{ "pdf_info": [ { "preproc_blocks": [ { "type": "title", "bbox": [ 147, 78, 463, 96 ], "lines": [ { "bbox": [ 146, 77, 465, 98 ], "spans": [ { "bbox": [ 146, 77, 465, 98 ], "score": 1.0, "content": "VISION TRANSFORMERS NEED REGISTERS", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 113, 113, 497, 127 ], "lines": [ { "bbox": [ 114, 113, 499, 129 ], "spans": [ { "bbox": [ 114, 114, 199, 127 ], "score": 1.0, "content": "Timothee Darcet ´ 1,2", "type": "text" }, { "bbox": [ 224, 114, 299, 129 ], "score": 1.0, "content": "Maxime Oquab1", "type": "text" }, { "bbox": [ 324, 113, 392, 128 ], "score": 1.0, "content": "Julien Mairal2", "type": "text" }, { "bbox": [ 417, 114, 499, 128 ], "score": 1.0, "content": "Piotr Bojanowski1", "type": "text" } ], "index": 1 } ], "index": 1 }, { "type": "text", "bbox": [ 114, 142, 445, 166 ], "lines": [ { "bbox": [ 111, 141, 171, 156 ], "spans": [ { "bbox": [ 111, 141, 171, 156 ], "score": 1.0, "content": "1 FAIR, Meta", "type": "text" } ], "index": 2 }, { "bbox": [ 110, 153, 448, 168 ], "spans": [ { "bbox": [ 110, 153, 448, 168 ], "score": 1.0, "content": "2 Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, 38000 Grenoble, France", "type": "text" } ], "index": 3 } ], "index": 2.5 }, { "type": "title", "bbox": [ 278, 195, 333, 207 ], "lines": [ { "bbox": [ 277, 195, 335, 208 ], "spans": [ { "bbox": [ 277, 195, 335, 208 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 142, 219, 469, 340 ], "lines": [ { "bbox": [ 141, 219, 469, 233 ], "spans": [ { "bbox": [ 141, 219, 469, 233 ], "score": 1.0, "content": "Transformers have recently emerged as a powerful tool for learning visual rep-", "type": "text" } ], "index": 5 }, { "bbox": [ 141, 231, 469, 244 ], "spans": [ { "bbox": [ 141, 231, 469, 244 ], "score": 1.0, "content": "resentations. In this paper, we identify and characterize artifacts in feature maps", "type": "text" } ], "index": 6 }, { "bbox": [ 142, 242, 469, 254 ], "spans": [ { "bbox": [ 142, 242, 469, 254 ], "score": 1.0, "content": "of both supervised and self-supervised ViT networks. The artifacts correspond to", "type": "text" } ], "index": 7 }, { "bbox": [ 141, 252, 470, 265 ], "spans": [ { "bbox": [ 141, 252, 470, 265 ], "score": 1.0, "content": "high-norm tokens appearing during inference primarily in low-informative back-", "type": "text" } ], "index": 8 }, { "bbox": [ 142, 264, 469, 276 ], "spans": [ { "bbox": [ 142, 264, 469, 276 ], "score": 1.0, "content": "ground areas of images, that are repurposed for internal computations. We propose", "type": "text" } ], "index": 9 }, { "bbox": [ 141, 275, 470, 287 ], "spans": [ { "bbox": [ 141, 275, 470, 287 ], "score": 1.0, "content": "a simple yet effective solution based on providing additional tokens to the input se-", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 286, 470, 298 ], "spans": [ { "bbox": [ 141, 286, 470, 298 ], "score": 1.0, "content": "quence of the Vision Transformer to fill that role. We show that this solution fixes", "type": "text" } ], "index": 11 }, { "bbox": [ 142, 297, 469, 308 ], "spans": [ { "bbox": [ 142, 297, 469, 308 ], "score": 1.0, "content": "that problem entirely for both supervised and self-supervised models, sets a new", "type": "text" } ], "index": 12 }, { "bbox": [ 142, 307, 469, 320 ], "spans": [ { "bbox": [ 142, 307, 469, 320 ], "score": 1.0, "content": "state of the art for self-supervised visual models on dense visual prediction tasks,", "type": "text" } ], "index": 13 }, { "bbox": [ 142, 319, 469, 330 ], "spans": [ { "bbox": [ 142, 319, 469, 330 ], "score": 1.0, "content": "enables object discovery methods with larger models, and most importantly leads", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 329, 462, 343 ], "spans": [ { "bbox": [ 141, 329, 462, 343 ], "score": 1.0, "content": "to smoother feature maps and attention maps for downstream visual processing.", "type": "text" } ], "index": 15 } ], "index": 10 }, { "type": "image", "bbox": [ 109, 358, 508, 536 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 358, 508, 536 ], "group_id": 0, "lines": [ { "bbox": [ 109, 358, 508, 536 ], "spans": [ { "bbox": [ 109, 358, 508, 536 ], "score": 0.966, "type": "image", "image_path": "5618f57fc336a126451f7196bd7c0dada9ee3c6f8862d5badc23645a5236e959.jpg" } ] } ], "index": 17, "virtual_lines": [ { "bbox": [ 109, 358, 508, 417.3333333333333 ], "spans": [], "index": 16 }, { "bbox": [ 109, 417.3333333333333, 508, 476.66666666666663 ], "spans": [], "index": 17 }, { "bbox": [ 109, 476.66666666666663, 508, 536.0 ], "spans": [], "index": 18 } ] }, { "type": "image_caption", "bbox": [ 106, 546, 506, 579 ], "group_id": 0, "lines": [ { "bbox": [ 105, 546, 506, 559 ], "spans": [ { "bbox": [ 105, 546, 506, 559 ], "score": 1.0, "content": "Figure 1: Register tokens enable interpretable attention maps in all vision transformers, similar to", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 558, 506, 570 ], "spans": [ { "bbox": [ 106, 558, 506, 570 ], "score": 1.0, "content": "the original DINO method (Caron et al., 2021). Attention maps are calculated in high resolution for", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 568, 398, 581 ], "spans": [ { "bbox": [ 106, 568, 398, 581 ], "score": 1.0, "content": "better visualisation. More qualitative results are available in appendix H.", "type": "text" } ], "index": 21 } ], "index": 20 } ], "index": 18.5 }, { "type": "title", "bbox": [ 108, 602, 206, 614 ], "lines": [ { "bbox": [ 105, 601, 208, 617 ], "spans": [ { "bbox": [ 105, 601, 208, 617 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 627, 505, 715 ], "lines": [ { "bbox": [ 105, 627, 505, 639 ], "spans": [ { "bbox": [ 105, 627, 505, 639 ], "score": 1.0, "content": "Embedding images into generic features that can serve multiple purposes in computer vision has", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 638, 506, 650 ], "spans": [ { "bbox": [ 105, 638, 506, 650 ], "score": 1.0, "content": "been a long-standing problem. First methods relied on handcrafted principles, such as SIFT (Lowe,", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 648, 505, 661 ], "spans": [ { "bbox": [ 105, 648, 505, 661 ], "score": 1.0, "content": "2004), before the scale of data and deep learning techniques allowed for end-to-end training. Pur-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 660, 506, 672 ], "spans": [ { "bbox": [ 105, 660, 506, 672 ], "score": 1.0, "content": "suing generic feature embeddings is still relevant today, as collecting valuable annotated data for", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 669, 506, 685 ], "spans": [ { "bbox": [ 105, 669, 506, 685 ], "score": 1.0, "content": "many specific tasks remains difficult. This difficulty arises because of the required expertise (e.g.,", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 682, 506, 694 ], "spans": [ { "bbox": [ 105, 682, 506, 694 ], "score": 1.0, "content": "medical data, or remote sensing) or the cost at scale. Today, it is common to pretrain a model for", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 693, 506, 705 ], "spans": [ { "bbox": [ 105, 693, 506, 705 ], "score": 1.0, "content": "a task for which plenty of data is available and extract a subset of the model to use as a feature", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 704, 505, 716 ], "spans": [ { "bbox": [ 105, 704, 505, 716 ], "score": 1.0, "content": "extractor. Multiple approaches offer this possibility; supervised methods, building on classification", "type": "text" } ], "index": 30 } ], "index": 26.5 } ], "page_idx": 0, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 123, 723, 291, 732 ], "lines": [ { "bbox": [ 123, 721, 293, 733 ], "spans": [ { "bbox": [ 123, 721, 293, 733 ], "score": 1.0, "content": "Correspondence to timdarcet@meta.com", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 303, 752, 309, 762 ], "spans": [ { "bbox": [ 303, 752, 309, 762 ], "score": 1.0, "content": "1", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 147, 78, 463, 96 ], "lines": [ { "bbox": [ 146, 77, 465, 98 ], "spans": [ { "bbox": [ 146, 77, 465, 98 ], "score": 1.0, "content": "VISION TRANSFORMERS NEED REGISTERS", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 113, 113, 497, 127 ], "lines": [ { "bbox": [ 114, 113, 499, 129 ], "spans": [ { "bbox": [ 114, 114, 199, 127 ], "score": 1.0, "content": "Timothee Darcet ´ 1,2", "type": "text" }, { "bbox": [ 224, 114, 299, 129 ], "score": 1.0, "content": "Maxime Oquab1", "type": "text" }, { "bbox": [ 324, 113, 392, 128 ], "score": 1.0, "content": "Julien Mairal2", "type": "text" }, { "bbox": [ 417, 114, 499, 128 ], "score": 1.0, "content": "Piotr Bojanowski1", "type": "text" } ], "index": 1 } ], "index": 1, "bbox_fs": [ 114, 113, 499, 129 ] }, { "type": "index", "bbox": [ 114, 142, 445, 166 ], "lines": [ { "bbox": [ 111, 141, 171, 156 ], "spans": [ { "bbox": [ 111, 141, 171, 156 ], "score": 1.0, "content": "1 FAIR, Meta", "type": "text" } ], "index": 2, "is_list_start_line": true }, { "bbox": [ 110, 153, 448, 168 ], "spans": [ { "bbox": [ 110, 153, 448, 168 ], "score": 1.0, "content": "2 Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, 38000 Grenoble, France", "type": "text" } ], "index": 3, "is_list_start_line": true } ], "index": 2.5, "bbox_fs": [ 110, 141, 448, 168 ] }, { "type": "title", "bbox": [ 278, 195, 333, 207 ], "lines": [ { "bbox": [ 277, 195, 335, 208 ], "spans": [ { "bbox": [ 277, 195, 335, 208 ], "score": 1.0, "content": "ABSTRACT", "type": "text" } ], "index": 4 } ], "index": 4 }, { "type": "text", "bbox": [ 142, 219, 469, 340 ], "lines": [ { "bbox": [ 141, 219, 469, 233 ], "spans": [ { "bbox": [ 141, 219, 469, 233 ], "score": 1.0, "content": "Transformers have recently emerged as a powerful tool for learning visual rep-", "type": "text" } ], "index": 5 }, { "bbox": [ 141, 231, 469, 244 ], "spans": [ { "bbox": [ 141, 231, 469, 244 ], "score": 1.0, "content": "resentations. In this paper, we identify and characterize artifacts in feature maps", "type": "text" } ], "index": 6 }, { "bbox": [ 142, 242, 469, 254 ], "spans": [ { "bbox": [ 142, 242, 469, 254 ], "score": 1.0, "content": "of both supervised and self-supervised ViT networks. The artifacts correspond to", "type": "text" } ], "index": 7 }, { "bbox": [ 141, 252, 470, 265 ], "spans": [ { "bbox": [ 141, 252, 470, 265 ], "score": 1.0, "content": "high-norm tokens appearing during inference primarily in low-informative back-", "type": "text" } ], "index": 8 }, { "bbox": [ 142, 264, 469, 276 ], "spans": [ { "bbox": [ 142, 264, 469, 276 ], "score": 1.0, "content": "ground areas of images, that are repurposed for internal computations. We propose", "type": "text" } ], "index": 9 }, { "bbox": [ 141, 275, 470, 287 ], "spans": [ { "bbox": [ 141, 275, 470, 287 ], "score": 1.0, "content": "a simple yet effective solution based on providing additional tokens to the input se-", "type": "text" } ], "index": 10 }, { "bbox": [ 141, 286, 470, 298 ], "spans": [ { "bbox": [ 141, 286, 470, 298 ], "score": 1.0, "content": "quence of the Vision Transformer to fill that role. We show that this solution fixes", "type": "text" } ], "index": 11 }, { "bbox": [ 142, 297, 469, 308 ], "spans": [ { "bbox": [ 142, 297, 469, 308 ], "score": 1.0, "content": "that problem entirely for both supervised and self-supervised models, sets a new", "type": "text" } ], "index": 12 }, { "bbox": [ 142, 307, 469, 320 ], "spans": [ { "bbox": [ 142, 307, 469, 320 ], "score": 1.0, "content": "state of the art for self-supervised visual models on dense visual prediction tasks,", "type": "text" } ], "index": 13 }, { "bbox": [ 142, 319, 469, 330 ], "spans": [ { "bbox": [ 142, 319, 469, 330 ], "score": 1.0, "content": "enables object discovery methods with larger models, and most importantly leads", "type": "text" } ], "index": 14 }, { "bbox": [ 141, 329, 462, 343 ], "spans": [ { "bbox": [ 141, 329, 462, 343 ], "score": 1.0, "content": "to smoother feature maps and attention maps for downstream visual processing.", "type": "text" } ], "index": 15 } ], "index": 10, "bbox_fs": [ 141, 219, 470, 343 ] }, { "type": "image", "bbox": [ 109, 358, 508, 536 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 358, 508, 536 ], "group_id": 0, "lines": [ { "bbox": [ 109, 358, 508, 536 ], "spans": [ { "bbox": [ 109, 358, 508, 536 ], "score": 0.966, "type": "image", "image_path": "5618f57fc336a126451f7196bd7c0dada9ee3c6f8862d5badc23645a5236e959.jpg" } ] } ], "index": 17, "virtual_lines": [ { "bbox": [ 109, 358, 508, 417.3333333333333 ], "spans": [], "index": 16 }, { "bbox": [ 109, 417.3333333333333, 508, 476.66666666666663 ], "spans": [], "index": 17 }, { "bbox": [ 109, 476.66666666666663, 508, 536.0 ], "spans": [], "index": 18 } ] }, { "type": "image_caption", "bbox": [ 106, 546, 506, 579 ], "group_id": 0, "lines": [ { "bbox": [ 105, 546, 506, 559 ], "spans": [ { "bbox": [ 105, 546, 506, 559 ], "score": 1.0, "content": "Figure 1: Register tokens enable interpretable attention maps in all vision transformers, similar to", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 558, 506, 570 ], "spans": [ { "bbox": [ 106, 558, 506, 570 ], "score": 1.0, "content": "the original DINO method (Caron et al., 2021). Attention maps are calculated in high resolution for", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 568, 398, 581 ], "spans": [ { "bbox": [ 106, 568, 398, 581 ], "score": 1.0, "content": "better visualisation. More qualitative results are available in appendix H.", "type": "text" } ], "index": 21 } ], "index": 20 } ], "index": 18.5 }, { "type": "title", "bbox": [ 108, 602, 206, 614 ], "lines": [ { "bbox": [ 105, 601, 208, 617 ], "spans": [ { "bbox": [ 105, 601, 208, 617 ], "score": 1.0, "content": "1 INTRODUCTION", "type": "text" } ], "index": 22 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 627, 505, 715 ], "lines": [ { "bbox": [ 105, 627, 505, 639 ], "spans": [ { "bbox": [ 105, 627, 505, 639 ], "score": 1.0, "content": "Embedding images into generic features that can serve multiple purposes in computer vision has", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 638, 506, 650 ], "spans": [ { "bbox": [ 105, 638, 506, 650 ], "score": 1.0, "content": "been a long-standing problem. First methods relied on handcrafted principles, such as SIFT (Lowe,", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 648, 505, 661 ], "spans": [ { "bbox": [ 105, 648, 505, 661 ], "score": 1.0, "content": "2004), before the scale of data and deep learning techniques allowed for end-to-end training. Pur-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 660, 506, 672 ], "spans": [ { "bbox": [ 105, 660, 506, 672 ], "score": 1.0, "content": "suing generic feature embeddings is still relevant today, as collecting valuable annotated data for", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 669, 506, 685 ], "spans": [ { "bbox": [ 105, 669, 506, 685 ], "score": 1.0, "content": "many specific tasks remains difficult. This difficulty arises because of the required expertise (e.g.,", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 682, 506, 694 ], "spans": [ { "bbox": [ 105, 682, 506, 694 ], "score": 1.0, "content": "medical data, or remote sensing) or the cost at scale. Today, it is common to pretrain a model for", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 693, 506, 705 ], "spans": [ { "bbox": [ 105, 693, 506, 705 ], "score": 1.0, "content": "a task for which plenty of data is available and extract a subset of the model to use as a feature", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 704, 505, 716 ], "spans": [ { "bbox": [ 105, 704, 505, 716 ], "score": 1.0, "content": "extractor. Multiple approaches offer this possibility; supervised methods, building on classification", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 379, 505, 391 ], "spans": [ { "bbox": [ 105, 379, 505, 391 ], "score": 1.0, "content": "or text-image alignment, allow training strong feature models to unlock downstream tasks. Alterna-", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 106, 390, 505, 402 ], "spans": [ { "bbox": [ 106, 390, 505, 402 ], "score": 1.0, "content": "tively, self-supervised methods building on the Transformer architecture have attracted significant", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 105, 401, 506, 414 ], "spans": [ { "bbox": [ 105, 401, 506, 414 ], "score": 1.0, "content": "attention due to their high prediction performance on downstream tasks and the intriguing ability of", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 105, 412, 400, 424 ], "spans": [ { "bbox": [ 105, 412, 400, 424 ], "score": 1.0, "content": "some models to provide unsupervised segmentations (Caron et al., 2021)", "type": "text", "cross_page": true } ], "index": 10 } ], "index": 26.5, "bbox_fs": [ 105, 627, 506, 716 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 110, 76, 501, 301 ], "blocks": [ { "type": "image_body", "bbox": [ 110, 76, 501, 301 ], "group_id": 0, "lines": [ { "bbox": [ 110, 76, 501, 301 ], "spans": [ { "bbox": [ 110, 76, 501, 301 ], "score": 0.975, "type": "image", "image_path": "29d5e019111549c93558ac7ee27d51fe01fbdce517916b5e61cf291389ad7f7e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 110, 76, 501, 151.0 ], "spans": [], "index": 0 }, { "bbox": [ 110, 151.0, 501, 226.0 ], "spans": [], "index": 1 }, { "bbox": [ 110, 226.0, 501, 301.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 312, 504, 356 ], "group_id": 0, "lines": [ { "bbox": [ 105, 311, 506, 325 ], "spans": [ { "bbox": [ 105, 311, 506, 325 ], "score": 1.0, "content": "Figure 2: Illustration of artifacts observed in the attention maps of modern vision transformers.", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 322, 506, 335 ], "spans": [ { "bbox": [ 105, 322, 506, 335 ], "score": 1.0, "content": "We consider ViTs trained with label supervision (DeiT-III), text-supervision (OpenCLIP) or self-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 334, 506, 347 ], "spans": [ { "bbox": [ 105, 334, 506, 347 ], "score": 1.0, "content": "supervision (DINO and DINOv2). Interestingly, all models but DINO exhibit peaky outlier values", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 344, 475, 358 ], "spans": [ { "bbox": [ 105, 344, 475, 358 ], "score": 1.0, "content": "in the attention maps. The goal of this work is to understand and mitigate this phenomenon.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "text", "bbox": [ 107, 379, 505, 423 ], "lines": [ { "bbox": [ 105, 379, 505, 391 ], "spans": [ { "bbox": [ 105, 379, 505, 391 ], "score": 1.0, "content": "or text-image alignment, allow training strong feature models to unlock downstream tasks. Alterna-", "type": "text" } ], "index": 7 }, { "bbox": [ 106, 390, 505, 402 ], "spans": [ { "bbox": [ 106, 390, 505, 402 ], "score": 1.0, "content": "tively, self-supervised methods building on the Transformer architecture have attracted significant", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 401, 506, 414 ], "spans": [ { "bbox": [ 105, 401, 506, 414 ], "score": 1.0, "content": "attention due to their high prediction performance on downstream tasks and the intriguing ability of", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 412, 400, 424 ], "spans": [ { "bbox": [ 105, 412, 400, 424 ], "score": 1.0, "content": "some models to provide unsupervised segmentations (Caron et al., 2021)", "type": "text" } ], "index": 10 } ], "index": 8.5 }, { "type": "text", "bbox": [ 107, 429, 505, 495 ], "lines": [ { "bbox": [ 106, 429, 505, 441 ], "spans": [ { "bbox": [ 106, 429, 505, 441 ], "score": 1.0, "content": "In particular, the DINO algorithm is shown to produce models that contain explicit information about", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 440, 504, 452 ], "spans": [ { "bbox": [ 106, 440, 504, 452 ], "score": 1.0, "content": "the semantic layout of an image. Indeed, qualitative results show that the last attention layer naturally", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 452, 505, 463 ], "spans": [ { "bbox": [ 106, 452, 505, 463 ], "score": 1.0, "content": "focuses on semantically consistent parts of images and often produces interpretable attention maps.", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 462, 505, 474 ], "spans": [ { "bbox": [ 106, 462, 505, 474 ], "score": 1.0, "content": "Exploiting these properties, object discovery algorithms such as LOST (Simeoni et al., 2021) build ´", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 473, 505, 485 ], "spans": [ { "bbox": [ 105, 473, 505, 485 ], "score": 1.0, "content": "on top of DINO. Such algorithms can detect objects without supervision by gathering information", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 484, 442, 496 ], "spans": [ { "bbox": [ 106, 484, 442, 496 ], "score": 1.0, "content": "in attention maps. They are effectively unlocking a new frontier in computer vision.", "type": "text" } ], "index": 16 } ], "index": 13.5 }, { "type": "text", "bbox": [ 107, 500, 505, 621 ], "lines": [ { "bbox": [ 105, 500, 505, 513 ], "spans": [ { "bbox": [ 105, 500, 505, 513 ], "score": 1.0, "content": "DINOv2 (Oquab et al., 2023), a follow-up to DINO, provides features that allow tackling dense", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 510, 505, 525 ], "spans": [ { "bbox": [ 105, 510, 505, 525 ], "score": 1.0, "content": "prediction tasks. DINOv2 features lead to successful monocular depth estimation and semantic seg-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 522, 505, 534 ], "spans": [ { "bbox": [ 105, 522, 505, 534 ], "score": 1.0, "content": "mentation with a frozen backbone and linear models. Despite the strong performance on dense tasks,", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 533, 505, 546 ], "spans": [ { "bbox": [ 105, 533, 505, 546 ], "score": 1.0, "content": "we observed that DINOv2 is surprisingly incompatible with LOST. When used to extract features, it", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 544, 506, 558 ], "spans": [ { "bbox": [ 105, 544, 506, 558 ], "score": 1.0, "content": "delivers disappointing performance, only on par with supervised alternative backbones in this sce-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 555, 505, 568 ], "spans": [ { "bbox": [ 105, 555, 505, 568 ], "score": 1.0, "content": "nario. This suggests that DINOv2 behaves differently than DINO. The investigation described in", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 567, 506, 579 ], "spans": [ { "bbox": [ 105, 567, 506, 579 ], "score": 1.0, "content": "this work notably exposes the presence of artefacts in the feature maps of DINOv2 that were not", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 577, 506, 590 ], "spans": [ { "bbox": [ 105, 577, 506, 590 ], "score": 1.0, "content": "present in the first version of this model. These are observable qualitatively using straightforward", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 589, 505, 601 ], "spans": [ { "bbox": [ 105, 589, 505, 601 ], "score": 1.0, "content": "methods. Also surprisingly, applying the same observations to supervised vision transformers ex-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 600, 505, 612 ], "spans": [ { "bbox": [ 105, 600, 505, 612 ], "score": 1.0, "content": "poses similar artifacts, as shown in Fig. 2. This suggests that DINO is, in fact, an exception, while", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 610, 382, 622 ], "spans": [ { "bbox": [ 106, 610, 382, 622 ], "score": 1.0, "content": "DINOv2 models match the baseline behavior of vision transformers.", "type": "text" } ], "index": 27 } ], "index": 22 }, { "type": "text", "bbox": [ 107, 627, 504, 693 ], "lines": [ { "bbox": [ 105, 626, 506, 640 ], "spans": [ { "bbox": [ 105, 626, 506, 640 ], "score": 1.0, "content": "In this work, we set out to better understand this phenomenon and develop methods to detect these", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 637, 506, 650 ], "spans": [ { "bbox": [ 105, 637, 324, 650 ], "score": 1.0, "content": "artifacts. We observe that they are tokens with roughly", "type": "text" }, { "bbox": [ 324, 639, 340, 648 ], "score": 0.41, "content": "1 0 \\mathrm { x }", "type": "inline_equation" }, { "bbox": [ 340, 637, 506, 650 ], "score": 1.0, "content": "higher norm at the output and correspond", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 648, 506, 662 ], "spans": [ { "bbox": [ 105, 648, 298, 662 ], "score": 1.0, "content": "to a small fraction of the total sequence (around", "type": "text" }, { "bbox": [ 298, 649, 313, 659 ], "score": 0.83, "content": "2 \\%", "type": "inline_equation" }, { "bbox": [ 313, 648, 506, 662 ], "score": 1.0, "content": "). We also show that these tokens appear around", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 658, 506, 674 ], "spans": [ { "bbox": [ 105, 658, 506, 674 ], "score": 1.0, "content": "the middle layers of the vision transformer, and that they only appear after a sufficiently long training", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 670, 506, 684 ], "spans": [ { "bbox": [ 105, 670, 506, 684 ], "score": 1.0, "content": "of a sufficiently big transformer. In particular, we show that these outlier tokens appear in patches", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 682, 444, 694 ], "spans": [ { "bbox": [ 106, 682, 444, 694 ], "score": 1.0, "content": "similar to their neighbors, meaning patches that convey little additional information.", "type": "text" } ], "index": 33 } ], "index": 30.5 }, { "type": "text", "bbox": [ 107, 699, 504, 732 ], "lines": [ { "bbox": [ 105, 698, 506, 711 ], "spans": [ { "bbox": [ 105, 698, 506, 711 ], "score": 1.0, "content": "As part of our investigation, we evaluate the outlier tokens with simple linear models to under-", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "stand the information they contain. We observe that, compared to non-outlier tokens, they hold less", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 720, 505, 734 ], "spans": [ { "bbox": [ 105, 720, 505, 734 ], "score": 1.0, "content": "information about their original position in the image or the original pixels in their patch. This ob-", "type": "text" } ], "index": 36 } ], "index": 35 } ], "page_idx": 1, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 303, 752, 308, 760 ], "lines": [] } ], "para_blocks": [ { "type": "image", "bbox": [ 110, 76, 501, 301 ], "blocks": [ { "type": "image_body", "bbox": [ 110, 76, 501, 301 ], "group_id": 0, "lines": [ { "bbox": [ 110, 76, 501, 301 ], "spans": [ { "bbox": [ 110, 76, 501, 301 ], "score": 0.975, "type": "image", "image_path": "29d5e019111549c93558ac7ee27d51fe01fbdce517916b5e61cf291389ad7f7e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 110, 76, 501, 151.0 ], "spans": [], "index": 0 }, { "bbox": [ 110, 151.0, 501, 226.0 ], "spans": [], "index": 1 }, { "bbox": [ 110, 226.0, 501, 301.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 107, 312, 504, 356 ], "group_id": 0, "lines": [ { "bbox": [ 105, 311, 506, 325 ], "spans": [ { "bbox": [ 105, 311, 506, 325 ], "score": 1.0, "content": "Figure 2: Illustration of artifacts observed in the attention maps of modern vision transformers.", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 322, 506, 335 ], "spans": [ { "bbox": [ 105, 322, 506, 335 ], "score": 1.0, "content": "We consider ViTs trained with label supervision (DeiT-III), text-supervision (OpenCLIP) or self-", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 334, 506, 347 ], "spans": [ { "bbox": [ 105, 334, 506, 347 ], "score": 1.0, "content": "supervision (DINO and DINOv2). Interestingly, all models but DINO exhibit peaky outlier values", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 344, 475, 358 ], "spans": [ { "bbox": [ 105, 344, 475, 358 ], "score": 1.0, "content": "in the attention maps. The goal of this work is to understand and mitigate this phenomenon.