbox": [ + 84, + 511, + 101, + 522 + ], + "score": 1.0, + "content": "374", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 521, + 101, + 532 + ], + "spans": [ + { + "bbox": [ + 84, + 521, + 101, + 532 + ], + "score": 1.0, + "content": "375", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 531, + 100, + 542 + ], + "spans": [ + { + "bbox": [ + 84, + 531, + 100, + 542 + ], + "score": 1.0, + "content": "376", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 545, + 100, + 556 + ], + "spans": [ + { + "bbox": [ + 84, + 545, + 100, + 556 + ], + "score": 1.0, + "content": "377", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 555, + 101, + 567 + ], + "spans": [ + { + "bbox": [ + 84, + 555, + 101, + 567 + ], + "score": 1.0, + "content": "378", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 565, + 100, + 577 + ], + "spans": [ + { + "bbox": [ + 84, + 565, + 100, + 577 + ], + "score": 1.0, + "content": "379", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 580, + 100, + 591 + ], + "spans": [ + { + "bbox": [ + 84, + 580, + 100, + 591 + ], + "score": 1.0, + "content": "380", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 590, + 100, + 603 + ], + "spans": [ + { + "bbox": [ + 84, + 590, + 100, + 603 + ], + "score": 1.0, + "content": "381", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 604, + 100, + 616 + ], + "spans": [ + { + "bbox": [ + 84, + 604, + 100, + 616 + ], + "score": 1.0, + "content": "382", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 614, + 100, + 626 + ], + "spans": [ + { + "bbox": [ + 84, + 614, + 100, + 626 + ], + "score": 1.0, + "content": "383", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 628, + 102, + 642 + ], + "spans": [ + { + "bbox": [ + 84, + 628, + 102, + 642 + ], + "score": 1.0, + "content": "384", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 639, + 100, + 651 + ], + "spans": [ + { + "bbox": [ + 84, + 639, + 100, + 651 + ], + "score": 1.0, + "content": "385", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 653, + 102, + 666 + ], + "spans": [ + { + "bbox": [ + 84, + 653, + 102, + 666 + ], + "score": 1.0, + "content": "386", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 664, + 100, + 675 + ], + "spans": [ + { + "bbox": [ + 86, + 664, + 100, + 675 + ], + "score": 1.0, + "content": "387", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 678, + 100, + 689 + ], + "spans": [ + { + "bbox": [ + 84, + 678, + 100, + 689 + ], + "score": 1.0, + "content": "388", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 688, + 100, + 700 + ], + "spans": [ + { + "bbox": [ + 84, + 688, + 100, + 700 + ], + "score": 1.0, + "content": "389", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 703, + 100, + 714 + ], + "spans": [ + { + "bbox": [ + 84, + 703, + 100, + 714 + ], + "score": 1.0, + "content": "390", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 713, + 100, + 724 + ], + "spans": [ + { + "bbox": [ + 84, + 713, + 100, + 724 + ], + "score": 1.0, + "content": "391", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 89, + 72, + 164, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 165, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 165, + 87 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "list", + "bbox": [ + 101, + 86, + 508, + 723 + ], + "lines": [ + { + "bbox": [ + 110, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 110, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "[1] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 101, + 234, + 115 + ], + "spans": [ + { + "bbox": [ + 125, + 101, + 234, + 115 + ], + "score": 1.0, + "content": "ICML, pages 214–223, 2017.", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 117, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 110, + 117, + 505, + 128 + ], + "score": 1.0, + "content": "[2] Emmanuel Boissard, Thibaut Le Gouic, and Jean-Michel Loubes. Distribution’s template estimate with", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 127, + 321, + 138 + ], + "spans": [ + { + "bbox": [ + 126, + 127, + 321, + 138 + ], + "score": 1.0, + "content": "wasserstein metrics. Bernoulli, 21(2):740–759, 2015.", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 141, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 110, + 141, + 505, + 153 + ], + "score": 1.0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative components", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 151, + 323, + 165 + ], + "spans": [ + { + "bbox": [ + 125, + 151, + 323, + 165 + ], + "score": 1.0, + "content": "with random forests. In ECCV, pages 446–461, 2014.", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 110, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "[4] Liqun Chen, Yizhe Zhang, Ruiyi Zhang, Chenyang Tao, Zhe Gan, Haichao Zhang, Bai Li, Dinghan Shen,", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 125, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "Changyou Chen, and Lawrence Carin. Improving sequence-to-sequence learning via optimal transport.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 124, + 185, + 273, + 198 + ], + "spans": [ + { + "bbox": [ + 124, + 185, + 273, + 198 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1901.06283, 2019.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 108, + 198, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 108, + 198, + 506, + 214 + ], + "score": 1.0, + "content": "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 209, + 328, + 222 + ], + "spans": [ + { + "bbox": [ + 124, + 209, + 328, + 222 + ], + "score": 1.0, + "content": "textures in the wild. In CVPR, pages 3606–3613, 2014.", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 224, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 110, + 224, + 506, + 237 + ], + "score": 1.0, + "content": "[6] Marco Cuturi. Sinkhorn distances: lightspeed computation of optimal transport. In NeurIPS, volume 2,", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 236, + 178, + 246 + ], + "spans": [ + { + "bbox": [ + 124, + 236, + 178, + 246 + ], + "score": 1.0, + "content": "page 4, 2013.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 249, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 109, + 249, + 506, + 262 + ], + "score": 1.0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 259, + 304, + 272 + ], + "spans": [ + { + "bbox": [ + 124, + 259, + 304, + 272 + ], + "score": 1.0, + "content": "image database. In CVPR, pages 248–255, 2009.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 108, + 272, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 108, + 272, + 506, + 288 + ], + "score": 1.0, + "content": "[8] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Zicheng", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 284, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 125, + 284, + 506, + 297 + ], + "score": 1.0, + "content": "Liu, Michael Zeng, et al. An empirical study of training end-to-end vision-and-language transformers.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 124, + 294, + 274, + 306 + ], + "spans": [ + { + "bbox": [ + 124, + 294, + 274, + 306 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2111.02387, 2021.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 307, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 109, + 307, + 506, + 322 + ], + "score": 1.0, + "content": "[9] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples:", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 318, + 493, + 331 + ], + "spans": [ + { + "bbox": [ + 125, + 318, + 493, + 331 + ], + "score": 1.0, + "content": "An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178–178, 2004.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 332, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 506, + 345 + ], + "score": 1.0, + "content": "[10] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li,", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 125, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. arXiv preprint", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 353, + 220, + 365 + ], + "spans": [ + { + "bbox": [ + 126, + 353, + 220, + 365 + ], + "score": 1.0, + "content": "arXiv:2110.04544, 2021.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 506, + 380 + ], + "score": 1.0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 377, + 245, + 390 + ], + "spans": [ + { + "bbox": [ + 125, + 377, + 245, + 390 + ], + "score": 1.0, + "content": "In CVPR, pages 770–778, 2016.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 391, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 104, + 391, + 506, + 405 + ], + "score": 1.0, + "content": "[12] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 124, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 412, + 365, + 423 + ], + "spans": [ + { + "bbox": [ + 126, + 412, + 365, + 423 + ], + "score": 1.0, + "content": "Earth Observations and Remote Sensing, 12(7):2217–2226, 2019.", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 427, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 505, + 439 + ], + "score": 1.0, + "content": "[13] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai,", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 436, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 126, + 436, + 506, + 449 + ], + "score": 1.0, + "content": "Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 125, + 445, + 397, + 458 + ], + "spans": [ + { + "bbox": [ + 125, + 445, + 397, + 458 + ], + "score": 1.0, + "content": "out-of-distribution generalization. arXiv preprint arXiv:2006.16241, 2020.