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "text", "bbox": [ 107, 379, 505, 423 ], "lines": [], "index": 8.5, "bbox_fs": [ 105, 379, 506, 424 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 429, 505, 495 ], "lines": [ { "bbox": [ 106, 429, 505, 441 ], "spans": [ { "bbox": [ 106, 429, 505, 441 ], "score": 1.0, "content": "In particular, the DINO algorithm is shown to produce models that contain explicit information about", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 440, 504, 452 ], "spans": [ { "bbox": [ 106, 440, 504, 452 ], "score": 1.0, "content": "the semantic layout of an image. Indeed, qualitative results show that the last attention layer naturally", "type": "text" } ], "index": 12 }, { "bbox": [ 106, 452, 505, 463 ], "spans": [ { "bbox": [ 106, 452, 505, 463 ], "score": 1.0, "content": "focuses on semantically consistent parts of images and often produces interpretable attention maps.", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 462, 505, 474 ], "spans": [ { "bbox": [ 106, 462, 505, 474 ], "score": 1.0, "content": "Exploiting these properties, object discovery algorithms such as LOST (Simeoni et al., 2021) build ´", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 473, 505, 485 ], "spans": [ { "bbox": [ 105, 473, 505, 485 ], "score": 1.0, "content": "on top of DINO. Such algorithms can detect objects without supervision by gathering information", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 484, 442, 496 ], "spans": [ { "bbox": [ 106, 484, 442, 496 ], "score": 1.0, "content": "in attention maps. They are effectively unlocking a new frontier in computer vision.", "type": "text" } ], "index": 16 } ], "index": 13.5, "bbox_fs": [ 105, 429, 505, 496 ] }, { "type": "text", "bbox": [ 107, 500, 505, 621 ], "lines": [ { "bbox": [ 105, 500, 505, 513 ], "spans": [ { "bbox": [ 105, 500, 505, 513 ], "score": 1.0, "content": "DINOv2 (Oquab et al., 2023), a follow-up to DINO, provides features that allow tackling dense", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 510, 505, 525 ], "spans": [ { "bbox": [ 105, 510, 505, 525 ], "score": 1.0, "content": "prediction tasks. DINOv2 features lead to successful monocular depth estimation and semantic seg-", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 522, 505, 534 ], "spans": [ { "bbox": [ 105, 522, 505, 534 ], "score": 1.0, "content": "mentation with a frozen backbone and linear models. Despite the strong performance on dense tasks,", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 533, 505, 546 ], "spans": [ { "bbox": [ 105, 533, 505, 546 ], "score": 1.0, "content": "we observed that DINOv2 is surprisingly incompatible with LOST. When used to extract features, it", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 544, 506, 558 ], "spans": [ { "bbox": [ 105, 544, 506, 558 ], "score": 1.0, "content": "delivers disappointing performance, only on par with supervised alternative backbones in this sce-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 555, 505, 568 ], "spans": [ { "bbox": [ 105, 555, 505, 568 ], "score": 1.0, "content": "nario. This suggests that DINOv2 behaves differently than DINO. The investigation described in", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 567, 506, 579 ], "spans": [ { "bbox": [ 105, 567, 506, 579 ], "score": 1.0, "content": "this work notably exposes the presence of artefacts in the feature maps of DINOv2 that were not", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 577, 506, 590 ], "spans": [ { "bbox": [ 105, 577, 506, 590 ], "score": 1.0, "content": "present in the first version of this model. These are observable qualitatively using straightforward", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 589, 505, 601 ], "spans": [ { "bbox": [ 105, 589, 505, 601 ], "score": 1.0, "content": "methods. Also surprisingly, applying the same observations to supervised vision transformers ex-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 600, 505, 612 ], "spans": [ { "bbox": [ 105, 600, 505, 612 ], "score": 1.0, "content": "poses similar artifacts, as shown in Fig. 2. This suggests that DINO is, in fact, an exception, while", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 610, 382, 622 ], "spans": [ { "bbox": [ 106, 610, 382, 622 ], "score": 1.0, "content": "DINOv2 models match the baseline behavior of vision transformers.", "type": "text" } ], "index": 27 } ], "index": 22, "bbox_fs": [ 105, 500, 506, 622 ] }, { "type": "text", "bbox": [ 107, 627, 504, 693 ], "lines": [ { "bbox": [ 105, 626, 506, 640 ], "spans": [ { "bbox": [ 105, 626, 506, 640 ], "score": 1.0, "content": "In this work, we set out to better understand this phenomenon and develop methods to detect these", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 637, 506, 650 ], "spans": [ { "bbox": [ 105, 637, 324, 650 ], "score": 1.0, "content": "artifacts. We observe that they are tokens with roughly", "type": "text" }, { "bbox": [ 324, 639, 340, 648 ], "score": 0.41, "content": "1 0 \\mathrm { x }", "type": "inline_equation" }, { "bbox": [ 340, 637, 506, 650 ], "score": 1.0, "content": "higher norm at the output and correspond", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 648, 506, 662 ], "spans": [ { "bbox": [ 105, 648, 298, 662 ], "score": 1.0, "content": "to a small fraction of the total sequence (around", "type": "text" }, { "bbox": [ 298, 649, 313, 659 ], "score": 0.83, "content": "2 \\%", "type": "inline_equation" }, { "bbox": [ 313, 648, 506, 662 ], "score": 1.0, "content": "). We also show that these tokens appear around", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 658, 506, 674 ], "spans": [ { "bbox": [ 105, 658, 506, 674 ], "score": 1.0, "content": "the middle layers of the vision transformer, and that they only appear after a sufficiently long training", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 670, 506, 684 ], "spans": [ { "bbox": [ 105, 670, 506, 684 ], "score": 1.0, "content": "of a sufficiently big transformer. In particular, we show that these outlier tokens appear in patches", "type": "text" } ], "index": 32 }, { "bbox": [ 106, 682, 444, 694 ], "spans": [ { "bbox": [ 106, 682, 444, 694 ], "score": 1.0, "content": "similar to their neighbors, meaning patches that convey little additional information.", "type": "text" } ], "index": 33 } ], "index": 30.5, "bbox_fs": [ 105, 626, 506, 694 ] }, { "type": "text", "bbox": [ 107, 699, 504, 732 ], "lines": [ { "bbox": [ 105, 698, 506, 711 ], "spans": [ { "bbox": [ 105, 698, 506, 711 ], "score": 1.0, "content": "As part of our investigation, we evaluate the outlier tokens with simple linear models to under-", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 710, 505, 722 ], "spans": [ { "bbox": [ 106, 710, 505, 722 ], "score": 1.0, "content": "stand the information they contain. We observe that, compared to non-outlier tokens, they hold less", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 720, 505, 734 ], "spans": [ { "bbox": [ 105, 720, 505, 734 ], "score": 1.0, "content": "information about their original position in the image or the original pixels in their patch. This ob-", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 235, 506, 250 ], "spans": [ { "bbox": [ 105, 235, 506, 250 ], "score": 1.0, "content": "servation suggests that the model discards the local information contained in these patches during", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 105, 247, 505, 261 ], "spans": [ { "bbox": [ 105, 247, 505, 261 ], "score": 1.0, "content": "inference. On the other hand, learning an image classifier on outlier patches yields significantly", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 105, 258, 505, 271 ], "spans": [ { "bbox": [ 105, 258, 505, 271 ], "score": 1.0, "content": "stronger accuracy than doing so on the other patches, suggesting that they contain global informa-", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 105, 269, 506, 283 ], "spans": [ { "bbox": [ 105, 269, 506, 283 ], "score": 1.0, "content": "tion about the image. We propose the following interpretation to these elements: the model learns", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 105, 281, 506, 293 ], "spans": [ { "bbox": [ 105, 281, 506, 293 ], "score": 1.0, "content": "to recognize patches containing little useful information, and recycle the corresponding tokens to", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 105, 291, 401, 305 ], "spans": [ { "bbox": [ 105, 291, 401, 305 ], "score": 1.0, "content": "aggregate global image information while discarding spatial information.", "type": "text", "cross_page": true } ], "index": 12 } ], "index": 35, "bbox_fs": [ 105, 698, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 108, 78, 509, 159 ], "blocks": [ { "type": "image_body", "bbox": [ 108, 78, 509, 159 ], "group_id": 0, "lines": [ { "bbox": [ 108, 78, 509, 159 ], "spans": [ { "bbox": [ 108, 78, 509, 159 ], "score": 0.963, "type": "image", "image_path": "fe26196bcac0e507f5fee2e2dc92dd8ad384ef11ab92073f969917b083f64c7e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 108, 78, 509, 105.0 ], "spans": [], "index": 0 }, { "bbox": [ 108, 105.0, 509, 132.0 ], "spans": [], "index": 1 }, { "bbox": [ 108, 132.0, 509, 159.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 169, 505, 214 ], "group_id": 0, "lines": [ { "bbox": [ 105, 169, 505, 181 ], "spans": [ { "bbox": [ 105, 169, 448, 181 ], "score": 1.0, "content": "Figure 3: Comparison of local feature norms for DINO ViT-B/16 and DINOv2 ViT-", "type": "text" }, { "bbox": [ 448, 169, 468, 181 ], "score": 0.43, "content": "\\mathrm { \\ g } / 1 4", "type": "inline_equation" }, { "bbox": [ 468, 169, 505, 181 ], "score": 1.0, "content": ". We ob-", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 180, 505, 192 ], "spans": [ { "bbox": [ 105, 180, 505, 192 ], "score": 1.0, "content": "serve that DINOv2 has a few outlier patches, whereas DINO does not present these artifacts. For", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 192, 505, 203 ], "spans": [ { "bbox": [ 105, 192, 505, 203 ], "score": 1.0, "content": "DINOv2, although most patch tokens have a norm between 0 and 100, a small proportion of tokens", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 202, 498, 214 ], "spans": [ { "bbox": [ 105, 203, 467, 214 ], "score": 1.0, "content": "have a very high norm. We measure the proportion of tokens with norm larger than 150 at", "type": "text" }, { "bbox": [ 467, 202, 494, 213 ], "score": 0.88, "content": "2 . 3 7 \\%", "type": "inline_equation" }, { "bbox": [ 494, 203, 498, 214 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "text", "bbox": [ 107, 236, 505, 303 ], "lines": [ { "bbox": [ 105, 235, 506, 250 ], "spans": [ { "bbox": [ 105, 235, 506, 250 ], "score": 1.0, "content": "servation suggests that the model discards the local information contained in these patches during", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 247, 505, 261 ], "spans": [ { "bbox": [ 105, 247, 505, 261 ], "score": 1.0, "content": "inference. On the other hand, learning an image classifier on outlier patches yields significantly", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 258, 505, 271 ], "spans": [ { "bbox": [ 105, 258, 505, 271 ], "score": 1.0, "content": "stronger accuracy than doing so on the other patches, suggesting that they contain global informa-", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 269, 506, 283 ], "spans": [ { "bbox": [ 105, 269, 506, 283 ], "score": 1.0, "content": "tion about the image. We propose the following interpretation to these elements: the model learns", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 281, 506, 293 ], "spans": [ { "bbox": [ 105, 281, 506, 293 ], "score": 1.0, "content": "to recognize patches containing little useful information, and recycle the corresponding tokens to", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 291, 401, 305 ], "spans": [ { "bbox": [ 105, 291, 401, 305 ], "score": 1.0, "content": "aggregate global image information while discarding spatial information.", "type": "text" } ], "index": 12 } ], "index": 9.5 }, { "type": "text", "bbox": [ 107, 308, 505, 385 ], "lines": [ { "bbox": [ 106, 308, 505, 321 ], "spans": [ { "bbox": [ 106, 308, 505, 321 ], "score": 1.0, "content": "This interpretation is consistent with an inner mechanism in transformer models that allows per-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 318, 506, 333 ], "spans": [ { "bbox": [ 105, 318, 506, 333 ], "score": 1.0, "content": "forming computations within a restricted set of tokens. In order to test this hypothesis, we append", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 330, 505, 343 ], "spans": [ { "bbox": [ 105, 330, 505, 343 ], "score": 1.0, "content": "additional tokens - that we call registers - to the token sequence, independent of the input image. We", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 340, 506, 355 ], "spans": [ { "bbox": [ 105, 340, 506, 355 ], "score": 1.0, "content": "train several models with and without this modification and observe that the outlier tokens disappear", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 352, 505, 365 ], "spans": [ { "bbox": [ 105, 352, 505, 365 ], "score": 1.0, "content": "from the sequence entirely. As a result, the performance of the models increases in dense prediction", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 363, 505, 376 ], "spans": [ { "bbox": [ 105, 363, 505, 376 ], "score": 1.0, "content": "tasks, and the resulting feature maps are significantly smoother. These smooth feature maps enable", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 375, 429, 387 ], "spans": [ { "bbox": [ 106, 375, 429, 387 ], "score": 1.0, "content": "object discovery methods like LOST mentioned above with the updated models.", "type": "text" } ], "index": 19 } ], "index": 16 }, { "type": "title", "bbox": [ 108, 403, 255, 415 ], "lines": [ { "bbox": [ 104, 401, 258, 417 ], "spans": [ { "bbox": [ 104, 401, 258, 417 ], "score": 1.0, "content": "2 PROBLEM FORMULATION", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 428, 505, 495 ], "lines": [ { "bbox": [ 106, 428, 505, 441 ], "spans": [ { "bbox": [ 106, 428, 505, 441 ], "score": 1.0, "content": "As shown in Fig. 2, most modern vision transformers exhibit artifacts in the attention maps. The", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 440, 506, 452 ], "spans": [ { "bbox": [ 105, 440, 506, 452 ], "score": 1.0, "content": "unsupervised DINO backbone (Caron et al., 2021) has been previously praised for the quality of", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 451, 506, 464 ], "spans": [ { "bbox": [ 106, 451, 506, 464 ], "score": 1.0, "content": "local features and interpretability of attention maps. Surprisingly, the outputs of the subsequent", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 462, 505, 474 ], "spans": [ { "bbox": [ 106, 462, 505, 474 ], "score": 1.0, "content": "DINOv2 models have been shown to hold good local information but exhibit undesirable artifacts in", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 472, 505, 485 ], "spans": [ { "bbox": [ 105, 472, 505, 485 ], "score": 1.0, "content": "attention maps. In this section, we propose to study why and when these artifacts appear. While this", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 484, 500, 496 ], "spans": [ { "bbox": [ 106, 484, 500, 496 ], "score": 1.0, "content": "work focuses on alleviating artefacts in all vision transformers, we focus our analysis on DINOv2.", "type": "text" } ], "index": 26 } ], "index": 23.5 }, { "type": "title", "bbox": [ 109, 510, 341, 521 ], "lines": [ { "bbox": [ 106, 510, 343, 523 ], "spans": [ { "bbox": [ 106, 510, 343, 523 ], "score": 1.0, "content": "2.1 ARTIFACTS IN THE LOCAL FEATURES OF DINOV2", "type": "text" } ], "index": 27 } ], "index": 27 }, { "type": "text", "bbox": [ 107, 531, 505, 641 ], "lines": [ { "bbox": [ 105, 530, 505, 544 ], "spans": [ { "bbox": [ 105, 530, 505, 544 ], "score": 1.0, "content": "Artifacts are high-norm outlier tokens. We want to find a quantitative way of characterizing", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 542, 505, 554 ], "spans": [ { "bbox": [ 105, 542, 505, 554 ], "score": 1.0, "content": "artefacts that appear in the local features. We observe that an important difference between “artifact”", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 553, 505, 565 ], "spans": [ { "bbox": [ 105, 553, 505, 565 ], "score": 1.0, "content": "patches and other patches is the norm of their token embedding at the output of the model. In Fig. 3", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 564, 506, 577 ], "spans": [ { "bbox": [ 105, 564, 506, 577 ], "score": 1.0, "content": "(left), we compare the norm of local features for a DINO and DINOv2 model given a reference", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 575, 506, 587 ], "spans": [ { "bbox": [ 105, 575, 506, 587 ], "score": 1.0, "content": "image. We clearly see that the norm of artifact patches is much higher than the norm of other", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 586, 506, 599 ], "spans": [ { "bbox": [ 105, 586, 506, 599 ], "score": 1.0, "content": "patches. We also plot the distribution of feature norms over a small dataset of images in Fig. 3", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 597, 506, 610 ], "spans": [ { "bbox": [ 105, 597, 506, 610 ], "score": 1.0, "content": "(right), which is clearly bimodal, allowing us to choose a simple criterion for the rest of this section:", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 608, 505, 620 ], "spans": [ { "bbox": [ 106, 608, 505, 620 ], "score": 1.0, "content": "tokens with norm higher than 150 will be considered as “high-norm” tokens, and we will study their", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 619, 506, 632 ], "spans": [ { "bbox": [ 105, 619, 506, 632 ], "score": 1.0, "content": "properties relative to regular tokens. This hand-picked cutoff value can vary across models. In the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 630, 382, 643 ], "spans": [ { "bbox": [ 105, 630, 382, 643 ], "score": 1.0, "content": "rest of this work, we use “high-norm” and “outlier” interchangeably.", "type": "text" } ], "index": 37 } ], "index": 32.5 }, { "type": "text", "bbox": [ 107, 655, 505, 732 ], "lines": [ { "bbox": [ 105, 654, 506, 668 ], "spans": [ { "bbox": [ 105, 654, 506, 668 ], "score": 1.0, "content": "Outliers appear during the training of large models. We make several additional observations", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 666, 505, 678 ], "spans": [ { "bbox": [ 105, 666, 505, 678 ], "score": 1.0, "content": "about the conditions in which these outlier patches appear during the training of DINOv2. This", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "analysis is illustrated in Fig. 4. First, these high-norm patches seem to differentiate themselves from", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "other patches around layer 15 of this 40-layer ViT (Fig. 4a). Second, when looking at the distribution", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "of norms along training of DINOv2, we see that these outliers only appear after one third of training", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 506, 723 ], "score": 1.0, "content": "(Fig. 4b). Finally, when analyzing more closely models of different size (Tiny, Small, Base, Large,", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 721, 441, 734 ], "spans": [ { "bbox": [ 106, 721, 441, 734 ], "score": 1.0, "content": "Huge and giant), we see that only the three largest models exhibit outliers (Fig. 4c).", "type": "text" } ], "index": 44 } ], "index": 41 } ], "page_idx": 2, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 750, 309, 762 ], "spans": [ { "bbox": [ 302, 750, 309, 762 ], "score": 1.0, "content": "3", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 108, 78, 509, 159 ], "blocks": [ { "type": "image_body", "bbox": [ 108, 78, 509, 159 ], "group_id": 0, "lines": [ { "bbox": [ 108, 78, 509, 159 ], "spans": [ { "bbox": [ 108, 78, 509, 159 ], "score": 0.963, "type": "image", "image_path": "fe26196bcac0e507f5fee2e2dc92dd8ad384ef11ab92073f969917b083f64c7e.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 108, 78, 509, 105.0 ], "spans": [], "index": 0 }, { "bbox": [ 108, 105.0, 509, 132.0 ], "spans": [], "index": 1 }, { "bbox": [ 108, 132.0, 509, 159.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 169, 505, 214 ], "group_id": 0, "lines": [ { "bbox": [ 105, 169, 505, 181 ], "spans": [ { "bbox": [ 105, 169, 448, 181 ], "score": 1.0, "content": "Figure 3: Comparison of local feature norms for DINO ViT-B/16 and DINOv2 ViT-", "type": "text" }, { "bbox": [ 448, 169, 468, 181 ], "score": 0.43, "content": "\\mathrm { \\ g } / 1 4", "type": "inline_equation" }, { "bbox": [ 468, 169, 505, 181 ], "score": 1.0, "content": ". We ob-", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 180, 505, 192 ], "spans": [ { "bbox": [ 105, 180, 505, 192 ], "score": 1.0, "content": "serve that DINOv2 has a few outlier patches, whereas DINO does not present these artifacts. For", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 192, 505, 203 ], "spans": [ { "bbox": [ 105, 192, 505, 203 ], "score": 1.0, "content": "DINOv2, although most patch tokens have a norm between 0 and 100, a small proportion of tokens", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 202, 498, 214 ], "spans": [ { "bbox": [ 105, 203, 467, 214 ], "score": 1.0, "content": "have a very high norm. We measure the proportion of tokens with norm larger than 150 at", "type": "text" }, { "bbox": [ 467, 202, 494, 213 ], "score": 0.88, "content": "2 . 3 7 \\%", "type": "inline_equation" }, { "bbox": [ 494, 203, 498, 214 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "text", "bbox": [ 107, 236, 505, 303 ], "lines": [], "index": 9.5, "bbox_fs": [ 105, 235, 506, 305 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 308, 505, 385 ], "lines": [ { "bbox": [ 106, 308, 505, 321 ], "spans": [ { "bbox": [ 106, 308, 505, 321 ], "score": 1.0, "content": "This interpretation is consistent with an inner mechanism in transformer models that allows per-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 318, 506, 333 ], "spans": [ { "bbox": [ 105, 318, 506, 333 ], "score": 1.0, "content": "forming computations within a restricted set of tokens. In order to test this hypothesis, we append", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 330, 505, 343 ], "spans": [ { "bbox": [ 105, 330, 505, 343 ], "score": 1.0, "content": "additional tokens - that we call registers - to the token sequence, independent of the input image. We", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 340, 506, 355 ], "spans": [ { "bbox": [ 105, 340, 506, 355 ], "score": 1.0, "content": "train several models with and without this modification and observe that the outlier tokens disappear", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 352, 505, 365 ], "spans": [ { "bbox": [ 105, 352, 505, 365 ], "score": 1.0, "content": "from the sequence entirely. As a result, the performance of the models increases in dense prediction", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 363, 505, 376 ], "spans": [ { "bbox": [ 105, 363, 505, 376 ], "score": 1.0, "content": "tasks, and the resulting feature maps are significantly smoother. These smooth feature maps enable", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 375, 429, 387 ], "spans": [ { "bbox": [ 106, 375, 429, 387 ], "score": 1.0, "content": "object discovery methods like LOST mentioned above with the updated models.", "type": "text" } ], "index": 19 } ], "index": 16, "bbox_fs": [ 105, 308, 506, 387 ] }, { "type": "title", "bbox": [ 108, 403, 255, 415 ], "lines": [ { "bbox": [ 104, 401, 258, 417 ], "spans": [ { "bbox": [ 104, 401, 258, 417 ], "score": 1.0, "content": "2 PROBLEM FORMULATION", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 428, 505, 495 ], "lines": [ { "bbox": [ 106, 428, 505, 441 ], "spans": [ { "bbox": [ 106, 428, 505, 441 ], "score": 1.0, "content": "As shown in Fig. 2, most modern vision transformers exhibit artifacts in the attention maps. The", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 440, 506, 452 ], "spans": [ { "bbox": [ 105, 440, 506, 452 ], "score": 1.0, "content": "unsupervised DINO backbone (Caron et al., 2021) has been previously praised for the quality of", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 451, 506, 464 ], "spans": [ { "bbox": [ 106, 451, 506, 464 ], "score": 1.0, "content": "local features and interpretability of attention maps. Surprisingly, the outputs of the subsequent", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 462, 505, 474 ], "spans": [ { "bbox": [ 106, 462, 505, 474 ], "score": 1.0, "content": "DINOv2 models have been shown to hold good local information but exhibit undesirable artifacts in", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 472, 505, 485 ], "spans": [ { "bbox": [ 105, 472, 505, 485 ], "score": 1.0, "content": "attention maps. In this section, we propose to study why and when these artifacts appear. While this", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 484, 500, 496 ], "spans": [ { "bbox": [ 106, 484, 500, 496 ], "score": 1.0, "content": "work focuses on alleviating artefacts in all vision transformers, we focus our analysis on DINOv2.", "type": "text" } ], "index": 26 } ], "index": 23.5, "bbox_fs": [ 105, 428, 506, 496 ] }, { "type": "title", "bbox": [ 109, 510, 341, 521 ], "lines": [ { "bbox": [ 106, 510, 343, 523 ], "spans": [ { "bbox": [ 106, 510, 343, 523 ], "score": 1.0, "content": "2.1 ARTIFACTS IN THE LOCAL FEATURES OF DINOV2", "type": "text" } ], "index": 27 } ], "index": 27 }, { "type": "text", "bbox": [ 107, 531, 505, 641 ], "lines": [ { "bbox": [ 105, 530, 505, 544 ], "spans": [ { "bbox": [ 105, 530, 505, 544 ], "score": 1.0, "content": "Artifacts are high-norm outlier tokens. We want to find a quantitative way of characterizing", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 542, 505, 554 ], "spans": [ { "bbox": [ 105, 542, 505, 554 ], "score": 1.0, "content": "artefacts that appear in the local features. We observe that an important difference between “artifact”", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 553, 505, 565 ], "spans": [ { "bbox": [ 105, 553, 505, 565 ], "score": 1.0, "content": "patches and other patches is the norm of their token embedding at the output of the model. In Fig. 3", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 564, 506, 577 ], "spans": [ { "bbox": [ 105, 564, 506, 577 ], "score": 1.0, "content": "(left), we compare the norm of local features for a DINO and DINOv2 model given a reference", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 575, 506, 587 ], "spans": [ { "bbox": [ 105, 575, 506, 587 ], "score": 1.0, "content": "image. We clearly see that the norm of artifact patches is much higher than the norm of other", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 586, 506, 599 ], "spans": [ { "bbox": [ 105, 586, 506, 599 ], "score": 1.0, "content": "patches. We also plot the distribution of feature norms over a small dataset of images in Fig. 3", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 597, 506, 610 ], "spans": [ { "bbox": [ 105, 597, 506, 610 ], "score": 1.0, "content": "(right), which is clearly bimodal, allowing us to choose a simple criterion for the rest of this section:", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 608, 505, 620 ], "spans": [ { "bbox": [ 106, 608, 505, 620 ], "score": 1.0, "content": "tokens with norm higher than 150 will be considered as “high-norm” tokens, and we will study their", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 619, 506, 632 ], "spans": [ { "bbox": [ 105, 619, 506, 632 ], "score": 1.0, "content": "properties relative to regular tokens. This hand-picked cutoff value can vary across models. In the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 630, 382, 643 ], "spans": [ { "bbox": [ 105, 630, 382, 643 ], "score": 1.0, "content": "rest of this work, we use “high-norm” and “outlier” interchangeably.", "type": "text" } ], "index": 37 } ], "index": 32.5, "bbox_fs": [ 105, 530, 506, 643 ] }, { "type": "text", "bbox": [ 107, 655, 505, 732 ], "lines": [ { "bbox": [ 105, 654, 506, 668 ], "spans": [ { "bbox": [ 105, 654, 506, 668 ], "score": 1.0, "content": "Outliers appear during the training of large models. We make several additional observations", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 666, 505, 678 ], "spans": [ { "bbox": [ 105, 666, 505, 678 ], "score": 1.0, "content": "about the conditions in which these outlier patches appear during the training of DINOv2. This", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "analysis is illustrated in Fig. 4. First, these high-norm patches seem to differentiate themselves from", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "other patches around layer 15 of this 40-layer ViT (Fig. 4a). Second, when looking at the distribution", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "of norms along training of DINOv2, we see that these outliers only appear after one third of training", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 506, 723 ], "score": 1.0, "content": "(Fig. 4b). Finally, when analyzing more closely models of different size (Tiny, Small, Base, Large,", "type": "text" } ], "index": 43 }, { "bbox": [ 106, 721, 441, 734 ], "spans": [ { "bbox": [ 106, 721, 441, 734 ], "score": 1.0, "content": "Huge and giant), we see that only the three largest models exhibit outliers (Fig. 4c).", "type": "text" } ], "index": 44 } ], "index": 41, "bbox_fs": [ 105, 654, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 118, 83, 496, 182 ], "blocks": [ { "type": "image_body", "bbox": [ 118, 83, 496, 182 ], "group_id": 0, "lines": [ { "bbox": [ 118, 83, 496, 182 ], "spans": [ { "bbox": [ 118, 83, 496, 182 ], "score": 0.965, "type": "image", "image_path": "6e4b51f8abf6433478ef962801a5e327f9fbd0a4441ec75c954b1df856560849.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 118, 83, 496, 116.0 ], "spans": [], "index": 0 }, { "bbox": [ 118, 116.0, 496, 149.0 ], "spans": [], "index": 1 }, { "bbox": [ 118, 149.0, 496, 182.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 190, 505, 235 ], "group_id": 0, "lines": [ { "bbox": [ 106, 190, 505, 202 ], "spans": [ { "bbox": [ 106, 190, 467, 202 ], "score": 1.0, "content": "Figure 4: Illustration of several properties of outlier tokens in the 40-layer DINOv2 ViT-", "type": "text" }, { "bbox": [ 468, 192, 474, 202 ], "score": 0.32, "content": "\\mathbf { g }", "type": "inline_equation" }, { "bbox": [ 474, 190, 505, 202 ], "score": 1.0, "content": "model.", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 200, 506, 214 ], "spans": [ { "bbox": [ 105, 200, 506, 214 ], "score": 1.0, "content": "(a): Distribution of output token norms along layers. (b): Distribution of norms along training", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 212, 506, 226 ], "spans": [ { "bbox": [ 105, 212, 506, 226 ], "score": 1.0, "content": "iterations. (c): Distribution of norms for different model sizes. The outliers appear around the", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 222, 503, 237 ], "spans": [ { "bbox": [ 105, 222, 503, 237 ], "score": 1.0, "content": "middle of the model during training; they appear with models larger than and including ViT-Large.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "table", "bbox": [ 259, 262, 504, 325 ], "blocks": [ { "type": "table_body", "bbox": [ 259, 262, 504, 325 ], "group_id": 0, "lines": [ { "bbox": [ 259, 262, 504, 325 ], "spans": [ { "bbox": [ 259, 262, 504, 325 ], "score": 0.962, "html": "