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 506, + 474 + ], + "score": 1.0, + "content": "[14] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 470, + 312, + 483 + ], + "spans": [ + { + "bbox": [ + 125, + 470, + 312, + 483 + ], + "score": 1.0, + "content": "examples. arXiv preprint arXiv:1907.07174, 2019.", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "[15] Yicong Hong, Qi Wu, Yuankai Qi, Cristian Rodriguez-Opazo, and Stephen Gould. Vln bert: A recurrent", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 496, + 401, + 509 + ], + "spans": [ + { + "bbox": [ + 126, + 496, + 401, + 509 + ], + "score": 1.0, + "content": "vision-and-language bert for navigation. In CVPR, pages 1643–1653, 2021.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "[16] Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei Yang, and", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 520, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 125, + 520, + 506, + 533 + ], + "score": 1.0, + "content": "Jason Baldridge. Mural: multimodal, multitask retrieval across languages. arXiv preprint arXiv:2109.05125,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 125, + 528, + 151, + 542 + ], + "spans": [ + { + "bbox": [ + 125, + 528, + 151, + 542 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 542, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 506, + 559 + ], + "score": 1.0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung,", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 554, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 126, + 554, + 506, + 568 + ], + "score": 1.0, + "content": "Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 125, + 564, + 300, + 577 + ], + "spans": [ + { + "bbox": [ + 125, + 564, + 300, + 577 + ], + "score": 1.0, + "content": "supervision. In ICML, pages 4904–4916, 2021.", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "[18] Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 589, + 245, + 601 + ], + "spans": [ + { + "bbox": [ + 126, + 589, + 245, + 601 + ], + "score": 1.0, + "content": "know? TACL, 8:423–438, 2020.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 615 + ], + "score": 1.0, + "content": "[19] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion.", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 614, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 126, + 614, + 506, + 625 + ], + "score": 1.0, + "content": "Mdetr-modulated detection for end-to-end multi-modal understanding. In ICCV, pages 1780–1790, 2021.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 641 + ], + "score": 1.0, + "content": "[20] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or", + "type": "text" + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 638, + 325, + 651 + ], + "spans": [ + { + "bbox": [ + 125, + 638, + 325, + 651 + ], + "score": 1.0, + "content": "region supervision. In ICML, pages 5583–5594, 2021.", + "type": "text" + } + ], + "index": 47, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 652, + 506, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 506, + 665 + ], + "score": 1.0, + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained", + "type": "text" + } + ], + "index": 48, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 662, + 306, + 676 + ], + "spans": [ + { + "bbox": [ + 125, + 662, + 306, + 676 + ], + "score": 1.0, + "content": "categorization. In ICCVW, pages 554–561, 2013.", + "type": "text" + } + ], + "index": 49, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 676, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 691 + ], + "score": 1.0, + "content": "[22] Charlotte Laclau, Ievgen Redko, Basarab Matei, Younes Bennani, and Vincent Brault. Co-clustering", + "type": "text" + } + ], + "index": 50, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 687, + 350, + 700 + ], + "spans": [ + { + "bbox": [ + 125, + 687, + 350, + 700 + ], + "score": 1.0, + "content": "through optimal transport. In ICML, pages 1955–1964, 2017.", + "type": "text" + } + ], + "index": 51, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 699, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 506, + 716 + ], + "score": 1.0, + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training", + "type": "text" + } + ], + "index": 52, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 711, + 484, + 725 + ], + "spans": [ + { + "bbox": [ + 125, + 711, + 484, + 725 + ], + "score": 1.0, + "content": "for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022.", + "type": "text" + } + ], + "index": 53, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 72, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 87 + ], + "score": 1.0, + "content": "[24] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS,", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 93, + 164, + 105 + ], + "spans": [ + { + "bbox": [ + 126, + 93, + 164, + 105 + ], + "score": 1.0, + "content": "34, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 108, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 108, + 506, + 119 + ], + "score": 1.0, + "content": "[25] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 117, + 437, + 130 + ], + "spans": [ + { + "bbox": [ + 125, + 117, + 437, + 130 + ], + "score": 1.0, + "content": "performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 131, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 145 + ], + "score": 1.0, + "content": "[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 141, + 252, + 153 + ], + "spans": [ + { + "bbox": [ + 124, + 141, + 252, + 153 + ], + "score": 1.0, + "content": "preprint arXiv:2101.00190, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 155, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 506, + 167 + ], + "score": 1.0, + "content": "[27] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train,", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 125, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 124, + 175, + 252, + 187 + ], + "spans": [ + { + "bbox": [ + 124, + 175, + 252, + 187 + ], + "score": 1.0, + "content": "preprint arXiv:2107.13586, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 189, + 507, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 507, + 202 + ], + "score": 1.0, + "content": "[28] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands,", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 199, + 290, + 209 + ], + "spans": [ + { + "bbox": [ + 125, + 199, + 290, + 209 + ], + "score": 1.0, + "content": "too. arXiv preprint arXiv:2103.10385, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 213, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 506, + 225 + ], + "score": 1.0, + "content": "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 223, + 358, + 234 + ], + "spans": [ + { + "bbox": [ + 126, + 223, + 358, + 234 + ], + "score": 1.0, + "content": "classification of aircraft. arXiv preprint arXiv:1306.5151, 2013.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 236, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 248 + ], + "score": 1.0, + "content": "[30] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Histoire de l’Académie Royale des", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 246, + 215, + 257 + ], + "spans": [ + { + "bbox": [ + 126, + 246, + 215, + 257 + ], + "score": 1.0, + "content": "Sciences de Paris, 1781.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "score": 1.0, + "content": "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 125, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 280, + 339, + 292 + ], + "spans": [ + { + "bbox": [ + 126, + 280, + 339, + 292 + ], + "score": 1.0, + "content": "diffusion models. arXiv preprint arXiv:2112.10741, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "score": 1.0, + "content": "[32] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 303, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 124, + 303, + 506, + 318 + ], + "score": 1.0, + "content": "classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 314, + 185, + 326 + ], + "spans": [ + { + "bbox": [ + 127, + 314, + 185, + 326 + ], + "score": 1.0, + "content": "722–729, 2008.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 325, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 104, + 325, + 506, + 342 + ], + "score": 1.0, + "content": "[33] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, pages", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 338, + 194, + 349 + ], + "spans": [ + { + "bbox": [ + 126, + 338, + 194, + 349 + ], + "score": 1.0, + "content": "3498–3505, 2012.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 352, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 506, + 364 + ], + "score": 1.0, + "content": "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen,", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 360, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 124, + 360, + 506, + 375 + ], + "score": 1.0, + "content": "Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep", + "type": "text", + "cross_page": true + } + ], + "index": 25 + }, + { + "bbox": [ + 125, + 371, + 246, + 385 + ], + "spans": [ + { + "bbox": [ + 125, + 371, + 246, + 385 + ], + "score": 1.0, + "content": "learning library. NeurIPS, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "[35] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 395, + 377, + 408 + ], + "spans": [ + { + "bbox": [ + 124, + 395, + 377, + 408 + ], + "score": 1.0, + "content": "manipulation of stylegan imagery. In ICCV, pages 2085–2094, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 410, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 506, + 421 + ], + "score": 1.0, + "content": "[36] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 419, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 125, + 419, + 482, + 432 + ], + "score": 1.0, + "content": "Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 434, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 445 + ], + "score": 1.0, + "content": "[37] Gabriel Peyre and Marco Cuturi. Computational optimal transport. Foundations and Trends in Machine", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 443, + 251, + 455 + ], + "spans": [ + { + "bbox": [ + 126, + 443, + 251, + 455 + ], + "score": 1.0, + "content": "Learning, 11(5-6):355–607, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "[38] Nina Poerner, Ulli Waltinger, and Hinrich Schütze. Bert is not a knowledge base (yet): Factual knowledge", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 468, + 439, + 479 + ], + "spans": [ + { + "bbox": [ + 126, + 468, + 439, + 479 + ], + "score": 1.0, + "content": "vs. name-based reasoning in unsupervised qa. arXiv preprint arXiv:1911.03681, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 491, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 126, + 491, + 506, + 503 + ], + "score": 1.0, + "content": "Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from", + "type": "text", + "cross_page": true + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 500, + 381, + 514 + ], + "spans": [ + { + "bbox": [ + 124, + 500, + 381, + 514 + ], + "score": 1.0, + "content": "natural language supervision. arXiv preprint arXiv:2103.00020, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "[40] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 524, + 386, + 536 + ], + "spans": [ + { + "bbox": [ + 126, + 524, + 386, + 536 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 538, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 538, + 506, + 551 + ], + "score": 1.0, + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 548, + 399, + 561 + ], + "spans": [ + { + "bbox": [ + 124, + 548, + 399, + 561 + ], + "score": 1.0, + "content": "image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "score": 1.0, + "content": "[42] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 571, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 124, + 571, + 506, + 586 + ], + "score": 1.0, + "content": "Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 43 + }, + { + "bbox": [ + 125, + 583, + 220, + 593 + ], + "spans": [ + { + "bbox": [ + 125, + 583, + 220, + 593 + ], + "score": 1.0, + "content": "arXiv:2112.01518, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "[43] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 605, + 362, + 618 + ], + "spans": [ + { + "bbox": [ + 124, + 605, + 362, + 618 + ], + "score": 1.0, + "content": "generalize to imagenet? arXiv preprint arXiv:1902.10811, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 46, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 618, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 103, + 618, + 506, + 634 + ], + "score": 1.0, + "content": "[44] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover’s distance as a metric for image", + "type": "text", + "cross_page": true + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 630, + 260, + 641 + ], + "spans": [ + { + "bbox": [ + 126, + 630, + 260, + 641 + ], + "score": 1.0, + "content": "retrieval. IJCV, 40(2):99–121, 2000.", + "type": "text", + "cross_page": true + } + ], + "index": 48, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 643, + 507, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 507, + 657 + ], + "score": 1.0, + "content": "[45] Tim Salimans, Han Zhang, Alec Radford, and Dimitris Metaxas. Improving gans using optimal transport.", + "type": "text", + "cross_page": true + } + ], + "index": 49, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 651, + 176, + 667 + ], + "spans": [ + { + "bbox": [ + 125, + 651, + 176, + 667 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 50, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "[46] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting", + "type": "text", + "cross_page": true + } + ], + "index": 51, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 125, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980,", + "type": "text", + "cross_page": true + } + ], + "index": 52 + }, + { + "bbox": [ + 125, + 686, + 151, + 701 + ], + "spans": [ + { + "bbox": [ + 125, + 686, + 151, + 701 + ], + "score": 1.0, + "content": "2020.", + "type": "text", + "cross_page": true + } + ], + "index": 53, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 700, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 714 + ], + "score": 1.0, + "content": "[47] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions", + "type": "text", + "cross_page": true + } + ], + "index": 54, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 711, + 385, + 723 + ], + "spans": [ + { + "bbox": [ + 124, + 711, + 385, + 723 + ], + "score": 1.0, + "content": "classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.", + "type": "text", + "cross_page": true + } + ], + "index": 55, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "[48] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 83, + 416, + 95 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 416, + 95 + ], + "score": 1.0, + "content": "human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 97, + 506, + 110 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 506, + 110 + ], + "score": 1.0, + "content": "[49] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 107, + 406, + 119 + ], + "spans": [ + { + "bbox": [ + 125, + 107, + 406, + 119 + ], + "score": 1.0, + "content": "few-shot learning with frozen language models. NeurIPS, 34:200–212, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "score": 1.0, + "content": "[50] Ruibo Tu, Kun Zhang, Hedvig Kjellström, and Cheng Zhang. Optimal transport for causal discovery.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 131, + 172, + 142 + ], + "spans": [ + { + "bbox": [ + 126, + 131, + 172, + 142 + ], + "score": 1.0, + "content": "ICLR, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 145, + 407, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 407, + 158 + ], + "score": 1.0, + "content": "[51] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009.", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 159, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 171 + ], + "score": 1.0, + "content": "[52] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 170, + 322, + 181 + ], + "spans": [ + { + "bbox": [ + 126, + 170, + 322, + 181 + ], + "score": 1.0, + "content": "penalizing local predictive power. NeurIPS, 32, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 182, + 507, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 507, + 196 + ], + "score": 1.0, + "content": "[53] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 193, + 273, + 204 + ], + "spans": [ + { + "bbox": [ + 126, + 193, + 273, + 204 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2109.08472, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "[54] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 217, + 380, + 229 + ], + "spans": [ + { + "bbox": [ + 125, + 217, + 380, + 229 + ], + "score": 1.0, + "content": "mixture-of-modality-experts. arXiv preprint arXiv:2111.02358, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "[55] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 241, + 409, + 252 + ], + "spans": [ + { + "bbox": [ + 125, + 241, + 409, + 252 + ], + "score": 1.0, + "content": "scale scene recognition from abbey to zoo. In CVPR, pages 3485–3492, 2010.