position predictionreconstruction
top-1 accavg. distance ↓L2 error↓
normal41.70.7918.38
outlier22.85.0925.23
", "type": "table", "image_path": "eeeea2700af305c2121364a07acef2df8ac83746cbe5ba0a657c2a974ccf61d6.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 259, 262, 504, 283.0 ], "spans": [], "index": 8 }, { "bbox": [ 259, 283.0, 504, 304.0 ], "spans": [], "index": 10 }, { "bbox": [ 259, 304.0, 504, 325.0 ], "spans": [], "index": 11 } ] }, { "type": "table_footnote", "bbox": [ 306, 329, 453, 339 ], "group_id": 0, "lines": [ { "bbox": [ 306, 328, 454, 340 ], "spans": [ { "bbox": [ 306, 328, 454, 340 ], "score": 1.0, "content": "(b) Linear probing for local information.", "type": "text" } ], "index": 13 } ], "index": 13 } ], "index": 11.5 }, { "type": "text", "bbox": [ 108, 329, 231, 339 ], "lines": [ { "bbox": [ 106, 327, 232, 342 ], "spans": [ { "bbox": [ 106, 327, 232, 342 ], "score": 1.0, "content": "(a) Cosine similarity to neighbors.", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "image", "bbox": [ 110, 246, 230, 323 ], "blocks": [ { "type": "image_body", "bbox": [ 110, 246, 230, 323 ], "group_id": 1, "lines": [ { "bbox": [ 110, 246, 230, 323 ], "spans": [ { "bbox": [ 110, 246, 230, 323 ], "score": 0.93, "type": "image", "image_path": "a21cad8254da9e01cfee6f7f5dfc9e9a241e8f73e93b0137a9e097e8e5654def.jpg" } ] } ], "index": 8.0, "virtual_lines": [ { "bbox": [ 110, 246, 230, 284.5 ], "spans": [], "index": 7 }, { "bbox": [ 110, 284.5, 230, 323.0 ], "spans": [], "index": 9 } ] }, { "type": "image_caption", "bbox": [ 106, 348, 505, 404 ], "group_id": 1, "lines": [ { "bbox": [ 105, 348, 505, 361 ], "spans": [ { "bbox": [ 105, 348, 505, 361 ], "score": 1.0, "content": "Figure 5: (a): Distribution of cosine similarity between input patches and their 4 neighbors. We", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 360, 504, 371 ], "spans": [ { "bbox": [ 106, 360, 504, 371 ], "score": 1.0, "content": "plot separately artifact patches (norm of the output token over 150) and normal patches. (b): Local", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 369, 506, 385 ], "spans": [ { "bbox": [ 105, 369, 506, 385 ], "score": 1.0, "content": "information probing on normal and outlier patch tokens. We train two models: one for predicting", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 381, 505, 394 ], "spans": [ { "bbox": [ 105, 381, 505, 394 ], "score": 1.0, "content": "position, and one for reconstructing the input patch. Outlier tokens have much lower scores than the", "type": "text" } ], "index": 17 }, { "bbox": [ 107, 393, 384, 405 ], "spans": [ { "bbox": [ 107, 393, 384, 405 ], "score": 1.0, "content": "other tokens, suggesting they are storing less local patch information.", "type": "text" } ], "index": 18 } ], "index": 16 } ], "index": 12.0 }, { "type": "text", "bbox": [ 107, 424, 504, 502 ], "lines": [ { "bbox": [ 105, 425, 506, 438 ], "spans": [ { "bbox": [ 105, 425, 506, 438 ], "score": 1.0, "content": "High-norm tokens appear where patch information is redundant. To verify this, we measure", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 436, 505, 448 ], "spans": [ { "bbox": [ 106, 436, 505, 448 ], "score": 1.0, "content": "the cosine similarity between high-norm tokens and their 4 neighbors right after the patch em-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 446, 506, 460 ], "spans": [ { "bbox": [ 105, 446, 506, 460 ], "score": 1.0, "content": "bedding layer (at the beginning of the vision transformer). We illustrate the density plot in Fig.", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 457, 506, 470 ], "spans": [ { "bbox": [ 105, 457, 506, 470 ], "score": 1.0, "content": "5a. We observe that high-norm tokens appear on patches that are very similar to their neighbors.", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 469, 506, 481 ], "spans": [ { "bbox": [ 106, 469, 506, 481 ], "score": 1.0, "content": "This suggests that these patches contrain redundant information and that the model could discard", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 479, 506, 492 ], "spans": [ { "bbox": [ 105, 479, 506, 492 ], "score": 1.0, "content": "their information without hurting the quality of the image representation. This matches qualitative", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 491, 418, 504 ], "spans": [ { "bbox": [ 106, 491, 418, 504 ], "score": 1.0, "content": "observations (see Fig. 2) that they often appear in uniform, background areas.", "type": "text" } ], "index": 25 } ], "index": 22 }, { "type": "text", "bbox": [ 108, 514, 505, 580 ], "lines": [ { "bbox": [ 106, 514, 505, 525 ], "spans": [ { "bbox": [ 106, 514, 505, 525 ], "score": 1.0, "content": "High-norm tokens hold little local information. In order to better understand the nature of these", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 525, 506, 537 ], "spans": [ { "bbox": [ 106, 525, 506, 537 ], "score": 1.0, "content": "tokens, we propose to probe the patch embeddings for different types of information. For that we", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 535, 505, 549 ], "spans": [ { "bbox": [ 106, 535, 505, 549 ], "score": 1.0, "content": "consider two different tasks: position prediction and pixel reconstruction. For each of these tasks,", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 547, 505, 559 ], "spans": [ { "bbox": [ 106, 547, 505, 559 ], "score": 1.0, "content": "we train a linear model on top of the patch embeddings, and measure the performance of this", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 558, 506, 571 ], "spans": [ { "bbox": [ 105, 558, 506, 571 ], "score": 1.0, "content": "model. We compare the performance achieved with high-norm tokens and with other tokens, to see", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 569, 399, 581 ], "spans": [ { "bbox": [ 106, 569, 399, 581 ], "score": 1.0, "content": "if high-norm tokens contain different information than “normal” tokens.", "type": "text" } ], "index": 31 } ], "index": 28.5 }, { "type": "text", "bbox": [ 133, 590, 505, 693 ], "lines": [ { "bbox": [ 133, 590, 505, 603 ], "spans": [ { "bbox": [ 133, 590, 505, 603 ], "score": 1.0, "content": "• Position prediction. We train a linear model to predict the position of each patch token in", "type": "text" } ], "index": 32 }, { "bbox": [ 141, 601, 506, 614 ], "spans": [ { "bbox": [ 141, 601, 506, 614 ], "score": 1.0, "content": "the image, and measure its accuracy. We note that this position information was injected", "type": "text" } ], "index": 33 }, { "bbox": [ 141, 612, 506, 624 ], "spans": [ { "bbox": [ 141, 612, 506, 624 ], "score": 1.0, "content": "in the tokens before the first ViT layer in the form of absolute position embeddings. We", "type": "text" } ], "index": 34 }, { "bbox": [ 141, 623, 505, 636 ], "spans": [ { "bbox": [ 141, 623, 505, 636 ], "score": 1.0, "content": "observe that high-norm tokens have much lower accuracy than the other tokens (Fig. 5b),", "type": "text" } ], "index": 35 }, { "bbox": [ 141, 633, 441, 648 ], "spans": [ { "bbox": [ 141, 633, 441, 648 ], "score": 1.0, "content": "suggesting they contain less information about their position in the image.", "type": "text" } ], "index": 36 }, { "bbox": [ 133, 648, 505, 661 ], "spans": [ { "bbox": [ 133, 648, 505, 661 ], "score": 1.0, "content": "• Pixel reconstruction. We train a linear model to predict the pixel values of the image from", "type": "text" } ], "index": 37 }, { "bbox": [ 141, 660, 506, 672 ], "spans": [ { "bbox": [ 141, 660, 506, 672 ], "score": 1.0, "content": "the patch embeddings, and measure the accuracy of this model. We observe again that", "type": "text" } ], "index": 38 }, { "bbox": [ 141, 670, 505, 684 ], "spans": [ { "bbox": [ 141, 670, 505, 684 ], "score": 1.0, "content": "high-norm tokens achieve much lower accuracy than other tokens (Fig. 5b). This suggests", "type": "text" } ], "index": 39 }, { "bbox": [ 141, 682, 497, 694 ], "spans": [ { "bbox": [ 141, 682, 497, 694 ], "score": 1.0, "content": "that high-norm tokens contain less information to reconstruct the image than the others.", "type": "text" } ], "index": 40 } ], "index": 36 }, { "type": "text", "bbox": [ 105, 709, 505, 732 ], "lines": [ { "bbox": [ 105, 708, 505, 722 ], "spans": [ { "bbox": [ 105, 708, 505, 722 ], "score": 1.0, "content": "Artifacts hold global information. In order to evaluate how much global information is gathered", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 720, 505, 734 ], "spans": [ { "bbox": [ 105, 720, 505, 734 ], "score": 1.0, "content": "in the high-norm tokens, we propose to evaluate them on standard image representation learning", "type": "text" } ], "index": 42 } ], "index": 41.5 } ], "page_idx": 3, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 310, 762 ], "spans": [ { "bbox": [ 302, 751, 310, 762 ], "score": 1.0, "content": "4", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 118, 83, 496, 182 ], "blocks": [ { "type": "image_body", "bbox": [ 118, 83, 496, 182 ], "group_id": 0, "lines": [ { "bbox": [ 118, 83, 496, 182 ], "spans": [ { "bbox": [ 118, 83, 496, 182 ], "score": 0.965, "type": "image", "image_path": "6e4b51f8abf6433478ef962801a5e327f9fbd0a4441ec75c954b1df856560849.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 118, 83, 496, 116.0 ], "spans": [], "index": 0 }, { "bbox": [ 118, 116.0, 496, 149.0 ], "spans": [], "index": 1 }, { "bbox": [ 118, 149.0, 496, 182.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 190, 505, 235 ], "group_id": 0, "lines": [ { "bbox": [ 106, 190, 505, 202 ], "spans": [ { "bbox": [ 106, 190, 467, 202 ], "score": 1.0, "content": "Figure 4: Illustration of several properties of outlier tokens in the 40-layer DINOv2 ViT-", "type": "text" }, { "bbox": [ 468, 192, 474, 202 ], "score": 0.32, "content": "\\mathbf { g }", "type": "inline_equation" }, { "bbox": [ 474, 190, 505, 202 ], "score": 1.0, "content": "model.", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 200, 506, 214 ], "spans": [ { "bbox": [ 105, 200, 506, 214 ], "score": 1.0, "content": "(a): Distribution of output token norms along layers. (b): Distribution of norms along training", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 212, 506, 226 ], "spans": [ { "bbox": [ 105, 212, 506, 226 ], "score": 1.0, "content": "iterations. (c): Distribution of norms for different model sizes. The outliers appear around the", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 222, 503, 237 ], "spans": [ { "bbox": [ 105, 222, 503, 237 ], "score": 1.0, "content": "middle of the model during training; they appear with models larger than and including ViT-Large.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "table", "bbox": [ 259, 262, 504, 325 ], "blocks": [ { "type": "table_body", "bbox": [ 259, 262, 504, 325 ], "group_id": 0, "lines": [ { "bbox": [ 259, 262, 504, 325 ], "spans": [ { "bbox": [ 259, 262, 504, 325 ], "score": 0.962, "html": "
position predictionreconstruction
top-1 accavg. distance ↓L2 error↓
normal41.70.7918.38
outlier22.85.0925.23
", "type": "table", "image_path": "eeeea2700af305c2121364a07acef2df8ac83746cbe5ba0a657c2a974ccf61d6.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 259, 262, 504, 283.0 ], "spans": [], "index": 8 }, { "bbox": [ 259, 283.0, 504, 304.0 ], "spans": [], "index": 10 }, { "bbox": [ 259, 304.0, 504, 325.0 ], "spans": [], "index": 11 } ] }, { "type": "table_footnote", "bbox": [ 306, 329, 453, 339 ], "group_id": 0, "lines": [ { "bbox": [ 306, 328, 454, 340 ], "spans": [ { "bbox": [ 306, 328, 454, 340 ], "score": 1.0, "content": "(b) Linear probing for local information.", "type": "text" } ], "index": 13 } ], "index": 13 } ], "index": 11.5 }, { "type": "text", "bbox": [ 108, 329, 231, 339 ], "lines": [ { "bbox": [ 106, 327, 232, 342 ], "spans": [ { "bbox": [ 106, 327, 232, 342 ], "score": 1.0, "content": "(a) Cosine similarity to neighbors.", "type": "text" } ], "index": 12 } ], "index": 12, "bbox_fs": [ 106, 327, 232, 342 ] }, { "type": "image", "bbox": [ 110, 246, 230, 323 ], "blocks": [ { "type": "image_body", "bbox": [ 110, 246, 230, 323 ], "group_id": 1, "lines": [ { "bbox": [ 110, 246, 230, 323 ], "spans": [ { "bbox": [ 110, 246, 230, 323 ], "score": 0.93, "type": "image", "image_path": "a21cad8254da9e01cfee6f7f5dfc9e9a241e8f73e93b0137a9e097e8e5654def.jpg" } ] } ], "index": 8.0, "virtual_lines": [ { "bbox": [ 110, 246, 230, 284.5 ], "spans": [], "index": 7 }, { "bbox": [ 110, 284.5, 230, 323.0 ], "spans": [], "index": 9 } ] }, { "type": "image_caption", "bbox": [ 106, 348, 505, 404 ], "group_id": 1, "lines": [ { "bbox": [ 105, 348, 505, 361 ], "spans": [ { "bbox": [ 105, 348, 505, 361 ], "score": 1.0, "content": "Figure 5: (a): Distribution of cosine similarity between input patches and their 4 neighbors. We", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 360, 504, 371 ], "spans": [ { "bbox": [ 106, 360, 504, 371 ], "score": 1.0, "content": "plot separately artifact patches (norm of the output token over 150) and normal patches. (b): Local", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 369, 506, 385 ], "spans": [ { "bbox": [ 105, 369, 506, 385 ], "score": 1.0, "content": "information probing on normal and outlier patch tokens. We train two models: one for predicting", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 381, 505, 394 ], "spans": [ { "bbox": [ 105, 381, 505, 394 ], "score": 1.0, "content": "position, and one for reconstructing the input patch. Outlier tokens have much lower scores than the", "type": "text" } ], "index": 17 }, { "bbox": [ 107, 393, 384, 405 ], "spans": [ { "bbox": [ 107, 393, 384, 405 ], "score": 1.0, "content": "other tokens, suggesting they are storing less local patch information.", "type": "text" } ], "index": 18 } ], "index": 16 } ], "index": 12.0 }, { "type": "text", "bbox": [ 107, 424, 504, 502 ], "lines": [ { "bbox": [ 105, 425, 506, 438 ], "spans": [ { "bbox": [ 105, 425, 506, 438 ], "score": 1.0, "content": "High-norm tokens appear where patch information is redundant. To verify this, we measure", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 436, 505, 448 ], "spans": [ { "bbox": [ 106, 436, 505, 448 ], "score": 1.0, "content": "the cosine similarity between high-norm tokens and their 4 neighbors right after the patch em-", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 446, 506, 460 ], "spans": [ { "bbox": [ 105, 446, 506, 460 ], "score": 1.0, "content": "bedding layer (at the beginning of the vision transformer). We illustrate the density plot in Fig.", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 457, 506, 470 ], "spans": [ { "bbox": [ 105, 457, 506, 470 ], "score": 1.0, "content": "5a. We observe that high-norm tokens appear on patches that are very similar to their neighbors.", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 469, 506, 481 ], "spans": [ { "bbox": [ 106, 469, 506, 481 ], "score": 1.0, "content": "This suggests that these patches contrain redundant information and that the model could discard", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 479, 506, 492 ], "spans": [ { "bbox": [ 105, 479, 506, 492 ], "score": 1.0, "content": "their information without hurting the quality of the image representation. This matches qualitative", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 491, 418, 504 ], "spans": [ { "bbox": [ 106, 491, 418, 504 ], "score": 1.0, "content": "observations (see Fig. 2) that they often appear in uniform, background areas.", "type": "text" } ], "index": 25 } ], "index": 22, "bbox_fs": [ 105, 425, 506, 504 ] }, { "type": "text", "bbox": [ 108, 514, 505, 580 ], "lines": [ { "bbox": [ 106, 514, 505, 525 ], "spans": [ { "bbox": [ 106, 514, 505, 525 ], "score": 1.0, "content": "High-norm tokens hold little local information. In order to better understand the nature of these", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 525, 506, 537 ], "spans": [ { "bbox": [ 106, 525, 506, 537 ], "score": 1.0, "content": "tokens, we propose to probe the patch embeddings for different types of information. For that we", "type": "text" } ], "index": 27 }, { "bbox": [ 106, 535, 505, 549 ], "spans": [ { "bbox": [ 106, 535, 505, 549 ], "score": 1.0, "content": "consider two different tasks: position prediction and pixel reconstruction. For each of these tasks,", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 547, 505, 559 ], "spans": [ { "bbox": [ 106, 547, 505, 559 ], "score": 1.0, "content": "we train a linear model on top of the patch embeddings, and measure the performance of this", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 558, 506, 571 ], "spans": [ { "bbox": [ 105, 558, 506, 571 ], "score": 1.0, "content": "model. We compare the performance achieved with high-norm tokens and with other tokens, to see", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 569, 399, 581 ], "spans": [ { "bbox": [ 106, 569, 399, 581 ], "score": 1.0, "content": "if high-norm tokens contain different information than “normal” tokens.", "type": "text" } ], "index": 31 } ], "index": 28.5, "bbox_fs": [ 105, 514, 506, 581 ] }, { "type": "text", "bbox": [ 133, 590, 505, 693 ], "lines": [ { "bbox": [ 133, 590, 505, 603 ], "spans": [ { "bbox": [ 133, 590, 505, 603 ], "score": 1.0, "content": "• Position prediction. We train a linear model to predict the position of each patch token in", "type": "text" } ], "index": 32 }, { "bbox": [ 141, 601, 506, 614 ], "spans": [ { "bbox": [ 141, 601, 506, 614 ], "score": 1.0, "content": "the image, and measure its accuracy. We note that this position information was injected", "type": "text" } ], "index": 33 }, { "bbox": [ 141, 612, 506, 624 ], "spans": [ { "bbox": [ 141, 612, 506, 624 ], "score": 1.0, "content": "in the tokens before the first ViT layer in the form of absolute position embeddings. We", "type": "text" } ], "index": 34 }, { "bbox": [ 141, 623, 505, 636 ], "spans": [ { "bbox": [ 141, 623, 505, 636 ], "score": 1.0, "content": "observe that high-norm tokens have much lower accuracy than the other tokens (Fig. 5b),", "type": "text" } ], "index": 35 }, { "bbox": [ 141, 633, 441, 648 ], "spans": [ { "bbox": [ 141, 633, 441, 648 ], "score": 1.0, "content": "suggesting they contain less information about their position in the image.", "type": "text" } ], "index": 36 }, { "bbox": [ 133, 648, 505, 661 ], "spans": [ { "bbox": [ 133, 648, 505, 661 ], "score": 1.0, "content": "• Pixel reconstruction. We train a linear model to predict the pixel values of the image from", "type": "text" } ], "index": 37 }, { "bbox": [ 141, 660, 506, 672 ], "spans": [ { "bbox": [ 141, 660, 506, 672 ], "score": 1.0, "content": "the patch embeddings, and measure the accuracy of this model. We observe again that", "type": "text" } ], "index": 38 }, { "bbox": [ 141, 670, 505, 684 ], "spans": [ { "bbox": [ 141, 670, 505, 684 ], "score": 1.0, "content": "high-norm tokens achieve much lower accuracy than other tokens (Fig. 5b). This suggests", "type": "text" } ], "index": 39 }, { "bbox": [ 141, 682, 497, 694 ], "spans": [ { "bbox": [ 141, 682, 497, 694 ], "score": 1.0, "content": "that high-norm tokens contain less information to reconstruct the image than the others.", "type": "text" } ], "index": 40 } ], "index": 36, "bbox_fs": [ 133, 590, 506, 694 ] }, { "type": "text", "bbox": [ 105, 709, 505, 732 ], "lines": [ { "bbox": [ 105, 708, 505, 722 ], "spans": [ { "bbox": [ 105, 708, 505, 722 ], "score": 1.0, "content": "Artifacts hold global information. In order to evaluate how much global information is gathered", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 720, 505, 734 ], "spans": [ { "bbox": [ 105, 720, 505, 734 ], "score": 1.0, "content": "in the high-norm tokens, we propose to evaluate them on standard image representation learning", "type": "text" } ], "index": 42 } ], "index": 41.5, "bbox_fs": [ 105, 708, 505, 734 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 107, 80, 504, 137 ], "blocks": [ { "type": "table_body", "bbox": [ 107, 80, 504, 137 ], "group_id": 0, "lines": [ { "bbox": [ 107, 80, 504, 137 ], "spans": [ { "bbox": [ 107, 80, 504, 137 ], "score": 0.971, "html": "
IN1k P205 Airc. CF10 CF100 CUB Cal101 Cars DTD Flow.Food Pets SUN VOC
[CLS]86.0 66.4 87.399.494.591.396.991.5 85.2 99.7 94.7 96.9 78.6 89.1
normal65.8 53.1 17.197.181.318.673.210.8 63.159.574.2 47.8 37.7 70.8
outlier69.0 55.179.199.393.784.997.685.2 84.999.6 93.5 94.1 78.589.7
", "type": "table", "image_path": "49e2e86cb5b67ef324d3ec8af60b38a0a51dd993441c9ea7d2a446f6ec43ef7b.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 80, 504, 99.0 ], "spans": [], "index": 0 }, { "bbox": [ 107, 99.0, 504, 118.0 ], "spans": [], "index": 1 }, { "bbox": [ 107, 118.0, 504, 137.0 ], "spans": [], "index": 2 } ] } ], "index": 1 }, { "type": "text", "bbox": [ 107, 145, 505, 178 ], "lines": [ { "bbox": [ 106, 145, 506, 158 ], "spans": [ { "bbox": [ 106, 145, 506, 158 ], "score": 1.0, "content": "Table 1: Image classification via linear probing on normal and outlier patch tokens. We also report", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 156, 504, 167 ], "spans": [ { "bbox": [ 106, 156, 504, 167 ], "score": 1.0, "content": "the accuracy of classifiers learnt on the class token. We see that outlier tokens have a much higher", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 167, 477, 180 ], "spans": [ { "bbox": [ 106, 167, 477, 180 ], "score": 1.0, "content": "accuracy than regular ones, suggesting they are effectively storing global image information.", "type": "text" } ], "index": 5 } ], "index": 4 }, { "type": "image", "bbox": [ 139, 183, 472, 310 ], "blocks": [ { "type": "image_body", "bbox": [ 139, 183, 472, 310 ], "group_id": 0, "lines": [ { "bbox": [ 139, 183, 472, 310 ], "spans": [ { "bbox": [ 139, 183, 472, 310 ], "score": 0.972, "type": "image", "image_path": "1e848b87cde847a24e97afbf9af6f8da69a1f37248c2a7b358255f6d2278602c.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 139, 183, 472, 225.33333333333334 ], "spans": [], "index": 6 }, { "bbox": [ 139, 225.33333333333334, 472, 267.6666666666667 ], "spans": [], "index": 7 }, { "bbox": [ 139, 267.6666666666667, 472, 310.0 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 106, 319, 505, 353 ], "group_id": 0, "lines": [ { "bbox": [ 106, 319, 504, 331 ], "spans": [ { "bbox": [ 106, 319, 425, 331 ], "score": 1.0, "content": "Figure 6: Illustration of the proposed remediation and resulting model. We add", "type": "text" }, { "bbox": [ 425, 320, 436, 330 ], "score": 0.74, "content": "N", "type": "inline_equation" }, { "bbox": [ 436, 319, 504, 331 ], "score": 1.0, "content": "additional learn-", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 330, 505, 342 ], "spans": [ { "bbox": [ 105, 330, 505, 342 ], "score": 1.0, "content": "able input tokens (depicted in yellow), that the model can use as registers. At the output of the", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 342, 481, 354 ], "spans": [ { "bbox": [ 106, 342, 481, 354 ], "score": 1.0, "content": "model, only the patch tokens and [CLS] tokens are used, both during training and inference.", "type": "text" } ], "index": 11 } ], "index": 10 } ], "index": 8.5 }, { "type": "text", "bbox": [ 108, 365, 505, 432 ], "lines": [ { "bbox": [ 105, 364, 505, 378 ], "spans": [ { "bbox": [ 105, 364, 464, 378 ], "score": 1.0, "content": "benchmarks. For each image in a classification dataset, we forward it through DINOv2-", "type": "text" }, { "bbox": [ 465, 367, 471, 377 ], "score": 0.51, "content": "\\mathbf { g }", "type": "inline_equation" }, { "bbox": [ 471, 364, 505, 378 ], "score": 1.0, "content": "and ex-", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 376, 506, 389 ], "spans": [ { "bbox": [ 105, 376, 506, 389 ], "score": 1.0, "content": "tract the patch embeddings. From those, we choose a single token at random, either high-norm or", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 387, 505, 401 ], "spans": [ { "bbox": [ 105, 387, 505, 401 ], "score": 1.0, "content": "normal. This token is then considered as the image representation. We then train a logistic regres-", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 399, 505, 411 ], "spans": [ { "bbox": [ 106, 399, 505, 411 ], "score": 1.0, "content": "sion classifier to predict the image class from this representation, and measure the accuracy. We", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 409, 505, 422 ], "spans": [ { "bbox": [ 106, 409, 505, 422 ], "score": 1.0, "content": "observe that the high-norm tokens have a much higher accuracy than the other tokens (Table 1). This", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 420, 452, 433 ], "spans": [ { "bbox": [ 105, 420, 452, 433 ], "score": 1.0, "content": "suggests that outlier tokens contain more global information than other patch tokens.", "type": "text" } ], "index": 17 } ], "index": 14.5 }, { "type": "title", "bbox": [ 109, 447, 270, 457 ], "lines": [ { "bbox": [ 106, 447, 272, 459 ], "spans": [ { "bbox": [ 106, 447, 272, 459 ], "score": 1.0, "content": "2.2 HYPOTHESIS AND REMEDIATION", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 108, 467, 504, 523 ], "lines": [ { "bbox": [ 106, 468, 505, 480 ], "spans": [ { "bbox": [ 106, 468, 505, 480 ], "score": 1.0, "content": "Having made these observations, we make the following hypothesis: large, sufficiently trained mod-", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 479, 505, 491 ], "spans": [ { "bbox": [ 106, 479, 505, 491 ], "score": 1.0, "content": "els learn to recognize redundant tokens, and to use them as places to store, process and retrieve", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 489, 506, 502 ], "spans": [ { "bbox": [ 105, 489, 506, 502 ], "score": 1.0, "content": "global information. Furthermore, we posit that while this behavior is not bad in itself, the fact that", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 500, 505, 512 ], "spans": [ { "bbox": [ 105, 500, 505, 512 ], "score": 1.0, "content": "it happens inside the patch tokens is undesirable. Indeed, it leads the model to discard local patch", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 512, 471, 524 ], "spans": [ { "bbox": [ 106, 512, 471, 524 ], "score": 1.0, "content": "information (Tab. 5b), possibly incurring decreased performance on dense prediction tasks.", "type": "text" } ], "index": 23 } ], "index": 21 }, { "type": "text", "bbox": [ 108, 528, 504, 606 ], "lines": [ { "bbox": [ 106, 529, 506, 541 ], "spans": [ { "bbox": [ 106, 529, 506, 541 ], "score": 1.0, "content": "We therefore propose a simple fix to this issue: we explicitly add new tokens to the sequence, that", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 540, 505, 551 ], "spans": [ { "bbox": [ 106, 540, 505, 551 ], "score": 1.0, "content": "the model can learn to use as registers. We add these tokens after the patch embedding layer, with a", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 550, 506, 563 ], "spans": [ { "bbox": [ 105, 550, 506, 563 ], "score": 1.0, "content": "learnable value, similarly to the [CLS] token. At the end of the vision transformer, these tokens are", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 561, 505, 574 ], "spans": [ { "bbox": [ 106, 561, 505, 574 ], "score": 1.0, "content": "discarded, and the [CLS] token and patch tokens are used as image representations, as usual. This", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 572, 505, 585 ], "spans": [ { "bbox": [ 105, 572, 505, 585 ], "score": 1.0, "content": "mechanism was first proposed in Memory Transformers (Burtsev et al., 2020), improving translation", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 583, 506, 596 ], "spans": [ { "bbox": [ 106, 583, 506, 596 ], "score": 1.0, "content": "tasks in NLP. Interestingly, we show here that this mechanism admits a natural justification for vision", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 595, 464, 607 ], "spans": [ { "bbox": [ 106, 595, 464, 607 ], "score": 1.0, "content": "transformers, fixing an interpretability and performance issue that was present otherwise.", "type": "text" } ], "index": 30 } ], "index": 27 }, { "type": "text", "bbox": [ 108, 611, 504, 656 ], "lines": [ { "bbox": [ 105, 610, 506, 624 ], "spans": [ { "bbox": [ 105, 610, 506, 624 ], "score": 1.0, "content": "We note that we have not been able to fully determine which aspects of the training led to the appear-", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 622, 506, 634 ], "spans": [ { "bbox": [ 105, 622, 506, 634 ], "score": 1.0, "content": "ance of artifacts in different models. The pretraining paradigm seems to play a role, as OpenCLIP", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 632, 506, 646 ], "spans": [ { "bbox": [ 105, 632, 506, 646 ], "score": 1.0, "content": "and DeiT-III exhibit outliers both at size B and L (Fig. 2). However, the model size and training", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 645, 327, 657 ], "spans": [ { "bbox": [ 106, 645, 327, 657 ], "score": 1.0, "content": "length also play important parts, as observed in Fig. 4.", "type": "text" } ], "index": 34 } ], "index": 32.5 }, { "type": "title", "bbox": [ 108, 673, 200, 685 ], "lines": [ { "bbox": [ 104, 672, 202, 688 ], "spans": [ { "bbox": [ 104, 672, 202, 688 ], "score": 1.0, "content": "3 EXPERIMENTS", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 106, 699, 504, 732 ], "lines": [ { "bbox": [ 105, 698, 505, 712 ], "spans": [ { "bbox": [ 105, 698, 505, 712 ], "score": 1.0, "content": "In this section, we validate the proposed solution by training vision transformers with additional", "type": "text" } ], "index": 36 }, { "bbox": [ 107, 710, 505, 722 ], "spans": [ { "bbox": [ 107, 710, 505, 722 ], "score": 1.0, "content": "[reg] register tokens. We evaluate the effectiveness of our approach by a quantitative and quali-", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 721, 505, 733 ], "spans": [ { "bbox": [ 106, 721, 505, 733 ], "score": 1.0, "content": "tative analysis. We then ablate the number of registers used for training, to check that they do not", "type": "text" } ], "index": 38 } ], "index": 37 } ], "page_idx": 4, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 750, 309, 763 ], "spans": [ { "bbox": [ 302, 750, 309, 763 ], "score": 1.0, "content": "5", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 107, 80, 504, 137 ], "blocks": [ { "type": "table_body", "bbox": [ 107, 80, 504, 137 ], "group_id": 0, "lines": [ { "bbox": [ 107, 80, 504, 137 ], "spans": [ { "bbox": [ 107, 80, 504, 137 ], "score": 0.971, "html": "