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "[56] Hongteng Xu, Dixin Luo, Hongyuan Zha, and Lawrence Carin Duke. Gromov-wasserstein learning for", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 266, + 391, + 276 + ], + "spans": [ + { + "bbox": [ + 126, + 266, + 391, + 276 + ], + "score": 1.0, + "content": "graph matching and node embedding. In ICML, pages 6932–6941, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "[57] Jingjing Xu, Hao Zhou, Chun Gan, Zaixiang Zheng, and Lei Li. Vocabulary learning via optimal transport", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 288, + 385, + 300 + ], + "spans": [ + { + "bbox": [ + 125, + 288, + 385, + 300 + ], + "score": 1.0, + "content": "for neural machine translation. arXiv preprint arXiv:2012.15671, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "[58] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hong-", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 125, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "sheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 124, + 323, + 219, + 334 + ], + "spans": [ + { + "bbox": [ + 124, + 323, + 219, + 334 + ], + "score": 1.0, + "content": "arXiv:2111.03930, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "[59] Renrui Zhang, Longtian Qiu, Wei Zhang, and Ziyao Zeng. Vt-clip: Enhancing vision-language models", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 348, + 363, + 358 + ], + "spans": [ + { + "bbox": [ + 126, + 348, + 363, + 358 + ], + "score": 1.0, + "content": "with visual-guided texts. arXiv preprint arXiv:2112.02399, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 360, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 373 + ], + "score": 1.0, + "content": "[60] He Zhao, Dinh Phung, Viet Huynh, Trung Le, and Wray Buntine. Neural topic model via optimal transport.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 371, + 172, + 381 + ], + "spans": [ + { + "bbox": [ + 126, + 371, + 172, + 381 + ], + "score": 1.0, + "content": "ICLR, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "[61] Wenliang Zhao, Yongming Rao, Ziyi Wang, Jiwen Lu, and Jie Zhou. Towards interpretable deep metric", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 394, + 375, + 406 + ], + "spans": [ + { + "bbox": [ + 126, + 394, + 375, + 406 + ], + "score": 1.0, + "content": "learning with structural matching. In ICCV, pages 9887–9896, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 409, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 505, + 420 + ], + "score": 1.0, + "content": "[62] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 419, + 220, + 429 + ], + "spans": [ + { + "bbox": [ + 124, + 419, + 220, + 429 + ], + "score": 1.0, + "content": "arXiv:2112.01071, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 431, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 107, + 431, + 505, + 446 + ], + "score": 1.0, + "content": "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 442, + 304, + 454 + ], + "spans": [ + { + "bbox": [ + 125, + 442, + 304, + 454 + ], + "score": 1.0, + "content": "models. arXiv preprint arXiv:2109.01134, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 467, + 276, + 477 + ], + "spans": [ + { + "bbox": [ + 126, + 467, + 276, + 477 + ], + "score": 1.0, + "content": "vision-language models. In CVPR, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + } + ], + "index": 27, + "bbox_fs": [ + 104, + 92, + 506, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 102, + 72, + 507, + 726 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 87 + ], + "score": 1.0, + "content": "[24] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 93, + 164, + 105 + ], + "spans": [ + { + "bbox": [ + 126, + 93, + 164, + 105 + ], + "score": 1.0, + "content": "34, 2021.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 108, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 108, + 506, + 119 + ], + "score": 1.0, + "content": "[25] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 117, + 437, + 130 + ], + "spans": [ + { + "bbox": [ + 125, + 117, + 437, + 130 + ], + "score": 1.0, + "content": "performant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 131, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 145 + ], + "score": 1.0, + "content": "[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 141, + 252, + 153 + ], + "spans": [ + { + "bbox": [ + 124, + 141, + 252, + 153 + ], + "score": 1.0, + "content": "preprint arXiv:2101.00190, 2021.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 155, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 506, + 167 + ], + "score": 1.0, + "content": "[27] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 125, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "prompt, and predict: A systematic survey of prompting methods in natural language processing. arXiv", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 124, + 175, + 252, + 187 + ], + "spans": [ + { + "bbox": [ + 124, + 175, + 252, + 187 + ], + "score": 1.0, + "content": "preprint arXiv:2107.13586, 2021.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 189, + 507, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 189, + 507, + 202 + ], + "score": 1.0, + "content": "[28] Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. Gpt understands,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 125, + 199, + 290, + 209 + ], + "spans": [ + { + "bbox": [ + 125, + 199, + 290, + 209 + ], + "score": 1.0, + "content": "too. arXiv preprint arXiv:2103.10385, 2021.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 213, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 506, + 225 + ], + "score": 1.0, + "content": "[29] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 223, + 358, + 234 + ], + "spans": [ + { + "bbox": [ + 126, + 223, + 358, + 234 + ], + "score": 1.0, + "content": "classification of aircraft. arXiv preprint arXiv:1306.5151, 2013.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 236, + 506, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 248 + ], + "score": 1.0, + "content": "[30] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Histoire de l’Académie Royale des", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 246, + 215, + 257 + ], + "spans": [ + { + "bbox": [ + 126, + 246, + 215, + 257 + ], + "score": 1.0, + "content": "Sciences de Paris, 1781.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "score": 1.0, + "content": "[31] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 125, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 125, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 280, + 339, + 292 + ], + "spans": [ + { + "bbox": [ + 126, + 280, + 339, + 292 + ], + "score": 1.0, + "content": "diffusion models. arXiv preprint arXiv:2112.10741, 2021.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 506, + 306 + ], + "score": 1.0, + "content": "[32] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 124, + 303, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 124, + 303, + 506, + 318 + ], + "score": 1.0, + "content": "classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 314, + 185, + 326 + ], + "spans": [ + { + "bbox": [ + 127, + 314, + 185, + 326 + ], + "score": 1.0, + "content": "722–729, 2008.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 325, + 506, + 342 + ], + "spans": [ + { + "bbox": [ + 104, + 325, + 506, + 342 + ], + "score": 1.0, + "content": "[33] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, pages", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 338, + 194, + 349 + ], + "spans": [ + { + "bbox": [ + 126, + 338, + 194, + 349 + ], + "score": 1.0, + "content": "3498–3505, 2012.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 352, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 506, + 364 + ], + "score": 1.0, + "content": "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 124, + 360, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 124, + 360, + 506, + 375 + ], + "score": 1.0, + "content": "Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 125, + 371, + 246, + 385 + ], + "spans": [ + { + "bbox": [ + 125, + 371, + 246, + 385 + ], + "score": 1.0, + "content": "learning library. NeurIPS, 2019.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "[35] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 124, + 395, + 377, + 408 + ], + "spans": [ + { + "bbox": [ + 124, + 395, + 377, + 408 + ], + "score": 1.0, + "content": "manipulation of stylegan imagery. In ICCV, pages 2085–2094, 2021.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 410, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 506, + 421 + ], + "score": 1.0, + "content": "[36] Fabio Petroni, Tim Rocktäschel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander H Miller, and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 125, + 419, + 482, + 432 + ], + "spans": [ + { + "bbox": [ + 125, + 419, + 482, + 432 + ], + "score": 1.0, + "content": "Sebastian Riedel. Language models as knowledge bases? arXiv preprint arXiv:1909.01066, 2019.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 434, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 445 + ], + "score": 1.0, + "content": "[37] Gabriel Peyre and Marco Cuturi. Computational optimal transport. Foundations and Trends in Machine", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 443, + 251, + 455 + ], + "spans": [ + { + "bbox": [ + 126, + 443, + 251, + 455 + ], + "score": 1.0, + "content": "Learning, 11(5-6):355–607, 2019.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "[38] Nina Poerner, Ulli Waltinger, and Hinrich Schütze. Bert is not a knowledge base (yet): Factual knowledge", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 468, + 439, + 479 + ], + "spans": [ + { + "bbox": [ + 126, + 468, + 439, + 479 + ], + "score": 1.0, + "content": "vs. name-based reasoning in unsupervised qa. arXiv preprint arXiv:1911.03681, 2019.