IN1k P205 Airc. CF10 CF100 CUB Cal101 Cars DTD Flow.Food Pets SUN VOC
[CLS]86.0 66.4 87.399.494.591.396.991.5 85.2 99.7 94.7 96.9 78.6 89.1
normal65.8 53.1 17.197.181.318.673.210.8 63.159.574.2 47.8 37.7 70.8
outlier69.0 55.179.199.393.784.997.685.2 84.999.6 93.5 94.1 78.589.7
", "type": "table", "image_path": "49e2e86cb5b67ef324d3ec8af60b38a0a51dd993441c9ea7d2a446f6ec43ef7b.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 107, 80, 504, 99.0 ], "spans": [], "index": 0 }, { "bbox": [ 107, 99.0, 504, 118.0 ], "spans": [], "index": 1 }, { "bbox": [ 107, 118.0, 504, 137.0 ], "spans": [], "index": 2 } ] } ], "index": 1 }, { "type": "text", "bbox": [ 107, 145, 505, 178 ], "lines": [ { "bbox": [ 106, 145, 506, 158 ], "spans": [ { "bbox": [ 106, 145, 506, 158 ], "score": 1.0, "content": "Table 1: Image classification via linear probing on normal and outlier patch tokens. We also report", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 156, 504, 167 ], "spans": [ { "bbox": [ 106, 156, 504, 167 ], "score": 1.0, "content": "the accuracy of classifiers learnt on the class token. We see that outlier tokens have a much higher", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 167, 477, 180 ], "spans": [ { "bbox": [ 106, 167, 477, 180 ], "score": 1.0, "content": "accuracy than regular ones, suggesting they are effectively storing global image information.", "type": "text" } ], "index": 5 } ], "index": 4, "bbox_fs": [ 106, 145, 506, 180 ] }, { "type": "image", "bbox": [ 139, 183, 472, 310 ], "blocks": [ { "type": "image_body", "bbox": [ 139, 183, 472, 310 ], "group_id": 0, "lines": [ { "bbox": [ 139, 183, 472, 310 ], "spans": [ { "bbox": [ 139, 183, 472, 310 ], "score": 0.972, "type": "image", "image_path": "1e848b87cde847a24e97afbf9af6f8da69a1f37248c2a7b358255f6d2278602c.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 139, 183, 472, 225.33333333333334 ], "spans": [], "index": 6 }, { "bbox": [ 139, 225.33333333333334, 472, 267.6666666666667 ], "spans": [], "index": 7 }, { "bbox": [ 139, 267.6666666666667, 472, 310.0 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 106, 319, 505, 353 ], "group_id": 0, "lines": [ { "bbox": [ 106, 319, 504, 331 ], "spans": [ { "bbox": [ 106, 319, 425, 331 ], "score": 1.0, "content": "Figure 6: Illustration of the proposed remediation and resulting model. We add", "type": "text" }, { "bbox": [ 425, 320, 436, 330 ], "score": 0.74, "content": "N", "type": "inline_equation" }, { "bbox": [ 436, 319, 504, 331 ], "score": 1.0, "content": "additional learn-", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 330, 505, 342 ], "spans": [ { "bbox": [ 105, 330, 505, 342 ], "score": 1.0, "content": "able input tokens (depicted in yellow), that the model can use as registers. At the output of the", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 342, 481, 354 ], "spans": [ { "bbox": [ 106, 342, 481, 354 ], "score": 1.0, "content": "model, only the patch tokens and [CLS] tokens are used, both during training and inference.", "type": "text" } ], "index": 11 } ], "index": 10 } ], "index": 8.5 }, { "type": "text", "bbox": [ 108, 365, 505, 432 ], "lines": [ { "bbox": [ 105, 364, 505, 378 ], "spans": [ { "bbox": [ 105, 364, 464, 378 ], "score": 1.0, "content": "benchmarks. For each image in a classification dataset, we forward it through DINOv2-", "type": "text" }, { "bbox": [ 465, 367, 471, 377 ], "score": 0.51, "content": "\\mathbf { g }", "type": "inline_equation" }, { "bbox": [ 471, 364, 505, 378 ], "score": 1.0, "content": "and ex-", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 376, 506, 389 ], "spans": [ { "bbox": [ 105, 376, 506, 389 ], "score": 1.0, "content": "tract the patch embeddings. From those, we choose a single token at random, either high-norm or", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 387, 505, 401 ], "spans": [ { "bbox": [ 105, 387, 505, 401 ], "score": 1.0, "content": "normal. This token is then considered as the image representation. We then train a logistic regres-", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 399, 505, 411 ], "spans": [ { "bbox": [ 106, 399, 505, 411 ], "score": 1.0, "content": "sion classifier to predict the image class from this representation, and measure the accuracy. We", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 409, 505, 422 ], "spans": [ { "bbox": [ 106, 409, 505, 422 ], "score": 1.0, "content": "observe that the high-norm tokens have a much higher accuracy than the other tokens (Table 1). This", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 420, 452, 433 ], "spans": [ { "bbox": [ 105, 420, 452, 433 ], "score": 1.0, "content": "suggests that outlier tokens contain more global information than other patch tokens.", "type": "text" } ], "index": 17 } ], "index": 14.5, "bbox_fs": [ 105, 364, 506, 433 ] }, { "type": "title", "bbox": [ 109, 447, 270, 457 ], "lines": [ { "bbox": [ 106, 447, 272, 459 ], "spans": [ { "bbox": [ 106, 447, 272, 459 ], "score": 1.0, "content": "2.2 HYPOTHESIS AND REMEDIATION", "type": "text" } ], "index": 18 } ], "index": 18 }, { "type": "text", "bbox": [ 108, 467, 504, 523 ], "lines": [ { "bbox": [ 106, 468, 505, 480 ], "spans": [ { "bbox": [ 106, 468, 505, 480 ], "score": 1.0, "content": "Having made these observations, we make the following hypothesis: large, sufficiently trained mod-", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 479, 505, 491 ], "spans": [ { "bbox": [ 106, 479, 505, 491 ], "score": 1.0, "content": "els learn to recognize redundant tokens, and to use them as places to store, process and retrieve", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 489, 506, 502 ], "spans": [ { "bbox": [ 105, 489, 506, 502 ], "score": 1.0, "content": "global information. Furthermore, we posit that while this behavior is not bad in itself, the fact that", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 500, 505, 512 ], "spans": [ { "bbox": [ 105, 500, 505, 512 ], "score": 1.0, "content": "it happens inside the patch tokens is undesirable. Indeed, it leads the model to discard local patch", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 512, 471, 524 ], "spans": [ { "bbox": [ 106, 512, 471, 524 ], "score": 1.0, "content": "information (Tab. 5b), possibly incurring decreased performance on dense prediction tasks.", "type": "text" } ], "index": 23 } ], "index": 21, "bbox_fs": [ 105, 468, 506, 524 ] }, { "type": "text", "bbox": [ 108, 528, 504, 606 ], "lines": [ { "bbox": [ 106, 529, 506, 541 ], "spans": [ { "bbox": [ 106, 529, 506, 541 ], "score": 1.0, "content": "We therefore propose a simple fix to this issue: we explicitly add new tokens to the sequence, that", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 540, 505, 551 ], "spans": [ { "bbox": [ 106, 540, 505, 551 ], "score": 1.0, "content": "the model can learn to use as registers. We add these tokens after the patch embedding layer, with a", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 550, 506, 563 ], "spans": [ { "bbox": [ 105, 550, 506, 563 ], "score": 1.0, "content": "learnable value, similarly to the [CLS] token. At the end of the vision transformer, these tokens are", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 561, 505, 574 ], "spans": [ { "bbox": [ 106, 561, 505, 574 ], "score": 1.0, "content": "discarded, and the [CLS] token and patch tokens are used as image representations, as usual. This", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 572, 505, 585 ], "spans": [ { "bbox": [ 105, 572, 505, 585 ], "score": 1.0, "content": "mechanism was first proposed in Memory Transformers (Burtsev et al., 2020), improving translation", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 583, 506, 596 ], "spans": [ { "bbox": [ 106, 583, 506, 596 ], "score": 1.0, "content": "tasks in NLP. Interestingly, we show here that this mechanism admits a natural justification for vision", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 595, 464, 607 ], "spans": [ { "bbox": [ 106, 595, 464, 607 ], "score": 1.0, "content": "transformers, fixing an interpretability and performance issue that was present otherwise.", "type": "text" } ], "index": 30 } ], "index": 27, "bbox_fs": [ 105, 529, 506, 607 ] }, { "type": "text", "bbox": [ 108, 611, 504, 656 ], "lines": [ { "bbox": [ 105, 610, 506, 624 ], "spans": [ { "bbox": [ 105, 610, 506, 624 ], "score": 1.0, "content": "We note that we have not been able to fully determine which aspects of the training led to the appear-", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 622, 506, 634 ], "spans": [ { "bbox": [ 105, 622, 506, 634 ], "score": 1.0, "content": "ance of artifacts in different models. The pretraining paradigm seems to play a role, as OpenCLIP", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 632, 506, 646 ], "spans": [ { "bbox": [ 105, 632, 506, 646 ], "score": 1.0, "content": "and DeiT-III exhibit outliers both at size B and L (Fig. 2). However, the model size and training", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 645, 327, 657 ], "spans": [ { "bbox": [ 106, 645, 327, 657 ], "score": 1.0, "content": "length also play important parts, as observed in Fig. 4.", "type": "text" } ], "index": 34 } ], "index": 32.5, "bbox_fs": [ 105, 610, 506, 657 ] }, { "type": "title", "bbox": [ 108, 673, 200, 685 ], "lines": [ { "bbox": [ 104, 672, 202, 688 ], "spans": [ { "bbox": [ 104, 672, 202, 688 ], "score": 1.0, "content": "3 EXPERIMENTS", "type": "text" } ], "index": 35 } ], "index": 35 }, { "type": "text", "bbox": [ 106, 699, 504, 732 ], "lines": [ { "bbox": [ 105, 698, 505, 712 ], "spans": [ { "bbox": [ 105, 698, 505, 712 ], "score": 1.0, "content": "In this section, we validate the proposed solution by training vision transformers with additional", "type": "text" } ], "index": 36 }, { "bbox": [ 107, 710, 505, 722 ], "spans": [ { "bbox": [ 107, 710, 505, 722 ], "score": 1.0, "content": "[reg] register tokens. We evaluate the effectiveness of our approach by a quantitative and quali-", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 721, 505, 733 ], "spans": [ { "bbox": [ 106, 721, 505, 733 ], "score": 1.0, "content": "tative analysis. We then ablate the number of registers used for training, to check that they do not", "type": "text" } ], "index": 38 } ], "index": 37, "bbox_fs": [ 105, 698, 505, 733 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 114, 84, 499, 154 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 84, 499, 154 ], "group_id": 0, "lines": [ { "bbox": [ 114, 84, 499, 154 ], "spans": [ { "bbox": [ 114, 84, 499, 154 ], "score": 0.955, "type": "image", "image_path": "446515d99f2641b90538058f2cf5805833a4609431bf97a4f7e791da412fd4ab.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 114, 84, 499, 107.33333333333333 ], "spans": [], "index": 0 }, { "bbox": [ 114, 107.33333333333333, 499, 130.66666666666666 ], "spans": [], "index": 1 }, { "bbox": [ 114, 130.66666666666666, 499, 154.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 167, 505, 190 ], "group_id": 0, "lines": [ { "bbox": [ 105, 167, 505, 180 ], "spans": [ { "bbox": [ 105, 167, 505, 180 ], "score": 1.0, "content": "Figure 7: Effect of register tokens on the distribution of output norms on DINOv2, OpenCLIP and", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 178, 499, 191 ], "spans": [ { "bbox": [ 106, 178, 499, 191 ], "score": 1.0, "content": "DeiT-III. Using register tokens effectively removes the norm outliers that were present previously.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "text", "bbox": [ 105, 212, 504, 235 ], "lines": [ { "bbox": [ 105, 211, 505, 225 ], "spans": [ { "bbox": [ 105, 211, 505, 225 ], "score": 1.0, "content": "cause a performance regression, evaluate an unsupervised object discovery method atop our features", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 223, 420, 236 ], "spans": [ { "bbox": [ 106, 223, 420, 236 ], "score": 1.0, "content": "and finally provide a qualitative analysis of the patterns learnt by the registers.", "type": "text" } ], "index": 6 } ], "index": 5.5 }, { "type": "title", "bbox": [ 108, 249, 281, 260 ], "lines": [ { "bbox": [ 105, 249, 282, 262 ], "spans": [ { "bbox": [ 105, 249, 282, 262 ], "score": 1.0, "content": "3.1 TRAINING ALGORITHMS AND DATA", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 108, 270, 504, 303 ], "lines": [ { "bbox": [ 106, 270, 504, 283 ], "spans": [ { "bbox": [ 106, 270, 504, 283 ], "score": 1.0, "content": "As the proposed solution is a simple architectural change, we can easily apply it to any training pro-", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 281, 505, 293 ], "spans": [ { "bbox": [ 106, 281, 505, 293 ], "score": 1.0, "content": "cedure. We try it on three different state-of-the-art training methods for supervised, text-supervised,", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 292, 315, 304 ], "spans": [ { "bbox": [ 106, 292, 315, 304 ], "score": 1.0, "content": "and unsupervised learning, shortly described below.", "type": "text" } ], "index": 10 } ], "index": 9 }, { "type": "text", "bbox": [ 107, 308, 504, 364 ], "lines": [ { "bbox": [ 106, 309, 505, 321 ], "spans": [ { "bbox": [ 106, 309, 505, 321 ], "score": 1.0, "content": "DEIT-III (Touvron et al., 2022) is a simple and robust supervised training recipe for classification", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 320, 505, 332 ], "spans": [ { "bbox": [ 106, 320, 505, 332 ], "score": 1.0, "content": "with ViTs on ImageNet-1k and ImageNet-22k. We choose this method as an example of label-", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 331, 506, 344 ], "spans": [ { "bbox": [ 105, 331, 506, 344 ], "score": 1.0, "content": "supervised training as it is simple, uses the base ViT architecture, achieves strong classification", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 342, 506, 355 ], "spans": [ { "bbox": [ 106, 342, 506, 355 ], "score": 1.0, "content": "results, and is easy to reproduce and modify with our improvements. We run this method on the", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 352, 456, 366 ], "spans": [ { "bbox": [ 106, 352, 456, 366 ], "score": 1.0, "content": "ImageNet-22k dataset, using the ViT-B settings, as provided in the official repository 1.", "type": "text" } ], "index": 15 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 369, 504, 436 ], "lines": [ { "bbox": [ 106, 370, 505, 383 ], "spans": [ { "bbox": [ 106, 370, 505, 383 ], "score": 1.0, "content": "OpenCLIP (Ilharco et al., 2021) is a strong training method for producing text-image aligned mod-", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 381, 505, 393 ], "spans": [ { "bbox": [ 106, 381, 505, 393 ], "score": 1.0, "content": "els, following the original CLIP work. We chose this method as an example of text-supervised", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 392, 505, 403 ], "spans": [ { "bbox": [ 106, 392, 505, 403 ], "score": 1.0, "content": "training because it is open-source, uses the base ViT architecture, and is easy to reproduce and mod-", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 402, 505, 415 ], "spans": [ { "bbox": [ 106, 402, 505, 415 ], "score": 1.0, "content": "ify with our improvements. We run the OpenCLIP method on a text-image-aligned corpus based on", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 412, 506, 427 ], "spans": [ { "bbox": [ 105, 412, 506, 427 ], "score": 1.0, "content": "Shutterstock that includes only licensed image and text data. We use a ViT-B/16 image encoder, as", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 424, 252, 437 ], "spans": [ { "bbox": [ 104, 424, 252, 437 ], "score": 1.0, "content": "proposed in the official repository 2.", "type": "text" } ], "index": 21 } ], "index": 18.5 }, { "type": "text", "bbox": [ 107, 441, 504, 475 ], "lines": [ { "bbox": [ 105, 441, 506, 454 ], "spans": [ { "bbox": [ 105, 441, 506, 454 ], "score": 1.0, "content": "DINOv2 (Oquab et al., 2023) is a self-supervised method for learning visual features, following the", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 452, 506, 465 ], "spans": [ { "bbox": [ 106, 452, 506, 465 ], "score": 1.0, "content": "DINO work. We apply our changes to this method as it is the main focus of our study. We run this", "type": "text" } ], "index": 23 }, { "bbox": [ 104, 462, 461, 477 ], "spans": [ { "bbox": [ 104, 462, 461, 477 ], "score": 1.0, "content": "method on ImageNet-22k with the ViT-L configuration. We use the official repository 3.", "type": "text" } ], "index": 24 } ], "index": 23 }, { "type": "title", "bbox": [ 108, 489, 313, 500 ], "lines": [ { "bbox": [ 106, 489, 315, 502 ], "spans": [ { "bbox": [ 106, 489, 315, 502 ], "score": 1.0, "content": "3.2 EVALUATION OF THE PROPOSED SOLUTION", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 108, 510, 505, 576 ], "lines": [ { "bbox": [ 105, 509, 505, 523 ], "spans": [ { "bbox": [ 105, 509, 505, 523 ], "score": 1.0, "content": "As shown in Fig. 1, we get rid of the artifacts by training models with additional register tokens.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 521, 506, 534 ], "spans": [ { "bbox": [ 105, 521, 506, 534 ], "score": 1.0, "content": "In the appendix, we provide additional qualitative results for more images in Fig. 19. In order to", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 531, 506, 545 ], "spans": [ { "bbox": [ 105, 531, 506, 545 ], "score": 1.0, "content": "quantitatively measure this effect, for each model, we probe the norm of features at the output of", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 543, 505, 555 ], "spans": [ { "bbox": [ 106, 543, 505, 555 ], "score": 1.0, "content": "the model. We report these norms for all three algorithms with and without registers in Fig. 7. We", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 554, 505, 566 ], "spans": [ { "bbox": [ 105, 554, 505, 566 ], "score": 1.0, "content": "see that when training with registers, models do not exhibit large-norm tokens at the output, which", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 565, 277, 577 ], "spans": [ { "bbox": [ 106, 565, 277, 577 ], "score": 1.0, "content": "confirms the initial qualitative assessment.", "type": "text" } ], "index": 31 } ], "index": 28.5 }, { "type": "text", "bbox": [ 107, 582, 504, 691 ], "lines": [ { "bbox": [ 105, 581, 505, 594 ], "spans": [ { "bbox": [ 105, 581, 505, 594 ], "score": 1.0, "content": "Performance regression. In the previous section, we have shown that the proposed approach re-", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 593, 506, 605 ], "spans": [ { "bbox": [ 105, 593, 506, 605 ], "score": 1.0, "content": "moves artifacts from local feature maps. In this experiment, we want to check that the use of register", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 604, 505, 617 ], "spans": [ { "bbox": [ 105, 604, 505, 617 ], "score": 1.0, "content": "tokens does not affect the representation quality of those features. We run linear probing on Im-", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 614, 506, 627 ], "spans": [ { "bbox": [ 105, 614, 506, 627 ], "score": 1.0, "content": "ageNet classification, ADE20k Segmentation, and NYUd monocular depth estimation. We follow", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 626, 505, 637 ], "spans": [ { "bbox": [ 106, 626, 505, 637 ], "score": 1.0, "content": "the experimental protocol outlined in Oquab et al. (2023). We summarize the performance of the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 636, 506, 649 ], "spans": [ { "bbox": [ 105, 636, 506, 649 ], "score": 1.0, "content": "models described in Sec. 3.1 with and without register tokens in Table 2a. We see that when us-", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 648, 506, 661 ], "spans": [ { "bbox": [ 105, 648, 506, 661 ], "score": 1.0, "content": "ing registers, models do not lose performance and sometimes even work better. For completeness,", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 658, 506, 672 ], "spans": [ { "bbox": [ 105, 658, 506, 672 ], "score": 1.0, "content": "we also provided the zero-shot classification performance on ImageNet for OpenCLIP (Table 2b),", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 668, 506, 682 ], "spans": [ { "bbox": [ 105, 668, 506, 682 ], "score": 1.0, "content": "which remains unchanged. Please note that the absolute performance of our OpenCLIP reproduction", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 681, 267, 692 ], "spans": [ { "bbox": [ 105, 681, 267, 692 ], "score": 1.0, "content": "is lower due to the data source we used.", "type": "text" } ], "index": 41 } ], "index": 36.5 } ], "page_idx": 5, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 118, 699, 351, 732 ], "lines": [ { "bbox": [ 119, 698, 340, 711 ], "spans": [ { "bbox": [ 119, 698, 340, 711 ], "score": 1.0, "content": "1https://github.com/facebookresearch/deit", "type": "text" } ] }, { "bbox": [ 118, 708, 352, 723 ], "spans": [ { "bbox": [ 118, 708, 352, 723 ], "score": 1.0, "content": "2https://github.com/mlfoundations/open_clip", "type": "text" } ] }, { "bbox": [ 119, 720, 351, 733 ], "spans": [ { "bbox": [ 119, 720, 351, 733 ], "score": 1.0, "content": "3https://github.com/facebookresearch/dinov2", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 294, 38 ], "spans": [ { "bbox": [ 106, 25, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "6", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 114, 84, 499, 154 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 84, 499, 154 ], "group_id": 0, "lines": [ { "bbox": [ 114, 84, 499, 154 ], "spans": [ { "bbox": [ 114, 84, 499, 154 ], "score": 0.955, "type": "image", "image_path": "446515d99f2641b90538058f2cf5805833a4609431bf97a4f7e791da412fd4ab.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 114, 84, 499, 107.33333333333333 ], "spans": [], "index": 0 }, { "bbox": [ 114, 107.33333333333333, 499, 130.66666666666666 ], "spans": [], "index": 1 }, { "bbox": [ 114, 130.66666666666666, 499, 154.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 167, 505, 190 ], "group_id": 0, "lines": [ { "bbox": [ 105, 167, 505, 180 ], "spans": [ { "bbox": [ 105, 167, 505, 180 ], "score": 1.0, "content": "Figure 7: Effect of register tokens on the distribution of output norms on DINOv2, OpenCLIP and", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 178, 499, 191 ], "spans": [ { "bbox": [ 106, 178, 499, 191 ], "score": 1.0, "content": "DeiT-III. Using register tokens effectively removes the norm outliers that were present previously.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "text", "bbox": [ 105, 212, 504, 235 ], "lines": [ { "bbox": [ 105, 211, 505, 225 ], "spans": [ { "bbox": [ 105, 211, 505, 225 ], "score": 1.0, "content": "cause a performance regression, evaluate an unsupervised object discovery method atop our features", "type": "text" } ], "index": 5 }, { "bbox": [ 106, 223, 420, 236 ], "spans": [ { "bbox": [ 106, 223, 420, 236 ], "score": 1.0, "content": "and finally provide a qualitative analysis of the patterns learnt by the registers.", "type": "text" } ], "index": 6 } ], "index": 5.5, "bbox_fs": [ 105, 211, 505, 236 ] }, { "type": "title", "bbox": [ 108, 249, 281, 260 ], "lines": [ { "bbox": [ 105, 249, 282, 262 ], "spans": [ { "bbox": [ 105, 249, 282, 262 ], "score": 1.0, "content": "3.1 TRAINING ALGORITHMS AND DATA", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 108, 270, 504, 303 ], "lines": [ { "bbox": [ 106, 270, 504, 283 ], "spans": [ { "bbox": [ 106, 270, 504, 283 ], "score": 1.0, "content": "As the proposed solution is a simple architectural change, we can easily apply it to any training pro-", "type": "text" } ], "index": 8 }, { "bbox": [ 106, 281, 505, 293 ], "spans": [ { "bbox": [ 106, 281, 505, 293 ], "score": 1.0, "content": "cedure. We try it on three different state-of-the-art training methods for supervised, text-supervised,", "type": "text" } ], "index": 9 }, { "bbox": [ 106, 292, 315, 304 ], "spans": [ { "bbox": [ 106, 292, 315, 304 ], "score": 1.0, "content": "and unsupervised learning, shortly described below.", "type": "text" } ], "index": 10 } ], "index": 9, "bbox_fs": [ 106, 270, 505, 304 ] }, { "type": "text", "bbox": [ 107, 308, 504, 364 ], "lines": [ { "bbox": [ 106, 309, 505, 321 ], "spans": [ { "bbox": [ 106, 309, 505, 321 ], "score": 1.0, "content": "DEIT-III (Touvron et al., 2022) is a simple and robust supervised training recipe for classification", "type": "text" } ], "index": 11 }, { "bbox": [ 106, 320, 505, 332 ], "spans": [ { "bbox": [ 106, 320, 505, 332 ], "score": 1.0, "content": "with ViTs on ImageNet-1k and ImageNet-22k. We choose this method as an example of label-", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 331, 506, 344 ], "spans": [ { "bbox": [ 105, 331, 506, 344 ], "score": 1.0, "content": "supervised training as it is simple, uses the base ViT architecture, achieves strong classification", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 342, 506, 355 ], "spans": [ { "bbox": [ 106, 342, 506, 355 ], "score": 1.0, "content": "results, and is easy to reproduce and modify with our improvements. We run this method on the", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 352, 456, 366 ], "spans": [ { "bbox": [ 106, 352, 456, 366 ], "score": 1.0, "content": "ImageNet-22k dataset, using the ViT-B settings, as provided in the official repository 1.", "type": "text" } ], "index": 15 } ], "index": 13, "bbox_fs": [ 105, 309, 506, 366 ] }, { "type": "text", "bbox": [ 107, 369, 504, 436 ], "lines": [ { "bbox": [ 106, 370, 505, 383 ], "spans": [ { "bbox": [ 106, 370, 505, 383 ], "score": 1.0, "content": "OpenCLIP (Ilharco et al., 2021) is a strong training method for producing text-image aligned mod-", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 381, 505, 393 ], "spans": [ { "bbox": [ 106, 381, 505, 393 ], "score": 1.0, "content": "els, following the original CLIP work. We chose this method as an example of text-supervised", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 392, 505, 403 ], "spans": [ { "bbox": [ 106, 392, 505, 403 ], "score": 1.0, "content": "training because it is open-source, uses the base ViT architecture, and is easy to reproduce and mod-", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 402, 505, 415 ], "spans": [ { "bbox": [ 106, 402, 505, 415 ], "score": 1.0, "content": "ify with our improvements. We run the OpenCLIP method on a text-image-aligned corpus based on", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 412, 506, 427 ], "spans": [ { "bbox": [ 105, 412, 506, 427 ], "score": 1.0, "content": "Shutterstock that includes only licensed image and text data. We use a ViT-B/16 image encoder, as", "type": "text" } ], "index": 20 }, { "bbox": [ 104, 424, 252, 437 ], "spans": [ { "bbox": [ 104, 424, 252, 437 ], "score": 1.0, "content": "proposed in the official repository 2.", "type": "text" } ], "index": 21 } ], "index": 18.5, "bbox_fs": [ 104, 370, 506, 437 ] }, { "type": "text", "bbox": [ 107, 441, 504, 475 ], "lines": [ { "bbox": [ 105, 441, 506, 454 ], "spans": [ { "bbox": [ 105, 441, 506, 454 ], "score": 1.0, "content": "DINOv2 (Oquab et al., 2023) is a self-supervised method for learning visual features, following the", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 452, 506, 465 ], "spans": [ { "bbox": [ 106, 452, 506, 465 ], "score": 1.0, "content": "DINO work. We apply our changes to this method as it is the main focus of our study. We run this", "type": "text" } ], "index": 23 }, { "bbox": [ 104, 462, 461, 477 ], "spans": [ { "bbox": [ 104, 462, 461, 477 ], "score": 1.0, "content": "method on ImageNet-22k with the ViT-L configuration. We use the official repository 3.", "type": "text" } ], "index": 24 } ], "index": 23, "bbox_fs": [ 104, 441, 506, 477 ] }, { "type": "title", "bbox": [ 108, 489, 313, 500 ], "lines": [ { "bbox": [ 106, 489, 315, 502 ], "spans": [ { "bbox": [ 106, 489, 315, 502 ], "score": 1.0, "content": "3.2 EVALUATION OF THE PROPOSED SOLUTION", "type": "text" } ], "index": 25 } ], "index": 25 }, { "type": "text", "bbox": [ 108, 510, 505, 576 ], "lines": [ { "bbox": [ 105, 509, 505, 523 ], "spans": [ { "bbox": [ 105, 509, 505, 523 ], "score": 1.0, "content": "As shown in Fig. 1, we get rid of the artifacts by training models with additional register tokens.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 521, 506, 534 ], "spans": [ { "bbox": [ 105, 521, 506, 534 ], "score": 1.0, "content": "In the appendix, we provide additional qualitative results for more images in Fig. 19. In order to", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 531, 506, 545 ], "spans": [ { "bbox": [ 105, 531, 506, 545 ], "score": 1.0, "content": "quantitatively measure this effect, for each model, we probe the norm of features at the output of", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 543, 505, 555 ], "spans": [ { "bbox": [ 106, 543, 505, 555 ], "score": 1.0, "content": "the model. We report these norms for all three algorithms with and without registers in Fig. 7. We", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 554, 505, 566 ], "spans": [ { "bbox": [ 105, 554, 505, 566 ], "score": 1.0, "content": "see that when training with registers, models do not exhibit large-norm tokens at the output, which", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 565, 277, 577 ], "spans": [ { "bbox": [ 106, 565, 277, 577 ], "score": 1.0, "content": "confirms the initial qualitative assessment.", "type": "text" } ], "index": 31 } ], "index": 28.5, "bbox_fs": [ 105, 509, 506, 577 ] }, { "type": "text", "bbox": [ 107, 582, 504, 691 ], "lines": [ { "bbox": [ 105, 581, 505, 594 ], "spans": [ { "bbox": [ 105, 581, 505, 594 ], "score": 1.0, "content": "Performance regression. In the previous section, we have shown that the proposed approach re-", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 593, 506, 605 ], "spans": [ { "bbox": [ 105, 593, 506, 605 ], "score": 1.0, "content": "moves artifacts from local feature maps. In this experiment, we want to check that the use of register", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 604, 505, 617 ], "spans": [ { "bbox": [ 105, 604, 505, 617 ], "score": 1.0, "content": "tokens does not affect the representation quality of those features. We run linear probing on Im-", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 614, 506, 627 ], "spans": [ { "bbox": [ 105, 614, 506, 627 ], "score": 1.0, "content": "ageNet classification, ADE20k Segmentation, and NYUd monocular depth estimation. We follow", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 626, 505, 637 ], "spans": [ { "bbox": [ 106, 626, 505, 637 ], "score": 1.0, "content": "the experimental protocol outlined in Oquab et al. (2023). We summarize the performance of the", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 636, 506, 649 ], "spans": [ { "bbox": [ 105, 636, 506, 649 ], "score": 1.0, "content": "models described in Sec. 3.1 with and without register tokens in Table 2a. We see that when us-", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 648, 506, 661 ], "spans": [ { "bbox": [ 105, 648, 506, 661 ], "score": 1.0, "content": "ing registers, models do not lose performance and sometimes even work better. For completeness,", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 658, 506, 672 ], "spans": [ { "bbox": [ 105, 658, 506, 672 ], "score": 1.0, "content": "we also provided the zero-shot classification performance on ImageNet for OpenCLIP (Table 2b),", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 668, 506, 682 ], "spans": [ { "bbox": [ 105, 668, 506, 682 ], "score": 1.0, "content": "which remains unchanged. Please note that the absolute performance of our OpenCLIP reproduction", "type": "text" } ], "index": 40 }, { "bbox": [ 105, 681, 267, 692 ], "spans": [ { "bbox": [ 105, 681, 267, 692 ], "score": 1.0, "content": "is lower due to the data source we used.", "type": "text" } ], "index": 41 } ], "index": 36.5, "bbox_fs": [ 105, 581, 506, 692 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 368, 133, 496, 191 ], "blocks": [ { "type": "table_body", "bbox": [ 368, 133, 496, 191 ], "group_id": 1, "lines": [ { "bbox": [ 368, 133, 496, 191 ], "spans": [ { "bbox": [ 368, 133, 496, 191 ], "score": 0.968, "html": "
ImageNet Top-1
OpenCLIP59.9
OpenCLIP+reg60.1
", "type": "table", "image_path": "10d6ede5ca6c5fea7577bc3c9c781c4b2ddadde96fb491fe3271a30dd682260e.jpg" } ] } ], "index": 7.5, "virtual_lines": [ { "bbox": [ 368, 133, 496, 162.0 ], "spans": [], "index": 6 }, { "bbox": [ 368, 162.0, 496, 191.0 ], "spans": [], "index": 9 } ] }, { "type": "table_footnote", "bbox": [ 380, 195, 480, 205 ], "group_id": 1, "lines": [ { "bbox": [ 379, 194, 482, 207 ], "spans": [ { "bbox": [ 379, 194, 482, 207 ], "score": 1.0, "content": "(b) Zero-shot classification.", "type": "text" } ], "index": 11 } ], "index": 11 } ], "index": 9.25 }, { "type": "table", "bbox": [ 117, 79, 331, 191 ], "blocks": [ { "type": "table_body", "bbox": [ 117, 79, 331, 191 ], "group_id": 0, "lines": [ { "bbox": [ 117, 79, 331, 191 ], "spans": [ { "bbox": [ 117, 79, 331, 191 ], "score": 0.975, "html": "
ImageNet Top-1ADE20k mIoUNYUd rmse ↓
DeiT-III84.738.90.511
DeiT-I+reg84.739.10.512
OpenCLIP78.226.60.702
OpenCLIP+reg78.126.70.661
DINOv284.346.60.378
DINOv2+reg84.847.90.366
", "type": "table", "image_path": "332f74eb05493a3de8c7d8ea740219ca428913c0a44cd5895b7b947891913775.jpg" } ] } ], "index": 3.5, "virtual_lines": [ { "bbox": [ 117, 79, 331, 93.0 ], "spans": [], "index": 0 }, { "bbox": [ 117, 93.0, 331, 107.0 ], "spans": [], "index": 1 }, { "bbox": [ 117, 107.0, 331, 121.0 ], "spans": [], "index": 2 }, { "bbox": [ 117, 121.0, 331, 135.0 ], "spans": [], "index": 3 }, { "bbox": [ 117, 135.0, 331, 149.0 ], "spans": [], "index": 4 }, { "bbox": [ 117, 149.0, 331, 163.0 ], "spans": [], "index": 5 }, { "bbox": [ 117, 163.0, 331, 177.0 ], "spans": [], "index": 7 }, { "bbox": [ 117, 177.0, 331, 191.0 ], "spans": [], "index": 8 } ] }, { "type": "table_footnote", "bbox": [ 146, 195, 299, 206 ], "group_id": 0, "lines": [ { "bbox": [ 146, 195, 300, 206 ], "spans": [ { "bbox": [ 146, 195, 300, 206 ], "score": 1.0, "content": "(a) Linear evaluation with frozen features.", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "table_caption", "bbox": [ 107, 215, 505, 259 ], "group_id": 0, "lines": [ { "bbox": [ 105, 214, 506, 227 ], "spans": [ { "bbox": [ 105, 214, 506, 227 ], "score": 1.0, "content": "Table 2: Evaluation of downstream performance of the models that we trained, with and without", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 226, 505, 238 ], "spans": [ { "bbox": [ 105, 226, 505, 238 ], "score": 1.0, "content": "registers. We consider linear probing of frozen features for all three models, and zero-shot evaluation", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 237, 505, 249 ], "spans": [ { "bbox": [ 105, 237, 505, 249 ], "score": 1.0, "content": "for the OpenCLIP model. We see that using register not only does not degrade performance, but even", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 248, 288, 261 ], "spans": [ { "bbox": [ 105, 248, 288, 261 ], "score": 1.0, "content": "improves it by a slight margin in some cases.", "type": "text" } ], "index": 15 } ], "index": 13.5 } ], "index": 10 }, { "type": "image", "bbox": [ 107, 272, 503, 428 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 272, 503, 428 ], "group_id": 0, "lines": [ { "bbox": [ 107, 272, 503, 428 ], "spans": [ { "bbox": [ 107, 272, 503, 428 ], "score": 0.968, "type": "image", "image_path": "fa1d33877e15d880d2c1a54171159e99b4bbebfb16192fc8dfcc61e6eef59dec.jpg" } ] } ], "index": 17, "virtual_lines": [ { "bbox": [ 107, 272, 503, 324.0 ], "spans": [], "index": 16 }, { "bbox": [ 107, 324.0, 503, 376.0 ], "spans": [], "index": 17 }, { "bbox": [ 107, 376.0, 503, 428.0 ], "spans": [], "index": 18 } ] }, { "type": "image_caption", "bbox": [ 106, 439, 504, 484 ], "group_id": 0, "lines": [ { "bbox": [ 105, 439, 506, 452 ], "spans": [ { "bbox": [ 105, 439, 506, 452 ], "score": 1.0, "content": "Figure 8: Ablation of the the number of register tokens used with a DINOv2 model. (top): qualita-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 450, 506, 464 ], "spans": [ { "bbox": [ 105, 450, 506, 464 ], "score": 1.0, "content": "tive visualization of artifacts appearing as a function of number of registers. (bottom): performance", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 461, 506, 474 ], "spans": [ { "bbox": [ 105, 461, 506, 474 ], "score": 1.0, "content": "on three tasks (ImageNet, ADE-20k and NYUd) as a function of number of registers used. While", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 473, 506, 486 ], "spans": [ { "bbox": [ 105, 473, 506, 486 ], "score": 1.0, "content": "one register is sufficient to remove artefacts, using more leads to improved downstream performance.", "type": "text" } ], "index": 22 } ], "index": 20.5 } ], "index": 18.75 }, { "type": "text", "bbox": [ 107, 507, 504, 618 ], "lines": [ { "bbox": [ 105, 507, 504, 520 ], "spans": [ { "bbox": [ 105, 507, 504, 520 ], "score": 1.0, "content": "Number of register tokens. As described in Sec. 2.2, we propose alleviating the feature maps’", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 518, 506, 531 ], "spans": [ { "bbox": [ 105, 518, 506, 531 ], "score": 1.0, "content": "artifacts by adding register tokens. In this experiment, we study the influence of the number of such", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 529, 506, 542 ], "spans": [ { "bbox": [ 105, 529, 506, 542 ], "score": 1.0, "content": "tokens on local features and downstream performance. We train DINOv2 ViT-L/14 models with 0, 1,", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 540, 506, 554 ], "spans": [ { "bbox": [ 105, 540, 506, 554 ], "score": 1.0, "content": "2, 4, 8 or 16 registers. In Fig. 8, we report the results of this analysis. In Fig. 8(top), we qualitatively", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 551, 506, 564 ], "spans": [ { "bbox": [ 105, 551, 506, 564 ], "score": 1.0, "content": "study the attention maps and observe that the visible artifacts disappear when adding at least one", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 562, 506, 575 ], "spans": [ { "bbox": [ 105, 562, 506, 575 ], "score": 1.0, "content": "register. We then examine in Fig. 8(bottom) performance on downstream evaluation benchmarks,", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 573, 506, 586 ], "spans": [ { "bbox": [ 105, 573, 506, 586 ], "score": 1.0, "content": "following the protocol from Oquab et al. (2023). There seems to be an optimal number of registers", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 583, 506, 598 ], "spans": [ { "bbox": [ 105, 583, 506, 598 ], "score": 1.0, "content": "for dense tasks, and adding one brings most of the benefit. This optimum is likely explained by", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 594, 506, 608 ], "spans": [ { "bbox": [ 105, 594, 506, 608 ], "score": 1.0, "content": "the disappearance of artifacts, leading to better local features. On ImageNet, however, performance", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 607, 453, 619 ], "spans": [ { "bbox": [ 106, 607, 453, 619 ], "score": 1.0, "content": "improves when using more registers. In all our experiments, we kept 4 register tokens.", "type": "text" } ], "index": 32 } ], "index": 27.5 }, { "type": "title", "bbox": [ 108, 634, 217, 644 ], "lines": [ { "bbox": [ 105, 632, 218, 646 ], "spans": [ { "bbox": [ 105, 632, 218, 646 ], "score": 1.0, "content": "3.3 OBJECT DISCOVERY", "type": "text" } ], "index": 33 } ], "index": 33 }, { "type": "text", "bbox": [ 108, 654, 504, 732 ], "lines": [ { "bbox": [ 106, 655, 506, 667 ], "spans": [ { "bbox": [ 106, 655, 506, 667 ], "score": 1.0, "content": "Recent unsupervised object discovery methods rely on the quality and smoothness of local feature", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 666, 506, 678 ], "spans": [ { "bbox": [ 106, 666, 506, 678 ], "score": 1.0, "content": "maps (Simeoni et al., 2021; Wang et al., 2023). By leveraging DINO Caron et al. (2021), these ´", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 677, 506, 689 ], "spans": [ { "bbox": [ 106, 677, 506, 689 ], "score": 1.0, "content": "methods have significantly surpassed the previous state of the art. However, the algorithm leads", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 688, 506, 700 ], "spans": [ { "bbox": [ 105, 688, 506, 700 ], "score": 1.0, "content": "to poor performance when applied to modern backbones such as DINOv2 Oquab et al. (2023) or", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 699, 506, 711 ], "spans": [ { "bbox": [ 105, 699, 506, 711 ], "score": 1.0, "content": "supervised ones Touvron et al. (2022). We posit that this can be alleviated by the method proposed", "type": "text" } ], "index": 38 }, { "bbox": [ 104, 707, 506, 723 ], "spans": [ { "bbox": [ 104, 707, 506, 723 ], "score": 1.0, "content": "in this work. We run LOST (Simeoni et al., 2021) on features extracted from backbones trained using ´", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 720, 506, 734 ], "spans": [ { "bbox": [ 106, 720, 506, 734 ], "score": 1.0, "content": "the algorithms described in Sec.3.1 with and without registers. We run object discovery on PASCAL", "type": "text" } ], "index": 40 } ], "index": 37 } ], "page_idx": 6, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 294, 38 ], "spans": [ { "bbox": [ 106, 25, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 303, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 763 ], "spans": [ { "bbox": [ 302, 751, 309, 763 ], "score": 1.0, "content": "", "type": "text", "height": 12, "width": 7 } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 368, 133, 496, 191 ], "blocks": [ { "type": "table_body", "bbox": [ 368, 133, 496, 191 ], "group_id": 1, "lines": [ { "bbox": [ 368, 133, 496, 191 ], "spans": [ { "bbox": [ 368, 133, 496, 191 ], "score": 0.968, "html": "
ImageNet Top-1
OpenCLIP59.9
OpenCLIP+reg60.1
", "type": "table", "image_path": "10d6ede5ca6c5fea7577bc3c9c781c4b2ddadde96fb491fe3271a30dd682260e.jpg" } ] } ], "index": 7.5, "virtual_lines": [ { "bbox": [ 368, 133, 496, 162.0 ], "spans": [], "index": 6 }, { "bbox": [ 368, 162.0, 496, 191.0 ], "spans": [], "index": 9 } ] }, { "type": "table_footnote", "bbox": [ 380, 195, 480, 205 ], "group_id": 1, "lines": [ { "bbox": [ 379, 194, 482, 207 ], "spans": [ { "bbox": [ 379, 194, 482, 207 ], "score": 1.0, "content": "(b) Zero-shot classification.", "type": "text" } ], "index": 11 } ], "index": 11 } ], "index": 9.25 }, { "type": "table", "bbox": [ 117, 79, 331, 191 ], "blocks": [ { "type": "table_body", "bbox": [ 117, 79, 331, 191 ], "group_id": 0, "lines": [ { "bbox": [ 117, 79, 331, 191 ], "spans": [ { "bbox": [ 117, 79, 331, 191 ], "score": 0.975, "html": "
ImageNet Top-1ADE20k mIoUNYUd rmse ↓
DeiT-III84.738.90.511
DeiT-I+reg84.739.10.512
OpenCLIP78.226.60.702
OpenCLIP+reg78.126.70.661
DINOv284.346.60.378
DINOv2+reg84.847.90.366
", "type": "table", "image_path": "332f74eb05493a3de8c7d8ea740219ca428913c0a44cd5895b7b947891913775.jpg" } ] } ], "index": 3.5, "virtual_lines": [ { "bbox": [ 117, 79, 331, 93.0 ], "spans": [], "index": 0 }, { "bbox": [ 117, 93.0, 331, 107.0 ], "spans": [], "index": 1 }, { "bbox": [ 117, 107.0, 331, 121.0 ], "spans": [], "index": 2 }, { "bbox": [ 117, 121.0, 331, 135.0 ], "spans": [], "index": 3 }, { "bbox": [ 117, 135.0, 331, 149.0 ], "spans": [], "index": 4 }, { "bbox": [ 117, 149.0, 331, 163.0 ], "spans": [], "index": 5 }, { "bbox": [ 117, 163.0, 331, 177.0 ], "spans": [], "index": 7 }, { "bbox": [ 117, 177.0, 331, 191.0 ], "spans": [], "index": 8 } ] }, { "type": "table_footnote", "bbox": [ 146, 195, 299, 206 ], "group_id": 0, "lines": [ { "bbox": [ 146, 195, 300, 206 ], "spans": [ { "bbox": [ 146, 195, 300, 206 ], "score": 1.0, "content": "(a) Linear evaluation with frozen features.", "type": "text" } ], "index": 10 } ], "index": 10 }, { "type": "table_caption", "bbox": [ 107, 215, 505, 259 ], "group_id": 0, "lines": [ { "bbox": [ 105, 214, 506, 227 ], "spans": [ { "bbox": [ 105, 214, 506, 227 ], "score": 1.0, "content": "Table 2: Evaluation of downstream performance of the models that we trained, with and without", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 226, 505, 238 ], "spans": [ { "bbox": [ 105, 226, 505, 238 ], "score": 1.0, "content": "registers. We consider linear probing of frozen features for all three models, and zero-shot evaluation", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 237, 505, 249 ], "spans": [ { "bbox": [ 105, 237, 505, 249 ], "score": 1.0, "content": "for the OpenCLIP model. We see that using register not only does not degrade performance, but even", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 248, 288, 261 ], "spans": [ { "bbox": [ 105, 248, 288, 261 ], "score": 1.0, "content": "improves it by a slight margin in some cases.", "type": "text" } ], "index": 15 } ], "index": 13.5 } ], "index": 10 }, { "type": "image", "bbox": [ 107, 272, 503, 428 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 272, 503, 428 ], "group_id": 0, "lines": [ { "bbox": [ 107, 272, 503, 428 ], "spans": [ { "bbox": [ 107, 272, 503, 428 ], "score": 0.968, "type": "image", "image_path": "fa1d33877e15d880d2c1a54171159e99b4bbebfb16192fc8dfcc61e6eef59dec.jpg" } ] } ], "index": 17, "virtual_lines": [ { "bbox": [ 107, 272, 503, 324.0 ], "spans": [], "index": 16 }, { "bbox": [ 107, 324.0, 503, 376.0 ], "spans": [], "index": 17 }, { "bbox": [ 107, 376.0, 503, 428.0 ], "spans": [], "index": 18 } ] }, { "type": "image_caption", "bbox": [ 106, 439, 504, 484 ], "group_id": 0, "lines": [ { "bbox": [ 105, 439, 506, 452 ], "spans": [ { "bbox": [ 105, 439, 506, 452 ], "score": 1.0, "content": "Figure 8: Ablation of the the number of register tokens used with a DINOv2 model. (top): qualita-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 450, 506, 464 ], "spans": [ { "bbox": [ 105, 450, 506, 464 ], "score": 1.0, "content": "tive visualization of artifacts appearing as a function of number of registers. (bottom): performance", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 461, 506, 474 ], "spans": [ { "bbox": [ 105, 461, 506, 474 ], "score": 1.0, "content": "on three tasks (ImageNet, ADE-20k and NYUd) as a function of number of registers used. While", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 473, 506, 486 ], "spans": [ { "bbox": [ 105, 473, 506, 486 ], "score": 1.0, "content": "one register is sufficient to remove artefacts, using more leads to improved downstream performance.", "type": "text" } ], "index": 22 } ], "index": 20.5 } ], "index": 18.75 }, { "type": "text", "bbox": [ 107, 507, 504, 618 ], "lines": [ { "bbox": [ 105, 507, 504, 520 ], "spans": [ { "bbox": [ 105, 507, 504, 520 ], "score": 1.0, "content": "Number of register tokens. As described in Sec. 2.2, we propose alleviating the feature maps’", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 518, 506, 531 ], "spans": [ { "bbox": [ 105, 518, 506, 531 ], "score": 1.0, "content": "artifacts by adding register tokens. In this experiment, we study the influence of the number of such", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 529, 506, 542 ], "spans": [ { "bbox": [ 105, 529, 506, 542 ], "score": 1.0, "content": "tokens on local features and downstream performance. We train DINOv2 ViT-L/14 models with 0, 1,", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 540, 506, 554 ], "spans": [ { "bbox": [ 105, 540, 506, 554 ], "score": 1.0, "content": "2, 4, 8 or 16 registers. In Fig. 8, we report the results of this analysis. In Fig. 8(top), we qualitatively", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 551, 506, 564 ], "spans": [ { "bbox": [ 105, 551, 506, 564 ], "score": 1.0, "content": "study the attention maps and observe that the visible artifacts disappear when adding at least one", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 562, 506, 575 ], "spans": [ { "bbox": [ 105, 562, 506, 575 ], "score": 1.0, "content": "register. We then examine in Fig. 8(bottom) performance on downstream evaluation benchmarks,", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 573, 506, 586 ], "spans": [ { "bbox": [ 105, 573, 506, 586 ], "score": 1.0, "content": "following the protocol from Oquab et al. (2023). There seems to be an optimal number of registers", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 583, 506, 598 ], "spans": [ { "bbox": [ 105, 583, 506, 598 ], "score": 1.0, "content": "for dense tasks, and adding one brings most of the benefit. This optimum is likely explained by", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 594, 506, 608 ], "spans": [ { "bbox": [ 105, 594, 506, 608 ], "score": 1.0, "content": "the disappearance of artifacts, leading to better local features. On ImageNet, however, performance", "type": "text" } ], "index": 31 }, { "bbox": [ 106, 607, 453, 619 ], "spans": [ { "bbox": [ 106, 607, 453, 619 ], "score": 1.0, "content": "improves when using more registers. In all our experiments, we kept 4 register tokens.", "type": "text" } ], "index": 32 } ], "index": 27.5, "bbox_fs": [ 105, 507, 506, 619 ] }, { "type": "title", "bbox": [ 108, 634, 217, 644 ], "lines": [ { "bbox": [ 105, 632, 218, 646 ], "spans": [ { "bbox": [ 105, 632, 218, 646 ], "score": 1.0, "content": "3.3 OBJECT DISCOVERY", "type": "text" } ], "index": 33 } ], "index": 33 }, { "type": "text", "bbox": [ 108, 654, 504, 732 ], "lines": [ { "bbox": [ 106, 655, 506, 667 ], "spans": [ { "bbox": [ 106, 655, 506, 667 ], "score": 1.0, "content": "Recent unsupervised object discovery methods rely on the quality and smoothness of local feature", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 666, 506, 678 ], "spans": [ { "bbox": [ 106, 666, 506, 678 ], "score": 1.0, "content": "maps (Simeoni et al., 2021; Wang et al., 2023). By leveraging DINO Caron et al. (2021), these ´", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 677, 506, 689 ], "spans": [ { "bbox": [ 106, 677, 506, 689 ], "score": 1.0, "content": "methods have significantly surpassed the previous state of the art. However, the algorithm leads", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 688, 506, 700 ], "spans": [ { "bbox": [ 105, 688, 506, 700 ], "score": 1.0, "content": "to poor performance when applied to modern backbones such as DINOv2 Oquab et al. (2023) or", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 699, 506, 711 ], "spans": [ { "bbox": [ 105, 699, 506, 711 ], "score": 1.0, "content": "supervised ones Touvron et al. (2022). We posit that this can be alleviated by the method proposed", "type": "text" } ], "index": 38 }, { "bbox": [ 104, 707, 506, 723 ], "spans": [ { "bbox": [ 104, 707, 506, 723 ], "score": 1.0, "content": "in this work. We run LOST (Simeoni et al., 2021) on features extracted from backbones trained using ´", "type": "text" } ], "index": 39 }, { "bbox": [ 106, 720, 506, 734 ], "spans": [ { "bbox": [ 106, 720, 506, 734 ], "score": 1.0, "content": "the algorithms described in Sec.3.1 with and without registers. We run object discovery on PASCAL", "type": "text" } ], "index": 40 } ], "index": 37, "bbox_fs": [ 104, 655, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 183, 79, 427, 180 ], "blocks": [ { "type": "table_body", "bbox": [ 183, 79, 427, 180 ], "group_id": 0, "lines": [ { "bbox": [ 183, 79, 427, 180 ], "spans": [ { "bbox": [ 183, 79, 427, 180 ], "score": 0.976, "html": "
VOC 2007VOC 2012COCO 20k
DeiT-II11.713.110.7
DeiT-III+reg27.132.725.1
OpenCLIP38.844.331.0
OpenCLIP+reg37.142.027.9
DINOv235.340.226.9
DINOv2+reg55.460.042.0