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 506, + 493 + ], + "score": 1.0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 126, + 491, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 126, + 491, + 506, + 503 + ], + "score": 1.0, + "content": "Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 500, + 381, + 514 + ], + "spans": [ + { + "bbox": [ + 124, + 500, + 381, + 514 + ], + "score": 1.0, + "content": "natural language supervision. arXiv preprint arXiv:2103.00020, 2021.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "[40] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 524, + 386, + 536 + ], + "spans": [ + { + "bbox": [ + 126, + 524, + 386, + 536 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 538, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 538, + 506, + 551 + ], + "score": 1.0, + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 124, + 548, + 399, + 561 + ], + "spans": [ + { + "bbox": [ + 124, + 548, + 399, + 561 + ], + "score": 1.0, + "content": "image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "score": 1.0, + "content": "[42] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 124, + 571, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 124, + 571, + 506, + 586 + ], + "score": 1.0, + "content": "Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. arXiv preprint", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 125, + 583, + 220, + 593 + ], + "spans": [ + { + "bbox": [ + 125, + 583, + 220, + 593 + ], + "score": 1.0, + "content": "arXiv:2112.01518, 2021.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "[43] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 124, + 605, + 362, + 618 + ], + "spans": [ + { + "bbox": [ + 124, + 605, + 362, + 618 + ], + "score": 1.0, + "content": "generalize to imagenet? arXiv preprint arXiv:1902.10811, 2019.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 103, + 618, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 103, + 618, + 506, + 634 + ], + "score": 1.0, + "content": "[44] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover’s distance as a metric for image", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 126, + 630, + 260, + 641 + ], + "spans": [ + { + "bbox": [ + 126, + 630, + 260, + 641 + ], + "score": 1.0, + "content": "retrieval. IJCV, 40(2):99–121, 2000.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 643, + 507, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 507, + 657 + ], + "score": 1.0, + "content": "[45] Tim Salimans, Han Zhang, Alec Radford, and Dimitris Metaxas. Improving gans using optimal transport.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 125, + 651, + 176, + 667 + ], + "spans": [ + { + "bbox": [ + 125, + 651, + 176, + 667 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "[46] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 125, + 678, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 125, + 678, + 506, + 690 + ], + "score": 1.0, + "content": "knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980,", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 125, + 686, + 151, + 701 + ], + "spans": [ + { + "bbox": [ + 125, + 686, + 151, + 701 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 700, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 714 + ], + "score": 1.0, + "content": "[47] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 124, + 711, + 385, + 723 + ], + "spans": [ + { + "bbox": [ + 124, + 711, + 385, + 723 + ], + "score": 1.0, + "content": "classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 27.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "11", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 102, + 72, + 507, + 726 + ], + "lines": [], + "index": 27.5, + "bbox_fs": [ + 103, + 72, + 507, + 723 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 70, + 507, + 482 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "[48] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 83, + 416, + 95 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 416, + 95 + ], + "score": 1.0, + "content": "human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 97, + 506, + 110 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 506, + 110 + ], + "score": 1.0, + "content": "[49] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 107, + 406, + 119 + ], + "spans": [ + { + "bbox": [ + 125, + 107, + 406, + 119 + ], + "score": 1.0, + "content": "few-shot learning with frozen language models. NeurIPS, 34:200–212, 2021.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 506, + 135 + ], + "score": 1.0, + "content": "[50] Ruibo Tu, Kun Zhang, Hedvig Kjellström, and Cheng Zhang. Optimal transport for causal discovery.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 131, + 172, + 142 + ], + "spans": [ + { + "bbox": [ + 126, + 131, + 172, + 142 + ], + "score": 1.0, + "content": "ICLR, 2022.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 145, + 407, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 407, + 158 + ], + "score": 1.0, + "content": "[51] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 159, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 505, + 171 + ], + "score": 1.0, + "content": "[52] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 170, + 322, + 181 + ], + "spans": [ + { + "bbox": [ + 126, + 170, + 322, + 181 + ], + "score": 1.0, + "content": "penalizing local predictive power. NeurIPS, 32, 2019.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 182, + 507, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 507, + 196 + ], + "score": 1.0, + "content": "[53] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 193, + 273, + 204 + ], + "spans": [ + { + "bbox": [ + 126, + 193, + 273, + 204 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2109.08472, 2021.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "[54] Wenhui Wang, Hangbo Bao, Li Dong, and Furu Wei. Vlmo: Unified vision-language pre-training with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 125, + 217, + 380, + 229 + ], + "spans": [ + { + "bbox": [ + 125, + 217, + 380, + 229 + ], + "score": 1.0, + "content": "mixture-of-modality-experts. arXiv preprint arXiv:2111.02358, 2021.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 243 + ], + "score": 1.0, + "content": "[55] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 241, + 409, + 252 + ], + "spans": [ + { + "bbox": [ + 125, + 241, + 409, + 252 + ], + "score": 1.0, + "content": "scale scene recognition from abbey to zoo. In CVPR, pages 3485–3492, 2010.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 505, + 267 + ], + "score": 1.0, + "content": "[56] Hongteng Xu, Dixin Luo, Hongyuan Zha, and Lawrence Carin Duke. Gromov-wasserstein learning for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 126, + 266, + 391, + 276 + ], + "spans": [ + { + "bbox": [ + 126, + 266, + 391, + 276 + ], + "score": 1.0, + "content": "graph matching and node embedding. In ICML, pages 6932–6941, 2019.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 505, + 291 + ], + "score": 1.0, + "content": "[57] Jingjing Xu, Hao Zhou, Chun Gan, Zaixiang Zheng, and Lei Li. Vocabulary learning via optimal transport", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 125, + 288, + 385, + 300 + ], + "spans": [ + { + "bbox": [ + 125, + 288, + 385, + 300 + ], + "score": 1.0, + "content": "for neural machine translation. arXiv preprint arXiv:2012.15671, 2020.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "[58] Renrui Zhang, Rongyao Fang, Peng Gao, Wei Zhang, Kunchang Li, Jifeng Dai, Yu Qiao, and Hong-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 125, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 125, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "sheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 124, + 323, + 219, + 334 + ], + "spans": [ + { + "bbox": [ + 124, + 323, + 219, + 334 + ], + "score": 1.0, + "content": "arXiv:2111.03930, 2021.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "[59] Renrui Zhang, Longtian Qiu, Wei Zhang, and Ziyao Zeng. Vt-clip: Enhancing vision-language models", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 348, + 363, + 358 + ], + "spans": [ + { + "bbox": [ + 126, + 348, + 363, + 358 + ], + "score": 1.0, + "content": "with visual-guided texts. arXiv preprint arXiv:2112.02399, 2021.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 360, + 506, + 373 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 373 + ], + "score": 1.0, + "content": "[60] He Zhao, Dinh Phung, Viet Huynh, Trung Le, and Wray Buntine. Neural topic model via optimal transport.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 371, + 172, + 381 + ], + "spans": [ + { + "bbox": [ + 126, + 371, + 172, + 381 + ], + "score": 1.0, + "content": "ICLR, 2021.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 397 + ], + "score": 1.0, + "content": "[61] Wenliang Zhao, Yongming Rao, Ziyi Wang, Jiwen Lu, and Jie Zhou. Towards interpretable deep metric", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 394, + 375, + 406 + ], + "spans": [ + { + "bbox": [ + 126, + 394, + 375, + 406 + ], + "score": 1.0, + "content": "learning with structural matching. In ICCV, pages 9887–9896, 2021.