", "type": "table", "image_path": "4261f3917619e8d2f913892fa0f1a0ac72327d8b2adc15fd7b81b41c3df9fbd6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 183, 79, 427, 112.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 183, 112.66666666666666, 427, 146.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 183, 146.33333333333331, 427, 179.99999999999997 ], "spans": [], "index": 2 } ] } ], "index": 1 }, { "type": "text", "bbox": [ 106, 187, 505, 233 ], "lines": [ { "bbox": [ 105, 188, 506, 201 ], "spans": [ { "bbox": [ 105, 188, 506, 201 ], "score": 1.0, "content": "Table 3: Unsupervised Object Discovery using LOST (Simeoni et al., 2021) on models with and ´", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 200, 505, 212 ], "spans": [ { "bbox": [ 106, 200, 505, 212 ], "score": 1.0, "content": "without registers. We evaluated three types of models trained with various amounts of supervision", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 209, 506, 223 ], "spans": [ { "bbox": [ 105, 209, 506, 223 ], "score": 1.0, "content": "on VOC 2007, 2012 and COCO. We measure performance using corloc. We observe that adding", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 221, 460, 235 ], "spans": [ { "bbox": [ 105, 221, 460, 235 ], "score": 1.0, "content": "register tokens makes all models significantly more viable for usage in object discovery.", "type": "text" } ], "index": 6 } ], "index": 4.5 }, { "type": "image", "bbox": [ 113, 245, 483, 309 ], "blocks": [ { "type": "image_body", "bbox": [ 113, 245, 483, 309 ], "group_id": 0, "lines": [ { "bbox": [ 113, 245, 483, 309 ], "spans": [ { "bbox": [ 113, 245, 483, 309 ], "score": 0.965, "type": "image", "image_path": "36c5b56960e84369a75b7d40990a2cb567d7aca1aa3ac7050a668615432a42be.jpg" } ] } ], "index": 8, "virtual_lines": [ { "bbox": [ 113, 245, 483, 266.3333333333333 ], "spans": [], "index": 7 }, { "bbox": [ 113, 266.3333333333333, 483, 287.66666666666663 ], "spans": [], "index": 8 }, { "bbox": [ 113, 287.66666666666663, 483, 308.99999999999994 ], "spans": [], "index": 9 } ] }, { "type": "image_caption", "bbox": [ 105, 320, 505, 354 ], "group_id": 0, "lines": [ { "bbox": [ 106, 321, 505, 333 ], "spans": [ { "bbox": [ 106, 321, 505, 333 ], "score": 1.0, "content": "Figure 9: Comparison of the attention maps of the [CLS] and register tokens. Register tokens", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 332, 506, 344 ], "spans": [ { "bbox": [ 105, 332, 506, 344 ], "score": 1.0, "content": "sometimes attend to different parts of the feature map, similarly to slot attention (Locatello et al.,", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 342, 491, 356 ], "spans": [ { "bbox": [ 105, 342, 491, 356 ], "score": 1.0, "content": "2020). This behaviour was never required from the model, and emerged naturally from training.", "type": "text" } ], "index": 12 } ], "index": 11 } ], "index": 9.5 }, { "type": "text", "bbox": [ 107, 377, 505, 454 ], "lines": [ { "bbox": [ 106, 376, 505, 390 ], "spans": [ { "bbox": [ 106, 376, 238, 390 ], "score": 1.0, "content": "VOC 2007 and 2012 and COCO", "type": "text" }, { "bbox": [ 239, 377, 255, 388 ], "score": 0.44, "content": "2 0 \\mathrm { k }", "type": "inline_equation" }, { "bbox": [ 255, 376, 505, 390 ], "score": 1.0, "content": ". We use values for DeiT and OpenCLIP, and for DINOv2, we", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 389, 505, 400 ], "spans": [ { "bbox": [ 106, 389, 505, 400 ], "score": 1.0, "content": "use keys. Because the output features may have different conditioning, we manually add a bias to", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 398, 506, 411 ], "spans": [ { "bbox": [ 105, 398, 506, 411 ], "score": 1.0, "content": "the gram matrix of features. The results of this experiment are presented in Table 3. For DINOv2", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 410, 506, 424 ], "spans": [ { "bbox": [ 105, 410, 506, 424 ], "score": 1.0, "content": "and DeiT-III, adding registers significantly improves the discovery performance. For OpenCLIP, the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 421, 505, 433 ], "spans": [ { "bbox": [ 105, 421, 505, 433 ], "score": 1.0, "content": "performance is slighty worse with registers (see Sec. C for analysis). The performance of DINOv2", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 432, 506, 445 ], "spans": [ { "bbox": [ 105, 432, 506, 445 ], "score": 1.0, "content": "on VOC2007 still does not match that of DINO as reported by Simeoni et al. (2021) ( ´ 61.9 corloc).", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 443, 466, 455 ], "spans": [ { "bbox": [ 106, 443, 466, 455 ], "score": 1.0, "content": "However, the model with registers gets an improvement of 20.1 corloc (55.4 versus 35.3).", "type": "text" } ], "index": 19 } ], "index": 16 }, { "type": "title", "bbox": [ 109, 470, 310, 480 ], "lines": [ { "bbox": [ 106, 469, 312, 483 ], "spans": [ { "bbox": [ 106, 469, 312, 483 ], "score": 1.0, "content": "3.4 QUALITATIVE EVALUATION OF REGISTERS", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 490, 505, 568 ], "lines": [ { "bbox": [ 105, 491, 505, 503 ], "spans": [ { "bbox": [ 105, 491, 505, 503 ], "score": 1.0, "content": "In this final experiment, we qualitatively probe for the behavior of register tokens. We want to verify", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 502, 505, 514 ], "spans": [ { "bbox": [ 105, 502, 505, 514 ], "score": 1.0, "content": "if they all exhibit similar attention patterns or whether a differentiation automatically emerges. To", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 512, 506, 525 ], "spans": [ { "bbox": [ 106, 512, 506, 525 ], "score": 1.0, "content": "this end, we plot the attention maps of the class and register tokens to patch tokens. The result of", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 523, 505, 536 ], "spans": [ { "bbox": [ 105, 523, 505, 536 ], "score": 1.0, "content": "this visualization is shown in Fig. 9. We see that registers do not have a completely aligned behavior.", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 534, 505, 547 ], "spans": [ { "bbox": [ 106, 534, 505, 547 ], "score": 1.0, "content": "Some selected registers exhibit interesting attention patterns, attending to the different objects in the", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 546, 505, 558 ], "spans": [ { "bbox": [ 105, 546, 505, 558 ], "score": 1.0, "content": "scene. While nothing enforced this behavior, their activations had some natural diversity. We leave", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 557, 341, 568 ], "spans": [ { "bbox": [ 106, 557, 341, 568 ], "score": 1.0, "content": "the study of the regularization of registers for future work.", "type": "text" } ], "index": 27 } ], "index": 24 }, { "type": "title", "bbox": [ 108, 585, 211, 598 ], "lines": [ { "bbox": [ 105, 584, 213, 599 ], "spans": [ { "bbox": [ 105, 584, 213, 599 ], "score": 1.0, "content": "4 RELATED WORK", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 107, 610, 504, 732 ], "lines": [ { "bbox": [ 105, 610, 505, 624 ], "spans": [ { "bbox": [ 105, 610, 505, 624 ], "score": 1.0, "content": "Feature extraction with pretrained models. Using pretrained neural network models for extracting", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 622, 505, 633 ], "spans": [ { "bbox": [ 106, 622, 505, 633 ], "score": 1.0, "content": "visual features has stood the test of time since the AlexNet (Krizhevsky et al., 2012) CNN model", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 633, 505, 646 ], "spans": [ { "bbox": [ 105, 633, 505, 646 ], "score": 1.0, "content": "pretrained on ImageNet-1k (Russakovsky et al., 2015). More recent models have upgraded the same", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 644, 506, 657 ], "spans": [ { "bbox": [ 105, 644, 506, 657 ], "score": 1.0, "content": "setup with modern architectures, such as ResNets (used in, e.g., DETR, Carion et al., 2020) or even", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 653, 506, 669 ], "spans": [ { "bbox": [ 105, 653, 506, 669 ], "score": 1.0, "content": "Vision Transformers. As Transformers are easily able to handle different modalities during training,", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 665, 506, 680 ], "spans": [ { "bbox": [ 105, 665, 506, 680 ], "score": 1.0, "content": "off-the-shelf backbones are now commonly trained on label supervision (e.g., DeiT-III on ImageNet-", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 675, 506, 691 ], "spans": [ { "bbox": [ 106, 677, 123, 687 ], "score": 0.59, "content": "2 2 \\mathrm { k }", "type": "inline_equation" }, { "bbox": [ 123, 675, 506, 691 ], "score": 1.0, "content": ", Touvron et al., 2022) or text supervision (e.g., CLIP (Radford et al., 2021)), providing strong", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 687, 506, 701 ], "spans": [ { "bbox": [ 105, 687, 506, 701 ], "score": 1.0, "content": "visual foundation models, scaling well with model sizes, and enabling excellent performance on", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 506, 712 ], "score": 1.0, "content": "a variety of tasks including detection (Carion et al., 2020) and segmentation (Zheng et al., 2021;", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 710, 506, 721 ], "spans": [ { "bbox": [ 106, 710, 506, 721 ], "score": 1.0, "content": "Kirillov et al., 2023). In this context, supervision relies on annotations in the form of labels or text", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 721, 506, 734 ], "spans": [ { "bbox": [ 106, 721, 506, 734 ], "score": 1.0, "content": "alignment; the dataset biases (Torralba & Efros, 2011) are not well characterized, yet they drive", "type": "text" } ], "index": 39 } ], "index": 34 } ], "page_idx": 7, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 300, 750, 309, 762 ], "spans": [ { "bbox": [ 300, 750, 309, 762 ], "score": 1.0, "content": "", "type": "text", "height": 12, "width": 9 } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 183, 79, 427, 180 ], "blocks": [ { "type": "table_body", "bbox": [ 183, 79, 427, 180 ], "group_id": 0, "lines": [ { "bbox": [ 183, 79, 427, 180 ], "spans": [ { "bbox": [ 183, 79, 427, 180 ], "score": 0.976, "html": "
VOC 2007VOC 2012COCO 20k
DeiT-II11.713.110.7
DeiT-III+reg27.132.725.1
OpenCLIP38.844.331.0
OpenCLIP+reg37.142.027.9
DINOv235.340.226.9
DINOv2+reg55.460.042.0
", "type": "table", "image_path": "4261f3917619e8d2f913892fa0f1a0ac72327d8b2adc15fd7b81b41c3df9fbd6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 183, 79, 427, 112.66666666666666 ], "spans": [], "index": 0 }, { "bbox": [ 183, 112.66666666666666, 427, 146.33333333333331 ], "spans": [], "index": 1 }, { "bbox": [ 183, 146.33333333333331, 427, 179.99999999999997 ], "spans": [], "index": 2 } ] } ], "index": 1 }, { "type": "text", "bbox": [ 106, 187, 505, 233 ], "lines": [ { "bbox": [ 105, 188, 506, 201 ], "spans": [ { "bbox": [ 105, 188, 506, 201 ], "score": 1.0, "content": "Table 3: Unsupervised Object Discovery using LOST (Simeoni et al., 2021) on models with and ´", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 200, 505, 212 ], "spans": [ { "bbox": [ 106, 200, 505, 212 ], "score": 1.0, "content": "without registers. We evaluated three types of models trained with various amounts of supervision", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 209, 506, 223 ], "spans": [ { "bbox": [ 105, 209, 506, 223 ], "score": 1.0, "content": "on VOC 2007, 2012 and COCO. We measure performance using corloc. We observe that adding", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 221, 460, 235 ], "spans": [ { "bbox": [ 105, 221, 460, 235 ], "score": 1.0, "content": "register tokens makes all models significantly more viable for usage in object discovery.", "type": "text" } ], "index": 6 } ], "index": 4.5, "bbox_fs": [ 105, 188, 506, 235 ] }, { "type": "image", "bbox": [ 113, 245, 483, 309 ], "blocks": [ { "type": "image_body", "bbox": [ 113, 245, 483, 309 ], "group_id": 0, "lines": [ { "bbox": [ 113, 245, 483, 309 ], "spans": [ { "bbox": [ 113, 245, 483, 309 ], "score": 0.965, "type": "image", "image_path": "36c5b56960e84369a75b7d40990a2cb567d7aca1aa3ac7050a668615432a42be.jpg" } ] } ], "index": 8, "virtual_lines": [ { "bbox": [ 113, 245, 483, 266.3333333333333 ], "spans": [], "index": 7 }, { "bbox": [ 113, 266.3333333333333, 483, 287.66666666666663 ], "spans": [], "index": 8 }, { "bbox": [ 113, 287.66666666666663, 483, 308.99999999999994 ], "spans": [], "index": 9 } ] }, { "type": "image_caption", "bbox": [ 105, 320, 505, 354 ], "group_id": 0, "lines": [ { "bbox": [ 106, 321, 505, 333 ], "spans": [ { "bbox": [ 106, 321, 505, 333 ], "score": 1.0, "content": "Figure 9: Comparison of the attention maps of the [CLS] and register tokens. Register tokens", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 332, 506, 344 ], "spans": [ { "bbox": [ 105, 332, 506, 344 ], "score": 1.0, "content": "sometimes attend to different parts of the feature map, similarly to slot attention (Locatello et al.,", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 342, 491, 356 ], "spans": [ { "bbox": [ 105, 342, 491, 356 ], "score": 1.0, "content": "2020). This behaviour was never required from the model, and emerged naturally from training.", "type": "text" } ], "index": 12 } ], "index": 11 } ], "index": 9.5 }, { "type": "text", "bbox": [ 107, 377, 505, 454 ], "lines": [ { "bbox": [ 106, 376, 505, 390 ], "spans": [ { "bbox": [ 106, 376, 238, 390 ], "score": 1.0, "content": "VOC 2007 and 2012 and COCO", "type": "text" }, { "bbox": [ 239, 377, 255, 388 ], "score": 0.44, "content": "2 0 \\mathrm { k }", "type": "inline_equation" }, { "bbox": [ 255, 376, 505, 390 ], "score": 1.0, "content": ". We use values for DeiT and OpenCLIP, and for DINOv2, we", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 389, 505, 400 ], "spans": [ { "bbox": [ 106, 389, 505, 400 ], "score": 1.0, "content": "use keys. Because the output features may have different conditioning, we manually add a bias to", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 398, 506, 411 ], "spans": [ { "bbox": [ 105, 398, 506, 411 ], "score": 1.0, "content": "the gram matrix of features. The results of this experiment are presented in Table 3. For DINOv2", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 410, 506, 424 ], "spans": [ { "bbox": [ 105, 410, 506, 424 ], "score": 1.0, "content": "and DeiT-III, adding registers significantly improves the discovery performance. For OpenCLIP, the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 421, 505, 433 ], "spans": [ { "bbox": [ 105, 421, 505, 433 ], "score": 1.0, "content": "performance is slighty worse with registers (see Sec. C for analysis). The performance of DINOv2", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 432, 506, 445 ], "spans": [ { "bbox": [ 105, 432, 506, 445 ], "score": 1.0, "content": "on VOC2007 still does not match that of DINO as reported by Simeoni et al. (2021) ( ´ 61.9 corloc).", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 443, 466, 455 ], "spans": [ { "bbox": [ 106, 443, 466, 455 ], "score": 1.0, "content": "However, the model with registers gets an improvement of 20.1 corloc (55.4 versus 35.3).", "type": "text" } ], "index": 19 } ], "index": 16, "bbox_fs": [ 105, 376, 506, 455 ] }, { "type": "title", "bbox": [ 109, 470, 310, 480 ], "lines": [ { "bbox": [ 106, 469, 312, 483 ], "spans": [ { "bbox": [ 106, 469, 312, 483 ], "score": 1.0, "content": "3.4 QUALITATIVE EVALUATION OF REGISTERS", "type": "text" } ], "index": 20 } ], "index": 20 }, { "type": "text", "bbox": [ 107, 490, 505, 568 ], "lines": [ { "bbox": [ 105, 491, 505, 503 ], "spans": [ { "bbox": [ 105, 491, 505, 503 ], "score": 1.0, "content": "In this final experiment, we qualitatively probe for the behavior of register tokens. We want to verify", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 502, 505, 514 ], "spans": [ { "bbox": [ 105, 502, 505, 514 ], "score": 1.0, "content": "if they all exhibit similar attention patterns or whether a differentiation automatically emerges. To", "type": "text" } ], "index": 22 }, { "bbox": [ 106, 512, 506, 525 ], "spans": [ { "bbox": [ 106, 512, 506, 525 ], "score": 1.0, "content": "this end, we plot the attention maps of the class and register tokens to patch tokens. The result of", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 523, 505, 536 ], "spans": [ { "bbox": [ 105, 523, 505, 536 ], "score": 1.0, "content": "this visualization is shown in Fig. 9. We see that registers do not have a completely aligned behavior.", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 534, 505, 547 ], "spans": [ { "bbox": [ 106, 534, 505, 547 ], "score": 1.0, "content": "Some selected registers exhibit interesting attention patterns, attending to the different objects in the", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 546, 505, 558 ], "spans": [ { "bbox": [ 105, 546, 505, 558 ], "score": 1.0, "content": "scene. While nothing enforced this behavior, their activations had some natural diversity. We leave", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 557, 341, 568 ], "spans": [ { "bbox": [ 106, 557, 341, 568 ], "score": 1.0, "content": "the study of the regularization of registers for future work.", "type": "text" } ], "index": 27 } ], "index": 24, "bbox_fs": [ 105, 491, 506, 568 ] }, { "type": "title", "bbox": [ 108, 585, 211, 598 ], "lines": [ { "bbox": [ 105, 584, 213, 599 ], "spans": [ { "bbox": [ 105, 584, 213, 599 ], "score": 1.0, "content": "4 RELATED WORK", "type": "text" } ], "index": 28 } ], "index": 28 }, { "type": "text", "bbox": [ 107, 610, 504, 732 ], "lines": [ { "bbox": [ 105, 610, 505, 624 ], "spans": [ { "bbox": [ 105, 610, 505, 624 ], "score": 1.0, "content": "Feature extraction with pretrained models. Using pretrained neural network models for extracting", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 622, 505, 633 ], "spans": [ { "bbox": [ 106, 622, 505, 633 ], "score": 1.0, "content": "visual features has stood the test of time since the AlexNet (Krizhevsky et al., 2012) CNN model", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 633, 505, 646 ], "spans": [ { "bbox": [ 105, 633, 505, 646 ], "score": 1.0, "content": "pretrained on ImageNet-1k (Russakovsky et al., 2015). More recent models have upgraded the same", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 644, 506, 657 ], "spans": [ { "bbox": [ 105, 644, 506, 657 ], "score": 1.0, "content": "setup with modern architectures, such as ResNets (used in, e.g., DETR, Carion et al., 2020) or even", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 653, 506, 669 ], "spans": [ { "bbox": [ 105, 653, 506, 669 ], "score": 1.0, "content": "Vision Transformers. As Transformers are easily able to handle different modalities during training,", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 665, 506, 680 ], "spans": [ { "bbox": [ 105, 665, 506, 680 ], "score": 1.0, "content": "off-the-shelf backbones are now commonly trained on label supervision (e.g., DeiT-III on ImageNet-", "type": "text" } ], "index": 34 }, { "bbox": [ 106, 675, 506, 691 ], "spans": [ { "bbox": [ 106, 677, 123, 687 ], "score": 0.59, "content": "2 2 \\mathrm { k }", "type": "inline_equation" }, { "bbox": [ 123, 675, 506, 691 ], "score": 1.0, "content": ", Touvron et al., 2022) or text supervision (e.g., CLIP (Radford et al., 2021)), providing strong", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 687, 506, 701 ], "spans": [ { "bbox": [ 105, 687, 506, 701 ], "score": 1.0, "content": "visual foundation models, scaling well with model sizes, and enabling excellent performance on", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 506, 712 ], "score": 1.0, "content": "a variety of tasks including detection (Carion et al., 2020) and segmentation (Zheng et al., 2021;", "type": "text" } ], "index": 37 }, { "bbox": [ 106, 710, 506, 721 ], "spans": [ { "bbox": [ 106, 710, 506, 721 ], "score": 1.0, "content": "Kirillov et al., 2023). In this context, supervision relies on annotations in the form of labels or text", "type": "text" } ], "index": 38 }, { "bbox": [ 106, 721, 506, 734 ], "spans": [ { "bbox": [ 106, 721, 506, 734 ], "score": 1.0, "content": "alignment; the dataset biases (Torralba & Efros, 2011) are not well characterized, yet they drive", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 83, 506, 95 ], "spans": [ { "bbox": [ 105, 83, 506, 95 ], "score": 1.0, "content": "learning and shape the learned models. An alternative approach consists of not using supervision", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 105, 93, 506, 107 ], "spans": [ { "bbox": [ 105, 93, 506, 107 ], "score": 1.0, "content": "and letting the models learn from the data via a pretext task that is designed to require understanding", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 105, 105, 506, 117 ], "spans": [ { "bbox": [ 105, 105, 506, 117 ], "score": 1.0, "content": "the content of images (Doersch et al., 2015). This self-supervised learning paradigm was explored in", "type": "text", "cross_page": true } ], "index": 2 }, { "bbox": [ 105, 114, 506, 129 ], "spans": [ { "bbox": [ 105, 114, 506, 129 ], "score": 1.0, "content": "multiple methods using Vision Transformers: MAE (He et al., 2022) trains a model at reconstructing", "type": "text", "cross_page": true } ], "index": 3 }, { "bbox": [ 105, 127, 506, 139 ], "spans": [ { "bbox": [ 105, 127, 506, 139 ], "score": 1.0, "content": "pixel values of hidden areas of an image and then applies fine-tuning to address a new task. With a", "type": "text", "cross_page": true } ], "index": 4 }, { "bbox": [ 105, 137, 506, 150 ], "spans": [ { "bbox": [ 105, 137, 506, 150 ], "score": 1.0, "content": "different approach, the self-distillation family of methods (He et al., 2020; Caron et al., 2021; Zhou", "type": "text", "cross_page": true } ], "index": 5 }, { "bbox": [ 105, 149, 506, 161 ], "spans": [ { "bbox": [ 105, 149, 506, 161 ], "score": 1.0, "content": "et al., 2022) showcase strong performance using frozen backbones, allowing for more robustness to", "type": "text", "cross_page": true } ], "index": 6 }, { "bbox": [ 106, 160, 505, 171 ], "spans": [ { "bbox": [ 106, 160, 505, 171 ], "score": 1.0, "content": "domain shifts for task-specific downstream models. In this work, we focused the analysis on self-", "type": "text", "cross_page": true } ], "index": 7 }, { "bbox": [ 105, 170, 506, 183 ], "spans": [ { "bbox": [ 105, 170, 506, 183 ], "score": 1.0, "content": "supervised learning, and more specifically on the DINOv2 approach (Oquab et al., 2023), which", "type": "text", "cross_page": true } ], "index": 8 }, { "bbox": [ 105, 181, 506, 194 ], "spans": [ { "bbox": [ 105, 181, 506, 194 ], "score": 1.0, "content": "has shown to be particularly effective for learning local features. We showed that despite excellent", "type": "text", "cross_page": true } ], "index": 9 }, { "bbox": [ 105, 192, 506, 205 ], "spans": [ { "bbox": [ 105, 192, 506, 205 ], "score": 1.0, "content": "benchmark scores, DINOv2 features exhibit undesirable artifacts and that correcting these artifacts", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 105, 204, 506, 215 ], "spans": [ { "bbox": [ 105, 204, 506, 215 ], "score": 1.0, "content": "in the learning process allows for further improvements in the benchmark performances. These", "type": "text", "cross_page": true } ], "index": 11 }, { "bbox": [ 105, 214, 506, 227 ], "spans": [ { "bbox": [ 105, 214, 506, 227 ], "score": 1.0, "content": "phenomenon is even more surprising as DINOv2 builds upon DINO (Caron et al., 2021), which", "type": "text", "cross_page": true } ], "index": 12 }, { "bbox": [ 105, 225, 506, 238 ], "spans": [ { "bbox": [ 105, 225, 506, 238 ], "score": 1.0, "content": "does not show signs of artifacts. We then further showed that the correction techniques hold for", "type": "text", "cross_page": true } ], "index": 13 }, { "bbox": [ 105, 236, 349, 248 ], "spans": [ { "bbox": [ 105, 236, 349, 248 ], "score": 1.0, "content": "supervised paradigms by testing on DeiT-III and OpenCLIP.", "type": "text", "cross_page": true } ], "index": 14 } ], "index": 34, "bbox_fs": [ 105, 610, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 107, 82, 504, 247 ], "lines": [ { "bbox": [ 105, 83, 506, 95 ], "spans": [ { "bbox": [ 105, 83, 506, 95 ], "score": 1.0, "content": "learning and shape the learned models. An alternative approach consists of not using supervision", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 93, 506, 107 ], "spans": [ { "bbox": [ 105, 93, 506, 107 ], "score": 1.0, "content": "and letting the models learn from the data via a pretext task that is designed to require understanding", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 105, 506, 117 ], "spans": [ { "bbox": [ 105, 105, 506, 117 ], "score": 1.0, "content": "the content of images (Doersch et al., 2015). This self-supervised learning paradigm was explored in", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 114, 506, 129 ], "spans": [ { "bbox": [ 105, 114, 506, 129 ], "score": 1.0, "content": "multiple methods using Vision Transformers: MAE (He et al., 2022) trains a model at reconstructing", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 127, 506, 139 ], "spans": [ { "bbox": [ 105, 127, 506, 139 ], "score": 1.0, "content": "pixel values of hidden areas of an image and then applies fine-tuning to address a new task. With a", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 137, 506, 150 ], "spans": [ { "bbox": [ 105, 137, 506, 150 ], "score": 1.0, "content": "different approach, the self-distillation family of methods (He et al., 2020; Caron et al., 2021; Zhou", "type": "text" } ], "index": 5 }, { "bbox": [ 105, 149, 506, 161 ], "spans": [ { "bbox": [ 105, 149, 506, 161 ], "score": 1.0, "content": "et al., 2022) showcase strong performance using frozen backbones, allowing for more robustness to", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 160, 505, 171 ], "spans": [ { "bbox": [ 106, 160, 505, 171 ], "score": 1.0, "content": "domain shifts for task-specific downstream models. In this work, we focused the analysis on self-", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 170, 506, 183 ], "spans": [ { "bbox": [ 105, 170, 506, 183 ], "score": 1.0, "content": "supervised learning, and more specifically on the DINOv2 approach (Oquab et al., 2023), which", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 181, 506, 194 ], "spans": [ { "bbox": [ 105, 181, 506, 194 ], "score": 1.0, "content": "has shown to be particularly effective for learning local features. We showed that despite excellent", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 192, 506, 205 ], "spans": [ { "bbox": [ 105, 192, 506, 205 ], "score": 1.0, "content": "benchmark scores, DINOv2 features exhibit undesirable artifacts and that correcting these artifacts", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 204, 506, 215 ], "spans": [ { "bbox": [ 105, 204, 506, 215 ], "score": 1.0, "content": "in the learning process allows for further improvements in the benchmark performances. These", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 214, 506, 227 ], "spans": [ { "bbox": [ 105, 214, 506, 227 ], "score": 1.0, "content": "phenomenon is even more surprising as DINOv2 builds upon DINO (Caron et al., 2021), which", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 225, 506, 238 ], "spans": [ { "bbox": [ 105, 225, 506, 238 ], "score": 1.0, "content": "does not show signs of artifacts. We then further showed that the correction techniques hold for", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 236, 349, 248 ], "spans": [ { "bbox": [ 105, 236, 349, 248 ], "score": 1.0, "content": "supervised paradigms by testing on DeiT-III and OpenCLIP.", "type": "text" } ], "index": 14 } ], "index": 7 }, { "type": "text", "bbox": [ 107, 253, 504, 483 ], "lines": [ { "bbox": [ 105, 253, 506, 266 ], "spans": [ { "bbox": [ 105, 253, 506, 266 ], "score": 1.0, "content": "Additional tokens in transformers. Extending the transformer sequence with special tokens was", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 263, 506, 278 ], "spans": [ { "bbox": [ 105, 263, 506, 278 ], "score": 1.0, "content": "popularized in BERT (Devlin et al., 2019). However, most approaches add new tokens either to pro-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 274, 505, 288 ], "spans": [ { "bbox": [ 105, 274, 505, 288 ], "score": 1.0, "content": "vide the network with new information as for example [SEP] tokens in BERT, provide opportunity", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 285, 506, 299 ], "spans": [ { "bbox": [ 105, 285, 506, 299 ], "score": 1.0, "content": "to spend more computation on the input as seen with the tape tokens in AdaTape (Xue et al., 2023),", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 297, 506, 309 ], "spans": [ { "bbox": [ 105, 297, 506, 309 ], "score": 1.0, "content": "or to gather information in these tokens, and use their output value as an output of the model: for", "type": "text" } ], "index": 19 }, { "bbox": [ 104, 306, 506, 322 ], "spans": [ { "bbox": [ 104, 306, 506, 322 ], "score": 1.0, "content": "classification, as [CLS] tokens in BERT and ViT (Dosovitskiy et al., 2021); for generative learning,", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 318, 506, 331 ], "spans": [ { "bbox": [ 105, 318, 506, 331 ], "score": 1.0, "content": "as [MASK] in BERT and BEiT (Bao et al., 2021); for detection, as object queries in DETR (Carion", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 330, 506, 343 ], "spans": [ { "bbox": [ 105, 330, 506, 343 ], "score": 1.0, "content": "et al., 2020), detection tokens in YOLOS (Fang et al., 2021), and ViDT (Song et al., 2021); or for", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 340, 505, 354 ], "spans": [ { "bbox": [ 105, 340, 505, 354 ], "score": 1.0, "content": "accumulating information from possibly multiple modalities before decoding, as latent token arrays", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 351, 505, 365 ], "spans": [ { "bbox": [ 105, 351, 505, 365 ], "score": 1.0, "content": "in Perceivers (Jaegle et al., 2021; 2022). Different to these works, the tokens we add to the sequence", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 362, 505, 376 ], "spans": [ { "bbox": [ 105, 362, 505, 376 ], "score": 1.0, "content": "add no information, and their output value is not used for any purpose. They are simply registers", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 374, 505, 387 ], "spans": [ { "bbox": [ 105, 374, 505, 387 ], "score": 1.0, "content": "where the model can learn to store and retrieve information during the forward pass. The Mem-", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 384, 506, 398 ], "spans": [ { "bbox": [ 105, 384, 506, 398 ], "score": 1.0, "content": "ory Transformer (Burtsev et al., 2020), closer to our work, presents a simple approach to improve", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 395, 505, 408 ], "spans": [ { "bbox": [ 105, 395, 505, 408 ], "score": 1.0, "content": "transformer models using memory tokens added to the token sequence, improving translation perfor-", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 406, 505, 419 ], "spans": [ { "bbox": [ 105, 406, 505, 419 ], "score": 1.0, "content": "mance. In follow-up work, Bulatov et al. (2022) address complex copy-repeat-reverse tasks. Sandler", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 417, 506, 430 ], "spans": [ { "bbox": [ 105, 417, 506, 430 ], "score": 1.0, "content": "et al. (2022) extend this line to the vision domain for fine-tuning but observe that such tokens do not", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 429, 506, 442 ], "spans": [ { "bbox": [ 105, 429, 506, 442 ], "score": 1.0, "content": "transfer well across tasks. In contrast, we do not perform fine-tuning and employ additional tokens", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 439, 506, 452 ], "spans": [ { "bbox": [ 105, 439, 506, 452 ], "score": 1.0, "content": "during pretraining to improve the features obtained for all tasks downstream. More importantly, our", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 449, 506, 464 ], "spans": [ { "bbox": [ 105, 449, 506, 464 ], "score": 1.0, "content": "study contributes the following new insight in Sec. 2: the mechanism implemented through memory", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 461, 506, 474 ], "spans": [ { "bbox": [ 105, 461, 506, 474 ], "score": 1.0, "content": "tokens already appears naturally in Vision Transformers; our study shows that such tokens allow us", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 472, 455, 485 ], "spans": [ { "bbox": [ 105, 472, 455, 485 ], "score": 1.0, "content": "not to create but to isolate this existing behavior, and thus avoid collateral side-effects.", "type": "text" } ], "index": 35 } ], "index": 25 }, { "type": "text", "bbox": [ 107, 489, 504, 566 ], "lines": [ { "bbox": [ 106, 489, 505, 501 ], "spans": [ { "bbox": [ 106, 489, 505, 501 ], "score": 1.0, "content": "Attention maps of vision transformers. Visualising the attention map from [CLS] token to patch", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 500, 505, 513 ], "spans": [ { "bbox": [ 105, 500, 505, 513 ], "score": 1.0, "content": "tokens was popularized in DINO (Caron et al., 2021). It was shown there that the attention maps", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 511, 505, 523 ], "spans": [ { "bbox": [ 105, 511, 505, 523 ], "score": 1.0, "content": "of DINO were clean of artifacts, as opposed to the attention maps of previous vision transformers.", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 521, 506, 536 ], "spans": [ { "bbox": [ 105, 521, 506, 536 ], "score": 1.0, "content": "Other works have since reported interesting attention maps using various techniques: by modifying", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 532, 506, 546 ], "spans": [ { "bbox": [ 105, 532, 506, 546 ], "score": 1.0, "content": "the optimisation procedure (Chen et al., 2022), by steering the attention scores towards useful image", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 544, 506, 557 ], "spans": [ { "bbox": [ 106, 544, 506, 557 ], "score": 1.0, "content": "parts (Shi et al., 2023), by modifying the architecture of the transformer layers (Yu et al., 2024), or", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 555, 451, 567 ], "spans": [ { "bbox": [ 105, 555, 451, 567 ], "score": 1.0, "content": "by introducing a learnable pooling to produce the [CLS] token (Psomas et al., 2023).", "type": "text" } ], "index": 42 } ], "index": 39 }, { "type": "title", "bbox": [ 107, 591, 195, 604 ], "lines": [ { "bbox": [ 104, 590, 198, 607 ], "spans": [ { "bbox": [ 104, 590, 198, 607 ], "score": 1.0, "content": "5 CONCLUSION", "type": "text" } ], "index": 43 } ], "index": 43 }, { "type": "text", "bbox": [ 107, 622, 504, 732 ], "lines": [ { "bbox": [ 105, 622, 505, 635 ], "spans": [ { "bbox": [ 105, 622, 505, 635 ], "score": 1.0, "content": "In this work, we exposed artifacts in the feature maps of DINOv2 models, and found this phe-", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 633, 506, 646 ], "spans": [ { "bbox": [ 105, 633, 506, 646 ], "score": 1.0, "content": "nomenon to be present in multiple existing popular models. We have described a simple method to", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 644, 506, 657 ], "spans": [ { "bbox": [ 105, 644, 506, 657 ], "score": 1.0, "content": "detect these artifacts by observing that they correspond to tokens with an outlier norm value at the", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 655, 506, 668 ], "spans": [ { "bbox": [ 105, 655, 506, 668 ], "score": 1.0, "content": "output of the Transformer model. Studying their location, we have proposed an interpretation that", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 666, 506, 679 ], "spans": [ { "bbox": [ 105, 666, 506, 679 ], "score": 1.0, "content": "models naturally recycle tokens from low-informative areas and repurpose them into a different role", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 676, 506, 690 ], "spans": [ { "bbox": [ 105, 676, 506, 690 ], "score": 1.0, "content": "for inference. Following this interpretation, we have proposed a simple fix, consisting of appending", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 688, 505, 701 ], "spans": [ { "bbox": [ 105, 688, 505, 701 ], "score": 1.0, "content": "additional tokens to the input sequence that are not used as outputs, and have found that this entirely", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 699, 506, 711 ], "spans": [ { "bbox": [ 105, 699, 506, 711 ], "score": 1.0, "content": "removes the artifacts, improving the performance in dense prediction and object discovery. More-", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 710, 506, 723 ], "spans": [ { "bbox": [ 105, 710, 506, 723 ], "score": 1.0, "content": "over, we have shown that the proposed solution also removes the same artifacts present in supervised", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 721, 435, 734 ], "spans": [ { "bbox": [ 105, 721, 435, 734 ], "score": 1.0, "content": "models such as DeiT-III and OpenCLIP, confirming the generality of our solution.", "type": "text" } ], "index": 53 } ], "index": 48.5 } ], "page_idx": 8, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 302, 752, 308, 760 ], "lines": [ { "bbox": [ 302, 751, 309, 762 ], "spans": [ { "bbox": [ 302, 751, 309, 762 ], "score": 1.0, "content": "9", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 107, 82, 504, 247 ], "lines": [], "index": 7, "bbox_fs": [ 105, 83, 506, 248 ], "lines_deleted": true }, { "type": "text", "bbox": [ 107, 253, 504, 483 ], "lines": [ { "bbox": [ 105, 253, 506, 266 ], "spans": [ { "bbox": [ 105, 253, 506, 266 ], "score": 1.0, "content": "Additional tokens in transformers. Extending the transformer sequence with special tokens was", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 263, 506, 278 ], "spans": [ { "bbox": [ 105, 263, 506, 278 ], "score": 1.0, "content": "popularized in BERT (Devlin et al., 2019). However, most approaches add new tokens either to pro-", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 274, 505, 288 ], "spans": [ { "bbox": [ 105, 274, 505, 288 ], "score": 1.0, "content": "vide the network with new information as for example [SEP] tokens in BERT, provide opportunity", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 285, 506, 299 ], "spans": [ { "bbox": [ 105, 285, 506, 299 ], "score": 1.0, "content": "to spend more computation on the input as seen with the tape tokens in AdaTape (Xue et al., 2023),", "type": "text" } ], "index": 18 }, { "bbox": [ 105, 297, 506, 309 ], "spans": [ { "bbox": [ 105, 297, 506, 309 ], "score": 1.0, "content": "or to gather information in these tokens, and use their output value as an output of the model: for", "type": "text" } ], "index": 19 }, { "bbox": [ 104, 306, 506, 322 ], "spans": [ { "bbox": [ 104, 306, 506, 322 ], "score": 1.0, "content": "classification, as [CLS] tokens in BERT and ViT (Dosovitskiy et al., 2021); for generative learning,", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 318, 506, 331 ], "spans": [ { "bbox": [ 105, 318, 506, 331 ], "score": 1.0, "content": "as [MASK] in BERT and BEiT (Bao et al., 2021); for detection, as object queries in DETR (Carion", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 330, 506, 343 ], "spans": [ { "bbox": [ 105, 330, 506, 343 ], "score": 1.0, "content": "et al., 2020), detection tokens in YOLOS (Fang et al., 2021), and ViDT (Song et al., 2021); or for", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 340, 505, 354 ], "spans": [ { "bbox": [ 105, 340, 505, 354 ], "score": 1.0, "content": "accumulating information from possibly multiple modalities before decoding, as latent token arrays", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 351, 505, 365 ], "spans": [ { "bbox": [ 105, 351, 505, 365 ], "score": 1.0, "content": "in Perceivers (Jaegle et al., 2021; 2022). Different to these works, the tokens we add to the sequence", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 362, 505, 376 ], "spans": [ { "bbox": [ 105, 362, 505, 376 ], "score": 1.0, "content": "add no information, and their output value is not used for any purpose. They are simply registers", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 374, 505, 387 ], "spans": [ { "bbox": [ 105, 374, 505, 387 ], "score": 1.0, "content": "where the model can learn to store and retrieve information during the forward pass. The Mem-", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 384, 506, 398 ], "spans": [ { "bbox": [ 105, 384, 506, 398 ], "score": 1.0, "content": "ory Transformer (Burtsev et al., 2020), closer to our work, presents a simple approach to improve", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 395, 505, 408 ], "spans": [ { "bbox": [ 105, 395, 505, 408 ], "score": 1.0, "content": "transformer models using memory tokens added to the token sequence, improving translation perfor-", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 406, 505, 419 ], "spans": [ { "bbox": [ 105, 406, 505, 419 ], "score": 1.0, "content": "mance. In follow-up work, Bulatov et al. (2022) address complex copy-repeat-reverse tasks. Sandler", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 417, 506, 430 ], "spans": [ { "bbox": [ 105, 417, 506, 430 ], "score": 1.0, "content": "et al. (2022) extend this line to the vision domain for fine-tuning but observe that such tokens do not", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 429, 506, 442 ], "spans": [ { "bbox": [ 105, 429, 506, 442 ], "score": 1.0, "content": "transfer well across tasks. In contrast, we do not perform fine-tuning and employ additional tokens", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 439, 506, 452 ], "spans": [ { "bbox": [ 105, 439, 506, 452 ], "score": 1.0, "content": "during pretraining to improve the features obtained for all tasks downstream. More importantly, our", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 449, 506, 464 ], "spans": [ { "bbox": [ 105, 449, 506, 464 ], "score": 1.0, "content": "study contributes the following new insight in Sec. 2: the mechanism implemented through memory", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 461, 506, 474 ], "spans": [ { "bbox": [ 105, 461, 506, 474 ], "score": 1.0, "content": "tokens already appears naturally in Vision Transformers; our study shows that such tokens allow us", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 472, 455, 485 ], "spans": [ { "bbox": [ 105, 472, 455, 485 ], "score": 1.0, "content": "not to create but to isolate this existing behavior, and thus avoid collateral side-effects.", "type": "text" } ], "index": 35 } ], "index": 25, "bbox_fs": [ 104, 253, 506, 485 ] }, { "type": "text", "bbox": [ 107, 489, 504, 566 ], "lines": [ { "bbox": [ 106, 489, 505, 501 ], "spans": [ { "bbox": [ 106, 489, 505, 501 ], "score": 1.0, "content": "Attention maps of vision transformers. Visualising the attention map from [CLS] token to patch", "type": "text" } ], "index": 36 }, { "bbox": [ 105, 500, 505, 513 ], "spans": [ { "bbox": [ 105, 500, 505, 513 ], "score": 1.0, "content": "tokens was popularized in DINO (Caron et al., 2021). It was shown there that the attention maps", "type": "text" } ], "index": 37 }, { "bbox": [ 105, 511, 505, 523 ], "spans": [ { "bbox": [ 105, 511, 505, 523 ], "score": 1.0, "content": "of DINO were clean of artifacts, as opposed to the attention maps of previous vision transformers.", "type": "text" } ], "index": 38 }, { "bbox": [ 105, 521, 506, 536 ], "spans": [ { "bbox": [ 105, 521, 506, 536 ], "score": 1.0, "content": "Other works have since reported interesting attention maps using various techniques: by modifying", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 532, 506, 546 ], "spans": [ { "bbox": [ 105, 532, 506, 546 ], "score": 1.0, "content": "the optimisation procedure (Chen et al., 2022), by steering the attention scores towards useful image", "type": "text" } ], "index": 40 }, { "bbox": [ 106, 544, 506, 557 ], "spans": [ { "bbox": [ 106, 544, 506, 557 ], "score": 1.0, "content": "parts (Shi et al., 2023), by modifying the architecture of the transformer layers (Yu et al., 2024), or", "type": "text" } ], "index": 41 }, { "bbox": [ 105, 555, 451, 567 ], "spans": [ { "bbox": [ 105, 555, 451, 567 ], "score": 1.0, "content": "by introducing a learnable pooling to produce the [CLS] token (Psomas et al., 2023).", "type": "text" } ], "index": 42 } ], "index": 39, "bbox_fs": [ 105, 489, 506, 567 ] }, { "type": "title", "bbox": [ 107, 591, 195, 604 ], "lines": [ { "bbox": [ 104, 590, 198, 607 ], "spans": [ { "bbox": [ 104, 590, 198, 607 ], "score": 1.0, "content": "5 CONCLUSION", "type": "text" } ], "index": 43 } ], "index": 43 }, { "type": "text", "bbox": [ 107, 622, 504, 732 ], "lines": [ { "bbox": [ 105, 622, 505, 635 ], "spans": [ { "bbox": [ 105, 622, 505, 635 ], "score": 1.0, "content": "In this work, we exposed artifacts in the feature maps of DINOv2 models, and found this phe-", "type": "text" } ], "index": 44 }, { "bbox": [ 105, 633, 506, 646 ], "spans": [ { "bbox": [ 105, 633, 506, 646 ], "score": 1.0, "content": "nomenon to be present in multiple existing popular models. We have described a simple method to", "type": "text" } ], "index": 45 }, { "bbox": [ 105, 644, 506, 657 ], "spans": [ { "bbox": [ 105, 644, 506, 657 ], "score": 1.0, "content": "detect these artifacts by observing that they correspond to tokens with an outlier norm value at the", "type": "text" } ], "index": 46 }, { "bbox": [ 105, 655, 506, 668 ], "spans": [ { "bbox": [ 105, 655, 506, 668 ], "score": 1.0, "content": "output of the Transformer model. Studying their location, we have proposed an interpretation that", "type": "text" } ], "index": 47 }, { "bbox": [ 105, 666, 506, 679 ], "spans": [ { "bbox": [ 105, 666, 506, 679 ], "score": 1.0, "content": "models naturally recycle tokens from low-informative areas and repurpose them into a different role", "type": "text" } ], "index": 48 }, { "bbox": [ 105, 676, 506, 690 ], "spans": [ { "bbox": [ 105, 676, 506, 690 ], "score": 1.0, "content": "for inference. Following this interpretation, we have proposed a simple fix, consisting of appending", "type": "text" } ], "index": 49 }, { "bbox": [ 105, 688, 505, 701 ], "spans": [ { "bbox": [ 105, 688, 505, 701 ], "score": 1.0, "content": "additional tokens to the input sequence that are not used as outputs, and have found that this entirely", "type": "text" } ], "index": 50 }, { "bbox": [ 105, 699, 506, 711 ], "spans": [ { "bbox": [ 105, 699, 506, 711 ], "score": 1.0, "content": "removes the artifacts, improving the performance in dense prediction and object discovery. More-", "type": "text" } ], "index": 51 }, { "bbox": [ 105, 710, 506, 723 ], "spans": [ { "bbox": [ 105, 710, 506, 723 ], "score": 1.0, "content": "over, we have shown that the proposed solution also removes the same artifacts present in supervised", "type": "text" } ], "index": 52 }, { "bbox": [ 105, 721, 435, 734 ], "spans": [ { "bbox": [ 105, 721, 435, 734 ], "score": 1.0, "content": "models such as DeiT-III and OpenCLIP, confirming the generality of our solution.", "type": "text" } ], "index": 53 } ], "index": 48.5, "bbox_fs": [ 105, 622, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 83, 200, 93 ], "lines": [ { "bbox": [ 107, 83, 200, 94 ], "spans": [ { "bbox": [ 107, 83, 200, 94 ], "score": 1.0, "content": "ACKNOWLEDGMENTS", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 108, 101, 504, 146 ], "lines": [ { "bbox": [ 106, 101, 505, 114 ], "spans": [ { "bbox": [ 106, 101, 146, 114 ], "score": 1.0, "content": "We thank", "type": "text" }, { "bbox": [ 147, 101, 175, 112 ], "score": 0.31, "content": "\\mathrm { H u } \\ X \\mathbf { u }", "type": "inline_equation" }, { "bbox": [ 176, 101, 505, 114 ], "score": 1.0, "content": ", Oriane Simeoni, Mido Assran and Armand Joulin for their insightful discussions ´", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 113, 506, 125 ], "spans": [ { "bbox": [ 106, 113, 506, 125 ], "score": 1.0, "content": "and help during the course of this work. We thank Pyrrhus for posing for fig 8. Julien Mairal was", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 124, 506, 136 ], "spans": [ { "bbox": [ 105, 124, 506, 136 ], "score": 1.0, "content": "supported by ANR 3IA MIAI@Grenoble Alpes (ANR-19-P3IA-0003) and by ERC grant number", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 134, 244, 147 ], "spans": [ { "bbox": [ 106, 134, 244, 147 ], "score": 1.0, "content": "101087696 (APHELEIA project).", "type": "text" } ], "index": 4 } ], "index": 2.5 }, { "type": "title", "bbox": [ 107, 163, 176, 174 ], "lines": [ { "bbox": [ 106, 163, 176, 176 ], "spans": [ { "bbox": [ 106, 163, 176, 176 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "text", "bbox": [ 104, 178, 506, 739 ], "lines": [ { "bbox": [ 105, 181, 506, 195 ], "spans": [ { "bbox": [ 105, 181, 506, 195 ], "score": 1.0, "content": "Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2021.", "type": "text" } ], "index": 6 }, { "bbox": [ 104, 198, 507, 214 ], "spans": [ { "bbox": [ 104, 198, 507, 214 ], "score": 1.0, "content": "Aydar Bulatov, Yury Kuratov, and Mikhail Burtsev. Recurrent memory transformer. In NeurIPS,", "type": "text" } ], "index": 7 }, { "bbox": [ 115, 211, 142, 224 ], "spans": [ { "bbox": [ 115, 211, 142, 224 ], "score": 1.0, "content": "2022.", "type": "text" } ], "index": 8 }, { "bbox": [ 104, 227, 505, 244 ], "spans": [ { "bbox": [ 104, 227, 505, 244 ], "score": 1.0, "content": "Mikhail S Burtsev, Yuri Kuratov, Anton Peganov, and Grigory V Sapunov. Memory transformer.", "type": "text" } ], "index": 9 }, { "bbox": [ 115, 240, 279, 254 ], "spans": [ { "bbox": [ 115, 240, 279, 254 ], "score": 1.0, "content": "arXiv preprint arXiv:2006.11527, 2020.", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 259, 506, 272 ], "spans": [ { "bbox": [ 105, 259, 506, 272 ], "score": 1.0, "content": "Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and", "type": "text" } ], "index": 11 }, { "bbox": [ 114, 270, 452, 283 ], "spans": [ { "bbox": [ 114, 270, 452, 283 ], "score": 1.0, "content": "Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020.", "type": "text" } ], "index": 12 }, { "bbox": [ 105, 288, 506, 302 ], "spans": [ { "bbox": [ 105, 288, 506, 302 ], "score": 1.0, "content": "Mathilde Caron, Hugo Touvron, Ishan Misra, Herve J ´ egou, Julien Mairal, Piotr Bojanowski, and ´", "type": "text" } ], "index": 13 }, { "bbox": [ 114, 299, 487, 315 ], "spans": [ { "bbox": [ 114, 299, 487, 315 ], "score": 1.0, "content": "Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021.", "type": "text" } ], "index": 14 }, { "bbox": [ 104, 316, 507, 333 ], "spans": [ { "bbox": [ 104, 316, 507, 333 ], "score": 1.0, "content": "Xiangning Chen, Cho-Jui Hsieh, and Boqing Gong. When vision transformers outperform resnets", "type": "text" } ], "index": 15 }, { "bbox": [ 114, 328, 383, 343 ], "spans": [ { "bbox": [ 114, 328, 383, 343 ], "score": 1.0, "content": "without pre-training or strong data augmentations. In ICLR, 2022.", "type": "text" } ], "index": 16 }, { "bbox": [ 104, 347, 505, 362 ], "spans": [ { "bbox": [ 104, 347, 505, 362 ], "score": 1.0, "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 359, 398, 372 ], "spans": [ { "bbox": [ 115, 359, 398, 372 ], "score": 1.0, "content": "bidirectional transformers for language understanding. NAACL, 2019.", "type": "text" } ], "index": 18 }, { "bbox": [ 104, 375, 505, 392 ], "spans": [ { "bbox": [ 104, 375, 505, 392 ], "score": 1.0, "content": "Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by", "type": "text" } ], "index": 19 }, { "bbox": [ 115, 388, 259, 402 ], "spans": [ { "bbox": [ 115, 388, 259, 402 ], "score": 1.0, "content": "context prediction. In ICCV, 2015.", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 407, 505, 420 ], "spans": [ { "bbox": [ 105, 407, 505, 420 ], "score": 1.0, "content": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas", "type": "text" } ], "index": 21 }, { "bbox": [ 114, 416, 507, 433 ], "spans": [ { "bbox": [ 114, 416, 507, 433 ], "score": 1.0, "content": "Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 429, 478, 443 ], "spans": [ { "bbox": [ 115, 429, 478, 443 ], "score": 1.0, "content": "image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021.", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 446, 506, 462 ], "spans": [ { "bbox": [ 105, 446, 506, 462 ], "score": 1.0, "content": "Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and", "type": "text" } ], "index": 24 }, { "bbox": [ 114, 457, 506, 473 ], "spans": [ { "bbox": [ 114, 457, 506, 473 ], "score": 1.0, "content": "Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 470, 234, 482 ], "spans": [ { "bbox": [ 115, 470, 234, 482 ], "score": 1.0, "content": "detection. In NeurIPS, 2021.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 488, 506, 502 ], "spans": [ { "bbox": [ 105, 488, 506, 502 ], "score": 1.0, "content": "Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 500, 361, 512 ], "spans": [ { "bbox": [ 115, 500, 361, 512 ], "score": 1.0, "content": "unsupervised visual representation learning. In CVPR, 2020.", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 517, 506, 531 ], "spans": [ { "bbox": [ 105, 517, 506, 531 ], "score": 1.0, "content": "Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked ´", "type": "text" } ], "index": 29 }, { "bbox": [ 115, 528, 351, 542 ], "spans": [ { "bbox": [ 115, 528, 351, 542 ], "score": 1.0, "content": "autoencoders are scalable vision learners. In CVPR, 2022.", "type": "text" } ], "index": 30 }, { "bbox": [ 106, 548, 505, 560 ], "spans": [ { "bbox": [ 106, 548, 505, 560 ], "score": 1.0, "content": "Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori,", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 558, 505, 572 ], "spans": [ { "bbox": [ 115, 558, 505, 572 ], "score": 1.0, "content": "Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 570, 308, 582 ], "spans": [ { "bbox": [ 115, 570, 308, 582 ], "score": 1.0, "content": "Farhadi, and Ludwig Schmidt. Openclip. 2021.", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 588, 505, 602 ], "spans": [ { "bbox": [ 105, 588, 505, 602 ], "score": 1.0, "content": "Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira.", "type": "text" } ], "index": 34 }, { "bbox": [ 115, 599, 399, 613 ], "spans": [ { "bbox": [ 115, 599, 399, 613 ], "score": 1.0, "content": "Perceiver: General perception with iterative attention. In ICML, 2021.", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 617, 506, 631 ], "spans": [ { "bbox": [ 105, 617, 506, 631 ], "score": 1.0, "content": "Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David", "type": "text" } ], "index": 36 }, { "bbox": [ 114, 628, 506, 642 ], "spans": [ { "bbox": [ 114, 628, 506, 642 ], "score": 1.0, "content": "Ding, Skanda Koppula, Andrew Brock, Evan Shelhamer, Olivier J. H’enaff, Matthew M.", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 640, 505, 652 ], "spans": [ { "bbox": [ 115, 640, 505, 652 ], "score": 1.0, "content": "Botvinick, Andrew Zisserman, Oriol Vinyals, and Joao Carreira. Perceiver io: A general ar- ˜", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 650, 350, 664 ], "spans": [ { "bbox": [ 115, 650, 350, 664 ], "score": 1.0, "content": "chitecture for structured inputs & outputs. In ICLR, 2022.", "type": "text" } ], "index": 39 }, { "bbox": [ 105, 669, 506, 682 ], "spans": [ { "bbox": [ 105, 669, 506, 682 ], "score": 1.0, "content": "Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete", "type": "text" } ], "index": 40 }, { "bbox": [ 115, 680, 505, 694 ], "spans": [ { "bbox": [ 115, 680, 505, 694 ], "score": 1.0, "content": "Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv", "type": "text" } ], "index": 41 }, { "bbox": [ 114, 691, 254, 705 ], "spans": [ { "bbox": [ 114, 691, 254, 705 ], "score": 1.0, "content": "preprint arXiv:2304.02643, 2023.", "type": "text" } ], "index": 42 }, { "bbox": [ 105, 709, 505, 724 ], "spans": [ { "bbox": [ 105, 709, 505, 724 ], "score": 1.0, "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", "type": "text" } ], "index": 43 }, { "bbox": [ 115, 721, 295, 734 ], "spans": [ { "bbox": [ 115, 721, 295, 734 ], "score": 1.0, "content": "lutional neural networks. In NeurIPS, 2012.", "type": "text" } ], "index": 44 } ], "index": 25 } ], "page_idx": 9, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 752, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 14 } ] } ] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 83, 200, 93 ], "lines": [ { "bbox": [ 107, 83, 200, 94 ], "spans": [ { "bbox": [ 107, 83, 200, 94 ], "score": 1.0, "content": "ACKNOWLEDGMENTS", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 108, 101, 504, 146 ], "lines": [ { "bbox": [ 106, 101, 505, 114 ], "spans": [ { "bbox": [ 106, 101, 146, 114 ], "score": 1.0, "content": "We thank", "type": "text" }, { "bbox": [ 147, 101, 175, 112 ], "score": 0.31, "content": "\\mathrm { H u } \\ X \\mathbf { u }", "type": "inline_equation" }, { "bbox": [ 176, 101, 505, 114 ], "score": 1.0, "content": ", Oriane Simeoni, Mido Assran and Armand Joulin for their insightful discussions ´", "type": "text" } ], "index": 1 }, { "bbox": [ 106, 113, 506, 125 ], "spans": [ { "bbox": [ 106, 113, 506, 125 ], "score": 1.0, "content": "and help during the course of this work. We thank Pyrrhus for posing for fig 8. Julien Mairal was", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 124, 506, 136 ], "spans": [ { "bbox": [ 105, 124, 506, 136 ], "score": 1.0, "content": "supported by ANR 3IA MIAI@Grenoble Alpes (ANR-19-P3IA-0003) and by ERC grant number", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 134, 244, 147 ], "spans": [ { "bbox": [ 106, 134, 244, 147 ], "score": 1.0, "content": "101087696 (APHELEIA project).", "type": "text" } ], "index": 4 } ], "index": 2.5, "bbox_fs": [ 105, 101, 506, 147 ] }, { "type": "title", "bbox": [ 107, 163, 176, 174 ], "lines": [ { "bbox": [ 106, 163, 176, 176 ], "spans": [ { "bbox": [ 106, 163, 176, 176 ], "score": 1.0, "content": "REFERENCES", "type": "text" } ], "index": 5 } ], "index": 5 }, { "type": "list", "bbox": [ 104, 178, 506, 739 ], "lines": [ { "bbox": [ 105, 181, 506, 195 ], "spans": [ { "bbox": [ 105, 181, 506, 195 ], "score": 1.0, "content": "Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2021.", "type": "text" } ], "index": 6, "is_list_start_line": true }, { "bbox": [ 104, 198, 507, 214 ], "spans": [ { "bbox": [ 104, 198, 507, 214 ], "score": 1.0, "content": "Aydar Bulatov, Yury Kuratov, and Mikhail Burtsev. Recurrent memory transformer. In NeurIPS,", "type": "text" } ], "index": 7, "is_list_start_line": true }, { "bbox": [ 115, 211, 142, 224 ], "spans": [ { "bbox": [ 115, 211, 142, 224 ], "score": 1.0, "content": "2022.", "type": "text" } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 104, 227, 505, 244 ], "spans": [ { "bbox": [ 104, 227, 505, 244 ], "score": 1.0, "content": "Mikhail S Burtsev, Yuri Kuratov, Anton Peganov, and Grigory V Sapunov. Memory transformer.", "type": "text" } ], "index": 9, "is_list_start_line": true }, { "bbox": [ 115, 240, 279, 254 ], "spans": [ { "bbox": [ 115, 240, 279, 254 ], "score": 1.0, "content": "arXiv preprint arXiv:2006.11527, 2020.", "type": "text" } ], "index": 10, "is_list_end_line": true }, { "bbox": [ 105, 259, 506, 272 ], "spans": [ { "bbox": [ 105, 259, 506, 272 ], "score": 1.0, "content": "Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and", "type": "text" } ], "index": 11, "is_list_start_line": true }, { "bbox": [ 114, 270, 452, 283 ], "spans": [ { "bbox": [ 114, 270, 452, 283 ], "score": 1.0, "content": "Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020.", "type": "text" } ], "index": 12, "is_list_end_line": true }, { "bbox": [ 105, 288, 506, 302 ], "spans": [ { "bbox": [ 105, 288, 506, 302 ], "score": 1.0, "content": "Mathilde Caron, Hugo Touvron, Ishan Misra, Herve J ´ egou, Julien Mairal, Piotr Bojanowski, and ´", "type": "text" } ], "index": 13, "is_list_start_line": true }, { "bbox": [ 114, 299, 487, 315 ], "spans": [ { "bbox": [ 114, 299, 487, 315 ], "score": 1.0, "content": "Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021.", "type": "text" } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 104, 316, 507, 333 ], "spans": [ { "bbox": [ 104, 316, 507, 333 ], "score": 1.0, "content": "Xiangning Chen, Cho-Jui Hsieh, and Boqing Gong. When vision transformers outperform resnets", "type": "text" } ], "index": 15, "is_list_start_line": true }, { "bbox": [ 114, 328, 383, 343 ], "spans": [ { "bbox": [ 114, 328, 383, 343 ], "score": 1.0, "content": "without pre-training or strong data augmentations. In ICLR, 2022.", "type": "text" } ], "index": 16, "is_list_end_line": true }, { "bbox": [ 104, 347, 505, 362 ], "spans": [ { "bbox": [ 104, 347, 505, 362 ], "score": 1.0, "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep", "type": "text" } ], "index": 17, "is_list_start_line": true }, { "bbox": [ 115, 359, 398, 372 ], "spans": [ { "bbox": [ 115, 359, 398, 372 ], "score": 1.0, "content": "bidirectional transformers for language understanding. NAACL, 2019.", "type": "text" } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 104, 375, 505, 392 ], "spans": [ { "bbox": [ 104, 375, 505, 392 ], "score": 1.0, "content": "Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by", "type": "text" } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 115, 388, 259, 402 ], "spans": [ { "bbox": [ 115, 388, 259, 402 ], "score": 1.0, "content": "context prediction. In ICCV, 2015.", "type": "text" } ], "index": 20, "is_list_end_line": true }, { "bbox": [ 105, 407, 505, 420 ], "spans": [ { "bbox": [ 105, 407, 505, 420 ], "score": 1.0, "content": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas", "type": "text" } ], "index": 21, "is_list_start_line": true }, { "bbox": [ 114, 416, 507, 433 ], "spans": [ { "bbox": [ 114, 416, 507, 433 ], "score": 1.0, "content": "Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 429, 478, 443 ], "spans": [ { "bbox": [ 115, 429, 478, 443 ], "score": 1.0, "content": "image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021.", "type": "text" } ], "index": 23, "is_list_end_line": true }, { "bbox": [ 105, 446, 506, 462 ], "spans": [ { "bbox": [ 105, 446, 506, 462 ], "score": 1.0, "content": "Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, and", "type": "text" } ], "index": 24, "is_list_start_line": true }, { "bbox": [ 114, 457, 506, 473 ], "spans": [ { "bbox": [ 114, 457, 506, 473 ], "score": 1.0, "content": "Wenyu Liu. You only look at one sequence: Rethinking transformer in vision through object", "type": "text" } ], "index": 25 }, { "bbox": [ 115, 470, 234, 482 ], "spans": [ { "bbox": [ 115, 470, 234, 482 ], "score": 1.0, "content": "detection. In NeurIPS, 2021.", "type": "text" } ], "index": 26, "is_list_end_line": true }, { "bbox": [ 105, 488, 506, 502 ], "spans": [ { "bbox": [ 105, 488, 506, 502 ], "score": 1.0, "content": "Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for", "type": "text" } ], "index": 27, "is_list_start_line": true }, { "bbox": [ 115, 500, 361, 512 ], "spans": [ { "bbox": [ 115, 500, 361, 512 ], "score": 1.0, "content": "unsupervised visual representation learning. In CVPR, 2020.", "type": "text" } ], "index": 28, "is_list_end_line": true }, { "bbox": [ 105, 517, 506, 531 ], "spans": [ { "bbox": [ 105, 517, 506, 531 ], "score": 1.0, "content": "Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked ´", "type": "text" } ], "index": 29, "is_list_start_line": true }, { "bbox": [ 115, 528, 351, 542 ], "spans": [ { "bbox": [ 115, 528, 351, 542 ], "score": 1.0, "content": "autoencoders are scalable vision learners. In CVPR, 2022.", "type": "text" } ], "index": 30, "is_list_end_line": true }, { "bbox": [ 106, 548, 505, 560 ], "spans": [ { "bbox": [ 106, 548, 505, 560 ], "score": 1.0, "content": "Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori,", "type": "text" } ], "index": 31, "is_list_start_line": true }, { "bbox": [ 115, 558, 505, 572 ], "spans": [ { "bbox": [ 115, 558, 505, 572 ], "score": 1.0, "content": "Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali", "type": "text" } ], "index": 32 }, { "bbox": [ 115, 570, 308, 582 ], "spans": [ { "bbox": [ 115, 570, 308, 582 ], "score": 1.0, "content": "Farhadi, and Ludwig Schmidt. Openclip. 