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 409, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 505, + 420 + ], + "score": 1.0, + "content": "[62] Chong Zhou, Chen Change Loy, and Bo Dai. Denseclip: Extract free dense labels from clip. arXiv preprint", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 124, + 419, + 220, + 429 + ], + "spans": [ + { + "bbox": [ + 124, + 419, + 220, + 429 + ], + "score": 1.0, + "content": "arXiv:2112.01071, 2021.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 107, + 431, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 107, + 431, + 505, + 446 + ], + "score": 1.0, + "content": "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 125, + 442, + 304, + 454 + ], + "spans": [ + { + "bbox": [ + 125, + 442, + 304, + 454 + ], + "score": 1.0, + "content": "models. arXiv preprint arXiv:2109.01134, 2021.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 126, + 467, + 276, + 477 + ], + "spans": [ + { + "bbox": [ + 126, + 467, + 276, + 477 + ], + "score": 1.0, + "content": "vision-language models. In CVPR, 2022.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 106, + 492, + 156, + 504 + ], + "lines": [ + { + "bbox": [ + 106, + 491, + 158, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 158, + 506 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 132, + 512, + 202, + 522 + ], + "lines": [ + { + "bbox": [ + 130, + 511, + 204, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 204, + 523 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 147, + 526, + 505, + 623 + ], + "lines": [ + { + "bbox": [ + 146, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 146, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contribu-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 161, + 535, + 246, + 548 + ], + "spans": [ + { + "bbox": [ + 161, + 535, + 246, + 548 + ], + "score": 1.0, + "content": "tions and scope? [Yes]", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 147, + 547, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 147, + 547, + 505, + 559 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes] See our analysis in Section 4.4. 1) Our", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 161, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 161, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "method is still need few-shot data for optimization, which cannot be applied in zero-shot setting.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 162, + 568, + 361, + 579 + ], + "spans": [ + { + "bbox": [ + 162, + 568, + 361, + 579 + ], + "score": 1.0, + "content": "2) The method needs more computing cost than CoOp.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 148, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 148, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [N/A] We propose", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 162, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 162, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "a general framework for using the vision-language pre-trained model. It is not for specific", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 162, + 600, + 381, + 612 + ], + "spans": [ + { + "bbox": [ + 162, + 600, + 381, + 612 + ], + "score": 1.0, + "content": "applications, which does not directly involve societal issues.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 146, + 611, + 504, + 624 + ], + "spans": [ + { + "bbox": [ + 146, + 611, + 504, + 624 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 132, + 626, + 286, + 636 + ], + "lines": [ + { + "bbox": [ + 130, + 625, + 288, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 625, + 288, + 638 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 147, + 640, + 423, + 663 + ], + "lines": [ + { + "bbox": [ + 146, + 639, + 424, + 651 + ], + "spans": [ + { + "bbox": [ + 146, + 639, + 424, + 651 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 146, + 651, + 398, + 663 + ], + "spans": [ + { + "bbox": [ + 146, + 651, + 398, + 663 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5 + }, + { + "type": "text", + "bbox": [ + 132, + 666, + 231, + 677 + ], + "lines": [ + { + "bbox": [ + 129, + 664, + 232, + 678 + ], + "spans": [ + { + "bbox": [ + 129, + 664, + 232, + 678 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 147, + 680, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 146, + 679, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 146, + 679, + 505, + 691 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 162, + 689, + 392, + 701 + ], + "spans": [ + { + "bbox": [ + 162, + 689, + 392, + 701 + ], + "score": 1.0, + "content": "results (either in the supplemental material or as a URL)? [Yes]", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 145, + 701, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 145, + 701, + 505, + 713 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 161, + 711, + 185, + 723 + ], + "spans": [ + { + "bbox": [ + 161, + 711, + 185, + 723 + ], + "score": 1.0, + "content": "[Yes]", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 84, + 67, + 102, + 490 + ], + "lines": [ + { + "bbox": [ + 85, + 74, + 100, + 85 + ], + "spans": [ + { + "bbox": [ + 85, + 74, + 100, + 85 + ], + "score": 1.0, + "content": "448", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 84, + 101, + 97 + ], + "spans": [ + { + "bbox": [ + 84, + 84, + 101, + 97 + ], + "score": 1.0, + "content": "449", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 97, + 100, + 109 + ], + "spans": [ + { + "bbox": [ + 85, + 97, + 100, + 109 + ], + "score": 1.0, + "content": "450", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 109, + 99, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 109, + 99, + 121 + ], + "score": 1.0, + "content": "451", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 122, + 100, + 134 + ], + "spans": [ + { + "bbox": [ + 85, + 122, + 100, + 134 + ], + "score": 1.0, + "content": "452", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 132, + 100, + 145 + ], + "spans": [ + { + "bbox": [ + 85, + 132, + 100, + 145 + ], + "score": 1.0, + "content": "453", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 146, + 100, + 159 + ], + "spans": [ + { + "bbox": [ + 85, + 146, + 100, + 159 + ], + "score": 1.0, + "content": "454", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 160, + 100, + 172 + ], + "spans": [ + { + "bbox": [ + 85, + 160, + 100, + 172 + ], + "score": 1.0, + "content": "455", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 170, + 100, + 183 + ], + "spans": [ + { + "bbox": [ + 85, + 170, + 100, + 183 + ], + "score": 1.0, + "content": "456", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 183, + 100, + 195 + ], + "spans": [ + { + "bbox": [ + 86, + 183, + 100, + 195 + ], + "score": 1.0, + "content": "457", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 194, + 100, + 207 + ], + "spans": [ + { + "bbox": [ + 85, + 194, + 100, + 207 + ], + "score": 1.0, + "content": "458", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 207, + 100, + 220 + ], + "spans": [ + { + "bbox": [ + 85, + 207, + 100, + 220 + ], + "score": 1.0, + "content": "459", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 218, + 100, + 231 + ], + "spans": [ + { + "bbox": [ + 85, + 218, + 100, + 231 + ], + "score": 1.0, + "content": "460", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 231, + 100, + 244 + ], + "spans": [ + { + "bbox": [ + 85, + 231, + 100, + 244 + ], + "score": 1.0, + "content": "461", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 242, + 101, + 255 + ], + "spans": [ + { + "bbox": [ + 85, + 242, + 101, + 255 + ], + "score": 1.0, + "content": "462", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 256, + 100, + 267 + ], + "spans": [ + { + "bbox": [ + 85, + 256, + 100, + 267 + ], + "score": 1.0, + "content": "463", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 266, + 101, + 279 + ], + "spans": [ + { + "bbox": [ + 84, + 266, + 101, + 279 + ], + "score": 1.0, + "content": "464", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 279, + 100, + 291 + ], + "spans": [ + { + "bbox": [ + 86, + 279, + 100, + 291 + ], + "score": 1.0, + "content": "465", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 290, + 100, + 303 + ], + "spans": [ + { + "bbox": [ + 85, + 290, + 100, + 303 + ], + "score": 1.0, + "content": "466", + "type": "text" + } + ] + }, + { + "bbox": [ + 86, + 303, + 100, + 315 + ], + "spans": [ + { + "bbox": [ + 86, + 303, + 100, + 315 + ], + "score": 1.0, + "content": "467", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 314, + 100, + 325 + ], + "spans": [ + { + "bbox": [ + 85, + 314, + 100, + 325 + ], + "score": 1.0, + "content": "468", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 324, + 100, + 336 + ], + "spans": [ + { + "bbox": [ + 85, + 324, + 100, + 336 + ], + "score": 1.0, + "content": "469", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 337, + 100, + 349 + ], + "spans": [ + { + "bbox": [ + 85, + 337, + 100, + 349 + ], + "score": 1.0, + "content": "470", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 347, + 100, + 361 + ], + "spans": [ + { + "bbox": [ + 85, + 347, + 100, + 361 + ], + "score": 1.0, + "content": "471", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 362, + 100, + 373 + ], + "spans": [ + { + "bbox": [ + 85, + 362, + 100, + 373 + ], + "score": 1.0, + "content": "472", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 371, + 100, + 385 + ], + "spans": [ + { + "bbox": [ + 85, + 371, + 100, + 385 + ], + "score": 1.0, + "content": "473", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 384, + 100, + 397 + ], + "spans": [ + { + "bbox": [ + 85, + 384, + 100, + 397 + ], + "score": 1.0, + "content": "474", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 396, + 100, + 408 + ], + "spans": [ + { + "bbox": [ + 85, + 396, + 100, + 408 + ], + "score": 1.0, + "content": "475", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 410, + 100, + 421 + ], + "spans": [ + { + "bbox": [ + 85, + 410, + 100, + 421 + ], + "score": 1.0, + "content": "476", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 420, + 100, + 433 + ], + "spans": [ + { + "bbox": [ + 84, + 420, + 100, + 433 + ], + "score": 1.0, + "content": "477", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 433, + 100, + 444 + ], + "spans": [ + { + "bbox": [ + 85, + 433, + 100, + 444 + ], + "score": 1.0, + "content": "478", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 444, + 100, + 455 + ], + "spans": [ + { + "bbox": [ + 85, + 444, + 100, + 455 + ], + "score": 1.0, + "content": "479", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 457, + 100, + 469 + ], + "spans": [ + { + "bbox": [ + 85, + 457, + 100, + 469 + ], + "score": 1.0, + "content": "480", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 467, + 100, + 479 + ], + "spans": [ + { + "bbox": [ + 85, + 467, + 100, + 479 + ], + "score": 1.0, + "content": "481", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 70, + 507, + 482 + ], + "lines": [], + "index": 16.5, + "bbox_fs": [ + 105, + 73, + 507, + 477 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 106, + 492, + 156, + 504 + ], + "lines": [ + { + "bbox": [ + 106, + 491, + 158, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 158, + 506 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 132, + 512, + 202, + 522 + ], + "lines": [ + { + "bbox": [ + 130, + 511, + 204, + 523 + ], + "spans": [ + { + "bbox": [ + 130, + 511, + 204, + 523 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 130, + 511, + 204, + 523 + ] + }, + { + "type": "list", + "bbox": [ + 147, + 526, + 505, + 623 + ], + "lines": [ + { + "bbox": [ + 146, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 146, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contribu-", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 535, + 246, + 548 + ], + "spans": [ + { + "bbox": [ + 161, + 535, + 246, + 548 + ], + "score": 1.0, + "content": "tions and scope? [Yes]", + "type": "text" + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 147, + 547, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 147, + 547, + 505, + 559 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes] See our analysis in Section 4.4. 1) Our", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 161, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "method is still need few-shot data for optimization, which cannot be applied in zero-shot setting.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 162, + 568, + 361, + 579 + ], + "spans": [ + { + "bbox": [ + 162, + 568, + 361, + 579 + ], + "score": 1.0, + "content": "2) The method needs more computing cost than CoOp.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 148, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 148, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [N/A] We propose", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 590, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 162, + 590, + 505, + 602 + ], + "score": 1.0, + "content": "a general framework for using the vision-language pre-trained model. It is not for specific", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 162, + 600, + 381, + 612 + ], + "spans": [ + { + "bbox": [ + 162, + 600, + 381, + 612 + ], + "score": 1.0, + "content": "applications, which does not directly involve societal issues.", + "type": "text" + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 611, + 504, + 624 + ], + "spans": [ + { + "bbox": [ + 146, + 611, + 504, + 624 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]", + "type": "text" + } + ], + "index": 44, + "is_list_start_line": true + } + ], + "index": 40, + "bbox_fs": [ + 146, + 525, + 506, + 624 + ] + }, + { + "type": "text", + "bbox": [ + 132, + 626, + 286, + 636 + ], + "lines": [ + { + "bbox": [ + 130, + 625, + 288, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 625, + 288, + 638 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45, + "bbox_fs": [ + 130, + 625, + 288, + 638 + ] + }, + { + "type": "text", + "bbox": [ + 147, + 640, + 423, + 663 + ], + "lines": [ + { + "bbox": [ + 146, + 639, + 424, + 651 + ], + "spans": [ + { + "bbox": [ + 146, + 639, + 424, + 651 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 146, + 651, + 398, + 663 + ], + "spans": [ + { + "bbox": [ + 146, + 651, + 398, + 663 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 146, + 639, + 424, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 132, + 666, + 231, + 677 + ], + "lines": [ + { + "bbox": [ + 129, + 664, + 232, + 678 + ], + "spans": [ + { + "bbox": [ + 129, + 664, + 232, + 678 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48, + "bbox_fs": [ + 129, + 664, + 232, + 678 + ] + }, + { + "type": "list", + "bbox": [ + 147, + 680, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 146, + 679, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 146, + 679, + 505, + 691 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental", + "type": "text" + } + ], + "index": 49, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 689, + 392, + 701 + ], + "spans": [ + { + "bbox": [ + 162, + 689, + 392, + 701 + ], + "score": 1.0, + "content": "results (either in the supplemental material or as a URL)? [Yes]", + "type": "text" + } + ], + "index": 50, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 701, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 145, + 701, + 505, + 713 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?", + "type": "text" + } + ], + "index": 51, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 711, + 185, + 723 + ], + "spans": [ + { + "bbox": [ + 161, + 711, + 185, + 723 + ], + "score": 1.0, + "content": "[Yes]", + "type": "text" + } + ], + "index": 52, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 146, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experiments", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 83, + 245, + 95 + ], + "spans": [ + { + "bbox": [ + 162, + 83, + 245, + 95 + ], + "score": 1.0, + "content": "multiple times)? [Yes]", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 95, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 146, + 95, + 507, + 108 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs,", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 105, + 373, + 116 + ], + "spans": [ + { + "bbox": [ + 161, + 105, + 373, + 116 + ], + "score": 1.0, + "content": "internal cluster, or cloud provider)? [Yes] See Section 4.2.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + } + ], + "index": 50.5, + "bbox_fs": [ + 145, + 679, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 147, + 73, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 146, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 146, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experiments", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 162, + 83, + 245, + 95 + ], + "spans": [ + { + "bbox": [ + 162, + 83, + 245, + 95 + ], + "score": 1.0, + "content": "multiple times)? [Yes]", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 146, + 95, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 146, + 95, + 507, + 108 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 105, + 373, + 116 + ], + "spans": [ + { + "bbox": [ + 161, + 105, + 373, + 116 + ], + "score": 1.0, + "content": "internal cluster, or cloud provider)? [Yes] See Section 4.2.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 130, + 119, + 470, + 130 + ], + "lines": [ + { + "bbox": [ + 129, + 118, + 471, + 132 + ], + "spans": [ + { + "bbox": [ + 129, + 118, + 471, + 132 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 147, + 133, + 506, + 212 + ], + "lines": [ + { + "bbox": [ + 146, + 133, + 397, + 145 + ], + "spans": [ + { + "bbox": [ + 146, + 133, + 397, + 145 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes]", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 145, + 340, + 157 + ], + "spans": [ + { + "bbox": [ + 146, + 145, + 340, + 157 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [N/A]", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 147, + 157, + 473, + 168 + ], + "spans": [ + { + "bbox": [ + 147, + 157, + 473, + 168 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [No]", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 147, + 168, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 147, + 168, + 506, + 181 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re us-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 161, + 179, + 236, + 191 + ], + "spans": [ + { + "bbox": [ + 161, + 179, + 236, + 191 + ], + "score": 1.0, + "content": "ing/curating? [N/A]", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 147, + 190, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 147, + 190, + 506, + 203 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable informa-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 161, + 201, + 281, + 212 + ], + "spans": [ + { + "bbox": [ + 161, + 201, + 281, + 212 + ], + "score": 1.0, + "content": "tion or offensive content? [N/A]", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 129, + 215, + 403, + 226 + ], + "lines": [ + { + "bbox": [ + 129, + 213, + 404, + 228 + ], + "spans": [ + { + "bbox": [ + 129, + 213, + 404, + 228 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 147, + 228, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 146, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 146, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable?", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 161, + 238, + 187, + 250 + ], + "spans": [ + { + "bbox": [ + 161, + 238, + 187, + 250 + ], + "score": 1.0, + "content": "[N/A]", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 147, + 250, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 147, + 250, + 506, + 262 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 161, + 261, + 277, + 272 + ], + "spans": [ + { + "bbox": [ + 161, + 261, + 277, + 272 + ], + "score": 1.0, + "content": "approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 148, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 148, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount spent on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 161, + 283, + 284, + 294 + ], + "spans": [ + { + "bbox": [ + 161, + 283, + 284, + 294 + ], + "score": 1.0, + "content": "participant compensation? [N/A]", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 84, + 66, + 102, + 310 + ], + "lines": [ + { + "bbox": [ + 85, + 74, + 100, + 85 + ], + "spans": [ + { + "bbox": [ + 85, + 74, + 100, + 85 + ], + "score": 1.0, + "content": "501", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 84, + 100, + 96 + ], + "spans": [ + { + "bbox": [ + 85, + 84, + 100, + 96 + ], + "score": 1.0, + "content": "502", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 96, + 100, + 107 + ], + "spans": [ + { + "bbox": [ + 85, + 96, + 100, + 107 + ], + "score": 1.0, + "content": "503", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 106, + 100, + 119 + ], + "spans": [ + { + "bbox": [ + 85, + 106, + 100, + 119 + ], + "score": 1.0, + "content": "504", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 120, + 100, + 132 + ], + "spans": [ + { + "bbox": [ + 85, + 120, + 100, + 132 + ], + "score": 1.0, + "content": "505", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 134, + 100, + 146 + ], + "spans": [ + { + "bbox": [ + 85, + 134, + 100, + 146 + ], + "score": 1.0, + "content": "506", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 146, + 100, + 158 + ], + "spans": [ + { + "bbox": [ + 85, + 146, + 100, + 158 + ], + "score": 1.0, + "content": "507", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 158, + 100, + 170 + ], + "spans": [ + { + "bbox": [ + 85, + 158, + 100, + 170 + ], + "score": 1.0, + "content": "508", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 169, + 100, + 181 + ], + "spans": [ + { + "bbox": [ + 85, + 169, + 100, + 181 + ], + "score": 1.0, + "content": "509", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 180, + 100, + 192 + ], + "spans": [ + { + "bbox": [ + 84, + 180, + 100, + 192 + ], + "score": 1.0, + "content": "510", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 191, + 100, + 203 + ], + "spans": [ + { + "bbox": [ + 85, + 191, + 100, + 203 + ], + "score": 1.0, + "content": "511", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 201, + 101, + 214 + ], + "spans": [ + { + "bbox": [ + 84, + 201, + 101, + 214 + ], + "score": 1.0, + "content": "512", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 216, + 100, + 229 + ], + "spans": [ + { + "bbox": [ + 84, + 216, + 100, + 229 + ], + "score": 1.0, + "content": "513", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 229, + 100, + 241 + ], + "spans": [ + { + "bbox": [ + 85, + 229, + 100, + 241 + ], + "score": 1.0, + "content": "514", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 240, + 100, + 252 + ], + "spans": [ + { + "bbox": [ + 84, + 240, + 100, + 252 + ], + "score": 1.0, + "content": "515", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 251, + 100, + 263 + ], + "spans": [ + { + "bbox": [ + 85, + 251, + 100, + 263 + ], + "score": 1.0, + "content": "516", + "type": "text" + } + ] + }, + { + "bbox": [ + 84, + 262, + 100, + 274 + ], + "spans": [ + { + "bbox": [ + 84, + 262, + 100, + 274 + ], + "score": 1.0, + "content": "517", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 273, + 100, + 285 + ], + "spans": [ + { + "bbox": [ + 85, + 273, + 100, + 285 + ], + "score": 1.0, + "content": "518", + "type": "text" + } + ] + }, + { + "bbox": [ + 85, + 283, + 100, + 295 + ], + "spans": [ + { + "bbox": [ + 85, + 283, + 100, + 295 + ], + "score": 1.0, + "content": "519", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 147, + 73, + 506, + 116 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 146, + 73, + 507, + 116 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 130, + 119, + 470, + 130 + ], + "lines": [ + { + "bbox": [ + 129, + 118, + 471, + 132 + ], + "spans": [ + { + "bbox": [ + 129, + 118, + 471, + 132 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 129, + 118, + 471, + 132 + ] + }, + { + "type": "list", + "bbox": [ + 147, + 133, + 506, + 212 + ], + "lines": [ + { + "bbox": [ + 146, + 133, + 397, + 145 + ], + "spans": [ + { + "bbox": [ + 146, + 133, + 397, + 145 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes]", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 145, + 340, + 157 + ], + "spans": [ + { + "bbox": [ + 146, + 145, + 340, + 157 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [N/A]", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 147, + 157, + 473, + 168 + ], + "spans": [ + { + "bbox": [ + 147, + 157, + 473, + 168 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [No]", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 147, + 168, + 506, + 181 + ], + "spans": [ + { + "bbox": [ + 147, + 168, + 506, + 181 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re us-", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 179, + 236, + 191 + ], + "spans": [ + { + "bbox": [ + 161, + 179, + 236, + 191 + ], + "score": 1.0, + "content": "ing/curating? [N/A]", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 147, + 190, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 147, + 190, + 506, + 203 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable informa-", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 201, + 281, + 212 + ], + "spans": [ + { + "bbox": [ + 161, + 201, + 281, + 212 + ], + "score": 1.0, + "content": "tion or offensive content? [N/A]", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + } + ], + "index": 8, + "bbox_fs": [ + 146, + 133, + 506, + 212 + ] + }, + { + "type": "text", + "bbox": [ + 129, + 215, + 403, + 226 + ], + "lines": [ + { + "bbox": [ + 129, + 213, + 404, + 228 + ], + "spans": [ + { + "bbox": [ + 129, + 213, + 404, + 228 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 129, + 213, + 404, + 228 + ] + }, + { + "type": "list", + "bbox": [ + 147, + 228, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 146, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 146, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable?", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 238, + 187, + 250 + ], + "spans": [ + { + "bbox": [ + 161, + 238, + 187, + 250 + ], + "score": 1.0, + "content": "[N/A]", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 147, + 250, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 147, + 250, + 506, + 262 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB)", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 261, + 277, + 272 + ], + "spans": [ + { + "bbox": [ + 161, + 261, + 277, + 272 + ], + "score": 1.0, + "content": "approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 148, + 272, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 148, + 272, + 505, + 284 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount spent on", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 283, + 284, + 294 + ], + "spans": [ + { + "bbox": [ + 161, + 283, + 284, + 294 + ], + "score": 1.0, + "content": "participant compensation? [N/A]", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + } + ], + "index": 15.5, + "bbox_fs": [ + 146, + 227, + 506, + 294 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file