2021.", "type": "text" } ], "index": 33, "is_list_end_line": true }, { "bbox": [ 105, 588, 505, 602 ], "spans": [ { "bbox": [ 105, 588, 505, 602 ], "score": 1.0, "content": "Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira.", "type": "text" } ], "index": 34, "is_list_start_line": true }, { "bbox": [ 115, 599, 399, 613 ], "spans": [ { "bbox": [ 115, 599, 399, 613 ], "score": 1.0, "content": "Perceiver: General perception with iterative attention. In ICML, 2021.", "type": "text" } ], "index": 35, "is_list_end_line": true }, { "bbox": [ 105, 617, 506, 631 ], "spans": [ { "bbox": [ 105, 617, 506, 631 ], "score": 1.0, "content": "Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David", "type": "text" } ], "index": 36, "is_list_start_line": true }, { "bbox": [ 114, 628, 506, 642 ], "spans": [ { "bbox": [ 114, 628, 506, 642 ], "score": 1.0, "content": "Ding, Skanda Koppula, Andrew Brock, Evan Shelhamer, Olivier J. H’enaff, Matthew M.", "type": "text" } ], "index": 37 }, { "bbox": [ 115, 640, 505, 652 ], "spans": [ { "bbox": [ 115, 640, 505, 652 ], "score": 1.0, "content": "Botvinick, Andrew Zisserman, Oriol Vinyals, and Joao Carreira. Perceiver io: A general ar- ˜", "type": "text" } ], "index": 38 }, { "bbox": [ 115, 650, 350, 664 ], "spans": [ { "bbox": [ 115, 650, 350, 664 ], "score": 1.0, "content": "chitecture for structured inputs & outputs. In ICLR, 2022.", "type": "text" } ], "index": 39, "is_list_end_line": true }, { "bbox": [ 105, 669, 506, 682 ], "spans": [ { "bbox": [ 105, 669, 506, 682 ], "score": 1.0, "content": "Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete", "type": "text" } ], "index": 40, "is_list_start_line": true }, { "bbox": [ 115, 680, 505, 694 ], "spans": [ { "bbox": [ 115, 680, 505, 694 ], "score": 1.0, "content": "Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv", "type": "text" } ], "index": 41 }, { "bbox": [ 114, 691, 254, 705 ], "spans": [ { "bbox": [ 114, 691, 254, 705 ], "score": 1.0, "content": "preprint arXiv:2304.02643, 2023.", "type": "text" } ], "index": 42, "is_list_end_line": true }, { "bbox": [ 105, 709, 505, 724 ], "spans": [ { "bbox": [ 105, 709, 505, 724 ], "score": 1.0, "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", "type": "text" } ], "index": 43, "is_list_start_line": true }, { "bbox": [ 115, 721, 295, 734 ], "spans": [ { "bbox": [ 115, 721, 295, 734 ], "score": 1.0, "content": "lutional neural networks. In NeurIPS, 2012.", "type": "text" } ], "index": 44, "is_list_end_line": true }, { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold,", "type": "text", "cross_page": true } ], "index": 0, "is_list_start_line": true }, { "bbox": [ 115, 94, 505, 107 ], "spans": [ { "bbox": [ 115, 94, 505, 107 ], "score": 1.0, "content": "Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot atten-", "type": "text", "cross_page": true } ], "index": 1 }, { "bbox": [ 115, 104, 213, 118 ], "spans": [ { "bbox": [ 115, 104, 213, 118 ], "score": 1.0, "content": "tion. In NeurIPS, 2020.", "type": "text", "cross_page": true } ], "index": 2, "is_list_end_line": true }, { "bbox": [ 106, 124, 455, 136 ], "spans": [ { "bbox": [ 106, 124, 455, 136 ], "score": 1.0, "content": "David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2004.", "type": "text", "cross_page": true } ], "index": 3, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 142, 505, 155 ], "spans": [ { "bbox": [ 105, 142, 505, 155 ], "score": 1.0, "content": "Maxime Oquab, Timothee Darcet, Th ´ eo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, ´", "type": "text", "cross_page": true } ], "index": 4, "is_list_start_line": true }, { "bbox": [ 114, 151, 506, 169 ], "spans": [ { "bbox": [ 114, 151, 506, 169 ], "score": 1.0, "content": "Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning", "type": "text", "cross_page": true } ], "index": 5 }, { "bbox": [ 115, 164, 451, 178 ], "spans": [ { "bbox": [ 115, 164, 451, 178 ], "score": 1.0, "content": "robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023.", "type": "text", "cross_page": true } ], "index": 6, "is_list_end_line": true }, { "bbox": [ 106, 183, 505, 197 ], "spans": [ { "bbox": [ 106, 183, 505, 197 ], "score": 1.0, "content": "Bill Psomas, Ioannis Kakogeorgiou, Konstantinos Karantzalos, and Yannis Avrithis. Keep it sim-", "type": "text", "cross_page": true } ], "index": 7, "is_list_start_line": true }, { "bbox": [ 114, 193, 461, 209 ], "spans": [ { "bbox": [ 114, 193, 461, 209 ], "score": 1.0, "content": "pool: Who said supervised transformers suffer from attention deficit? In ICCV, 2023.", "type": "text", "cross_page": true } ], "index": 8, "is_list_end_line": true }, { "bbox": [ 105, 212, 505, 227 ], "spans": [ { "bbox": [ 105, 212, 505, 227 ], "score": 1.0, "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,", "type": "text", "cross_page": true } ], "index": 9, "is_list_start_line": true }, { "bbox": [ 114, 223, 506, 239 ], "spans": [ { "bbox": [ 114, 223, 506, 239 ], "score": 1.0, "content": "Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual", "type": "text", "cross_page": true } ], "index": 10 }, { "bbox": [ 115, 236, 355, 249 ], "spans": [ { "bbox": [ 115, 236, 355, 249 ], "score": 1.0, "content": "models from natural language supervision. In ICML, 2021.", "type": "text", "cross_page": true } ], "index": 11, "is_list_end_line": true }, { "bbox": [ 104, 253, 506, 269 ], "spans": [ { "bbox": [ 104, 253, 506, 269 ], "score": 1.0, "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng", "type": "text", "cross_page": true } ], "index": 12, "is_list_start_line": true }, { "bbox": [ 114, 264, 506, 279 ], "spans": [ { "bbox": [ 114, 264, 506, 279 ], "score": 1.0, "content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C Berg, and Li Fei-Fei.", "type": "text", "cross_page": true } ], "index": 13 }, { "bbox": [ 115, 275, 370, 291 ], "spans": [ { "bbox": [ 115, 275, 370, 291 ], "score": 1.0, "content": "Imagenet large scale visual recognition challenge. IJCV, 2015.", "type": "text", "cross_page": true } ], "index": 14, "is_list_end_line": true }, { "bbox": [ 104, 293, 506, 310 ], "spans": [ { "bbox": [ 104, 293, 506, 310 ], "score": 1.0, "content": "Mark Sandler, Andrey Zhmoginov, Max Vladymyrov, and Andrew Jackson. Fine-tuning image", "type": "text", "cross_page": true } ], "index": 15, "is_list_start_line": true }, { "bbox": [ 114, 305, 338, 320 ], "spans": [ { "bbox": [ 114, 305, 338, 320 ], "score": 1.0, "content": "transformers using learnable memory. In CVPR, 2022.", "type": "text", "cross_page": true } ], "index": 16, "is_list_end_line": true }, { "bbox": [ 104, 323, 506, 340 ], "spans": [ { "bbox": [ 104, 323, 506, 340 ], "score": 1.0, "content": "Baifeng Shi, Siyu Gai, Trevor Darrell, and Xin Wang. Toast: Transfer learning via attention steering,", "type": "text", "cross_page": true } ], "index": 17, "is_list_start_line": true }, { "bbox": [ 115, 335, 142, 348 ], "spans": [ { "bbox": [ 115, 335, 142, 348 ], "score": 1.0, "content": "2023.", "type": "text", "cross_page": true } ], "index": 18, "is_list_end_line": true }, { "bbox": [ 106, 354, 506, 368 ], "spans": [ { "bbox": [ 106, 354, 506, 368 ], "score": 1.0, "content": "Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick ´", "type": "text", "cross_page": true } ], "index": 19, "is_list_start_line": true }, { "bbox": [ 115, 366, 506, 379 ], "spans": [ { "bbox": [ 115, 366, 506, 379 ], "score": 1.0, "content": "Perez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and ´", "type": "text", "cross_page": true } ], "index": 20 }, { "bbox": [ 115, 377, 227, 390 ], "spans": [ { "bbox": [ 115, 377, 227, 390 ], "score": 1.0, "content": "no labels. In BMVC, 2021.", "type": "text", "cross_page": true } ], "index": 21, "is_list_end_line": true }, { "bbox": [ 105, 396, 505, 409 ], "spans": [ { "bbox": [ 105, 396, 505, 409 ], "score": 1.0, "content": "Hwanjun Song, Deqing Sun, Sanghyuk Chun, Varun Jampani, Dongyoon Han, Byeongho Heo,", "type": "text", "cross_page": true } ], "index": 22, "is_list_start_line": true }, { "bbox": [ 115, 406, 506, 421 ], "spans": [ { "bbox": [ 115, 406, 506, 421 ], "score": 1.0, "content": "Wonjae Kim, and Ming-Hsuan Yang. Vidt: An efficient and effective fully transformer-based", "type": "text", "cross_page": true } ], "index": 23 }, { "bbox": [ 116, 418, 243, 430 ], "spans": [ { "bbox": [ 116, 418, 243, 430 ], "score": 1.0, "content": "object detector. In ICLR, 2021.", "type": "text", "cross_page": true } ], "index": 24, "is_list_end_line": true }, { "bbox": [ 106, 436, 452, 450 ], "spans": [ { "bbox": [ 106, 436, 452, 450 ], "score": 1.0, "content": "Antonio Torralba and Alexei A. Efros. Unbiased look at dataset bias. In CVPR, 2011.", "type": "text", "cross_page": true } ], "index": 25, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 105, 454, 486, 469 ], "spans": [ { "bbox": [ 105, 454, 486, 469 ], "score": 1.0, "content": "Hugo Touvron, Matthieu Cord, and Herve J ´ egou. Deit iii: Revenge of the vit. In ´ ECCV, 2022.", "type": "text", "cross_page": true } ], "index": 26, "is_list_start_line": true, "is_list_end_line": true }, { "bbox": [ 106, 474, 506, 488 ], "spans": [ { "bbox": [ 106, 474, 506, 488 ], "score": 1.0, "content": "Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object", "type": "text", "cross_page": true } ], "index": 27, "is_list_start_line": true }, { "bbox": [ 115, 485, 333, 499 ], "spans": [ { "bbox": [ 115, 485, 333, 499 ], "score": 1.0, "content": "detection and instance segmentation. In CVPR, 2023.", "type": "text", "cross_page": true } ], "index": 28, "is_list_end_line": true }, { "bbox": [ 105, 504, 505, 518 ], "spans": [ { "bbox": [ 105, 504, 505, 518 ], "score": 1.0, "content": "Fuzhao Xue, Valerii Likhosherstov, Anurag Arnab, Neil Houlsby, Mostafa Dehghani, and Yang You.", "type": "text", "cross_page": true } ], "index": 29, "is_list_start_line": true }, { "bbox": [ 116, 516, 385, 528 ], "spans": [ { "bbox": [ 116, 516, 385, 528 ], "score": 1.0, "content": "Adaptive computation with elastic input sequence. In ICML, 2023.", "type": "text", "cross_page": true } ], "index": 30, "is_list_end_line": true }, { "bbox": [ 105, 532, 506, 549 ], "spans": [ { "bbox": [ 105, 532, 506, 549 ], "score": 1.0, "content": "Yaodong Yu, Tianzhe Chu, Shengbang Tong, Ziyang Wu, Druv Pai, Sam Buchanan, and Yi Ma.", "type": "text", "cross_page": true } ], "index": 31, "is_list_start_line": true }, { "bbox": [ 115, 545, 466, 558 ], "spans": [ { "bbox": [ 115, 545, 466, 558 ], "score": 1.0, "content": "Emergence of segmentation with minimalistic white-box transformers. In CPAL, 2024.", "type": "text", "cross_page": true } ], "index": 32, "is_list_end_line": true }, { "bbox": [ 105, 564, 505, 577 ], "spans": [ { "bbox": [ 105, 564, 505, 577 ], "score": 1.0, "content": "Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei", "type": "text", "cross_page": true } ], "index": 33, "is_list_start_line": true }, { "bbox": [ 115, 575, 506, 588 ], "spans": [ { "bbox": [ 115, 575, 506, 588 ], "score": 1.0, "content": "Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from", "type": "text", "cross_page": true } ], "index": 34 }, { "bbox": [ 114, 586, 403, 599 ], "spans": [ { "bbox": [ 114, 586, 403, 599 ], "score": 1.0, "content": "a sequence-to-sequence perspective with transformers. In CVPR, 2021.", "type": "text", "cross_page": true } ], "index": 35, "is_list_end_line": true }, { "bbox": [ 105, 604, 506, 619 ], "spans": [ { "bbox": [ 105, 604, 506, 619 ], "score": 1.0, "content": "Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot:", "type": "text", "cross_page": true } ], "index": 36, "is_list_start_line": true }, { "bbox": [ 115, 616, 363, 630 ], "spans": [ { "bbox": [ 115, 616, 363, 630 ], "score": 1.0, "content": "Image bert pre-training with online tokenizer. In ICLR, 2022.", "type": "text", "cross_page": true } ], "index": 37, "is_list_end_line": true } ], "index": 25, "bbox_fs": [ 104, 181, 507, 734 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 100, 74, 507, 627 ], "lines": [ { "bbox": [ 105, 82, 505, 96 ], "spans": [ { "bbox": [ 105, 82, 505, 96 ], "score": 1.0, "content": "Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold,", "type": "text" } ], "index": 0 }, { "bbox": [ 115, 94, 505, 107 ], "spans": [ { "bbox": [ 115, 94, 505, 107 ], "score": 1.0, "content": "Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot atten-", "type": "text" } ], "index": 1 }, { "bbox": [ 115, 104, 213, 118 ], "spans": [ { "bbox": [ 115, 104, 213, 118 ], "score": 1.0, "content": "tion. In NeurIPS, 2020.", "type": "text" } ], "index": 2 }, { "bbox": [ 106, 124, 455, 136 ], "spans": [ { "bbox": [ 106, 124, 455, 136 ], "score": 1.0, "content": "David G Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2004.", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 142, 505, 155 ], "spans": [ { "bbox": [ 105, 142, 505, 155 ], "score": 1.0, "content": "Maxime Oquab, Timothee Darcet, Th ´ eo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, ´", "type": "text" } ], "index": 4 }, { "bbox": [ 114, 151, 506, 169 ], "spans": [ { "bbox": [ 114, 151, 506, 169 ], "score": 1.0, "content": "Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning", "type": "text" } ], "index": 5 }, { "bbox": [ 115, 164, 451, 178 ], "spans": [ { "bbox": [ 115, 164, 451, 178 ], "score": 1.0, "content": "robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023.", "type": "text" } ], "index": 6 }, { "bbox": [ 106, 183, 505, 197 ], "spans": [ { "bbox": [ 106, 183, 505, 197 ], "score": 1.0, "content": "Bill Psomas, Ioannis Kakogeorgiou, Konstantinos Karantzalos, and Yannis Avrithis. Keep it sim-", "type": "text" } ], "index": 7 }, { "bbox": [ 114, 193, 461, 209 ], "spans": [ { "bbox": [ 114, 193, 461, 209 ], "score": 1.0, "content": "pool: Who said supervised transformers suffer from attention deficit? In ICCV, 2023.", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 212, 505, 227 ], "spans": [ { "bbox": [ 105, 212, 505, 227 ], "score": 1.0, "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal,", "type": "text" } ], "index": 9 }, { "bbox": [ 114, 223, 506, 239 ], "spans": [ { "bbox": [ 114, 223, 506, 239 ], "score": 1.0, "content": "Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual", "type": "text" } ], "index": 10 }, { "bbox": [ 115, 236, 355, 249 ], "spans": [ { "bbox": [ 115, 236, 355, 249 ], "score": 1.0, "content": "models from natural language supervision. In ICML, 2021.", "type": "text" } ], "index": 11 }, { "bbox": [ 104, 253, 506, 269 ], "spans": [ { "bbox": [ 104, 253, 506, 269 ], "score": 1.0, "content": "Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng", "type": "text" } ], "index": 12 }, { "bbox": [ 114, 264, 506, 279 ], "spans": [ { "bbox": [ 114, 264, 506, 279 ], "score": 1.0, "content": "Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C Berg, and Li Fei-Fei.", "type": "text" } ], "index": 13 }, { "bbox": [ 115, 275, 370, 291 ], "spans": [ { "bbox": [ 115, 275, 370, 291 ], "score": 1.0, "content": "Imagenet large scale visual recognition challenge. IJCV, 2015.", "type": "text" } ], "index": 14 }, { "bbox": [ 104, 293, 506, 310 ], "spans": [ { "bbox": [ 104, 293, 506, 310 ], "score": 1.0, "content": "Mark Sandler, Andrey Zhmoginov, Max Vladymyrov, and Andrew Jackson. Fine-tuning image", "type": "text" } ], "index": 15 }, { "bbox": [ 114, 305, 338, 320 ], "spans": [ { "bbox": [ 114, 305, 338, 320 ], "score": 1.0, "content": "transformers using learnable memory. In CVPR, 2022.", "type": "text" } ], "index": 16 }, { "bbox": [ 104, 323, 506, 340 ], "spans": [ { "bbox": [ 104, 323, 506, 340 ], "score": 1.0, "content": "Baifeng Shi, Siyu Gai, Trevor Darrell, and Xin Wang. Toast: Transfer learning via attention steering,", "type": "text" } ], "index": 17 }, { "bbox": [ 115, 335, 142, 348 ], "spans": [ { "bbox": [ 115, 335, 142, 348 ], "score": 1.0, "content": "2023.", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 354, 506, 368 ], "spans": [ { "bbox": [ 106, 354, 506, 368 ], "score": 1.0, "content": "Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick ´", "type": "text" } ], "index": 19 }, { "bbox": [ 115, 366, 506, 379 ], "spans": [ { "bbox": [ 115, 366, 506, 379 ], "score": 1.0, "content": "Perez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and ´", "type": "text" } ], "index": 20 }, { "bbox": [ 115, 377, 227, 390 ], "spans": [ { "bbox": [ 115, 377, 227, 390 ], "score": 1.0, "content": "no labels. In BMVC, 2021.", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 396, 505, 409 ], "spans": [ { "bbox": [ 105, 396, 505, 409 ], "score": 1.0, "content": "Hwanjun Song, Deqing Sun, Sanghyuk Chun, Varun Jampani, Dongyoon Han, Byeongho Heo,", "type": "text" } ], "index": 22 }, { "bbox": [ 115, 406, 506, 421 ], "spans": [ { "bbox": [ 115, 406, 506, 421 ], "score": 1.0, "content": "Wonjae Kim, and Ming-Hsuan Yang. Vidt: An efficient and effective fully transformer-based", "type": "text" } ], "index": 23 }, { "bbox": [ 116, 418, 243, 430 ], "spans": [ { "bbox": [ 116, 418, 243, 430 ], "score": 1.0, "content": "object detector. In ICLR, 2021.", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 436, 452, 450 ], "spans": [ { "bbox": [ 106, 436, 452, 450 ], "score": 1.0, "content": "Antonio Torralba and Alexei A. Efros. Unbiased look at dataset bias. In CVPR, 2011.", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 454, 486, 469 ], "spans": [ { "bbox": [ 105, 454, 486, 469 ], "score": 1.0, "content": "Hugo Touvron, Matthieu Cord, and Herve J ´ egou. Deit iii: Revenge of the vit. In ´ ECCV, 2022.", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 474, 506, 488 ], "spans": [ { "bbox": [ 106, 474, 506, 488 ], "score": 1.0, "content": "Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object", "type": "text" } ], "index": 27 }, { "bbox": [ 115, 485, 333, 499 ], "spans": [ { "bbox": [ 115, 485, 333, 499 ], "score": 1.0, "content": "detection and instance segmentation. In CVPR, 2023.", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 504, 505, 518 ], "spans": [ { "bbox": [ 105, 504, 505, 518 ], "score": 1.0, "content": "Fuzhao Xue, Valerii Likhosherstov, Anurag Arnab, Neil Houlsby, Mostafa Dehghani, and Yang You.", "type": "text" } ], "index": 29 }, { "bbox": [ 116, 516, 385, 528 ], "spans": [ { "bbox": [ 116, 516, 385, 528 ], "score": 1.0, "content": "Adaptive computation with elastic input sequence. In ICML, 2023.", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 532, 506, 549 ], "spans": [ { "bbox": [ 105, 532, 506, 549 ], "score": 1.0, "content": "Yaodong Yu, Tianzhe Chu, Shengbang Tong, Ziyang Wu, Druv Pai, Sam Buchanan, and Yi Ma.", "type": "text" } ], "index": 31 }, { "bbox": [ 115, 545, 466, 558 ], "spans": [ { "bbox": [ 115, 545, 466, 558 ], "score": 1.0, "content": "Emergence of segmentation with minimalistic white-box transformers. In CPAL, 2024.", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 564, 505, 577 ], "spans": [ { "bbox": [ 105, 564, 505, 577 ], "score": 1.0, "content": "Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei", "type": "text" } ], "index": 33 }, { "bbox": [ 115, 575, 506, 588 ], "spans": [ { "bbox": [ 115, 575, 506, 588 ], "score": 1.0, "content": "Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from", "type": "text" } ], "index": 34 }, { "bbox": [ 114, 586, 403, 599 ], "spans": [ { "bbox": [ 114, 586, 403, 599 ], "score": 1.0, "content": "a sequence-to-sequence perspective with transformers. In CVPR, 2021.", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 604, 506, 619 ], "spans": [ { "bbox": [ 105, 604, 506, 619 ], "score": 1.0, "content": "Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot:", "type": "text" } ], "index": 36 }, { "bbox": [ 115, 616, 363, 630 ], "spans": [ { "bbox": [ 115, 616, 363, 630 ], "score": 1.0, "content": "Image bert pre-training with online tokenizer. In ICLR, 2022.", "type": "text" } ], "index": 37 } ], "index": 18.5 } ], "page_idx": 10, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 752, 310, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 765 ], "spans": [ { "bbox": [ 299, 750, 312, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "list", "bbox": [ 100, 74, 507, 627 ], "lines": [], "index": 18.5, "bbox_fs": [ 104, 82, 506, 630 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 163, 78, 445, 165 ], "blocks": [ { "type": "image_body", "bbox": [ 163, 78, 445, 165 ], "group_id": 0, "lines": [ { "bbox": [ 163, 78, 445, 165 ], "spans": [ { "bbox": [ 163, 78, 445, 165 ], "score": 0.963, "type": "image", "image_path": "75ec5aed27f6c58f47faea7645f37cca25a669603c12a1caa3c314687863319f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 163, 78, 445, 107.0 ], "spans": [], "index": 0 }, { "bbox": [ 163, 107.0, 445, 136.0 ], "spans": [], "index": 1 }, { "bbox": [ 163, 136.0, 445, 165.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 173, 505, 207 ], "group_id": 0, "lines": [ { "bbox": [ 106, 173, 506, 185 ], "spans": [ { "bbox": [ 106, 173, 506, 185 ], "score": 1.0, "content": "Figure 10: Feature norms along locations: proportion of tokens with norm larger than the cutoff", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 184, 505, 196 ], "spans": [ { "bbox": [ 106, 184, 505, 196 ], "score": 1.0, "content": "value at a given location. Left: official DINOv2 model (no antialiasing), right: our models (with", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 195, 414, 208 ], "spans": [ { "bbox": [ 106, 195, 280, 208 ], "score": 1.0, "content": "antialiasing). At some positions, more than", "type": "text" }, { "bbox": [ 281, 195, 300, 206 ], "score": 0.88, "content": "20 \\%", "type": "inline_equation" }, { "bbox": [ 301, 195, 414, 208 ], "score": 1.0, "content": "of tokens have a high norm.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "image", "bbox": [ 225, 230, 387, 352 ], "blocks": [ { "type": "image_body", "bbox": [ 225, 230, 387, 352 ], "group_id": 1, "lines": [ { "bbox": [ 225, 230, 387, 352 ], "spans": [ { "bbox": [ 225, 230, 387, 352 ], "score": 0.966, "type": "image", "image_path": "512234d5b9cf40198f2dbec953145d2191a347a6c3dcc4e164c9eb81b2c6267e.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 225, 230, 387, 243.55555555555554 ], "spans": [], "index": 6 }, { "bbox": [ 225, 243.55555555555554, 387, 257.1111111111111 ], "spans": [], "index": 7 }, { "bbox": [ 225, 257.1111111111111, 387, 270.66666666666663 ], "spans": [], "index": 8 }, { "bbox": [ 225, 270.66666666666663, 387, 284.2222222222222 ], "spans": [], "index": 9 }, { "bbox": [ 225, 284.2222222222222, 387, 297.7777777777777 ], "spans": [], "index": 10 }, { "bbox": [ 225, 297.7777777777777, 387, 311.33333333333326 ], "spans": [], "index": 11 }, { "bbox": [ 225, 311.33333333333326, 387, 324.8888888888888 ], "spans": [], "index": 12 }, { "bbox": [ 225, 324.8888888888888, 387, 338.44444444444434 ], "spans": [], "index": 13 }, { "bbox": [ 225, 338.44444444444434, 387, 351.9999999999999 ], "spans": [], "index": 14 } ] }, { "type": "image_caption", "bbox": [ 103, 362, 505, 385 ], "group_id": 1, "lines": [ { "bbox": [ 106, 361, 505, 374 ], "spans": [ { "bbox": [ 106, 361, 392, 374 ], "score": 1.0, "content": "Figure 11: Propagating unit gradients through a bicubic interpolation", "type": "text" }, { "bbox": [ 392, 362, 469, 373 ], "score": 0.87, "content": "( 1 6 \\times 1 6 \\to 7 \\times 7 )", "type": "inline_equation" }, { "bbox": [ 469, 361, 505, 374 ], "score": 1.0, "content": "without", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 372, 416, 385 ], "spans": [ { "bbox": [ 106, 372, 416, 385 ], "score": 1.0, "content": "antialiasing. We observe a striping pattern similar to the one of Fig. 10 (left).", "type": "text" } ], "index": 16 } ], "index": 15.5 } ], "index": 12.75 }, { "type": "title", "bbox": [ 108, 417, 472, 429 ], "lines": [ { "bbox": [ 105, 416, 474, 431 ], "spans": [ { "bbox": [ 105, 416, 474, 431 ], "score": 1.0, "content": "A INTERPOLATION ARTIFACTS AND OUTLIER POSITION DISTRIBUTION", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 108, 448, 505, 482 ], "lines": [ { "bbox": [ 106, 448, 505, 462 ], "spans": [ { "bbox": [ 106, 448, 505, 462 ], "score": 1.0, "content": "We plot in Figure 10 (left) the proportion of outlier tokens, characterized by a norm larger than the", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 460, 504, 472 ], "spans": [ { "bbox": [ 106, 460, 504, 472 ], "score": 1.0, "content": "cutoff value defined manually, following the distribution of norms shown in Fig. 3 (main text). We", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 471, 204, 483 ], "spans": [ { "bbox": [ 106, 471, 204, 483 ], "score": 1.0, "content": "make two observations:", "type": "text" } ], "index": 20 } ], "index": 19 }, { "type": "text", "bbox": [ 107, 487, 505, 565 ], "lines": [ { "bbox": [ 105, 487, 506, 500 ], "spans": [ { "bbox": [ 105, 487, 506, 500 ], "score": 1.0, "content": "First, the distribution has a vertical-striped pattern. We investigate this phenomenon and notice that", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 498, 506, 513 ], "spans": [ { "bbox": [ 105, 498, 506, 513 ], "score": 1.0, "content": "in the original DINOv2 implementation, during training the position embeddings are interpolated", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 509, 505, 523 ], "spans": [ { "bbox": [ 105, 509, 137, 523 ], "score": 1.0, "content": "from a", "type": "text" }, { "bbox": [ 137, 510, 172, 521 ], "score": 0.87, "content": "1 6 \\times 1 6", "type": "inline_equation" }, { "bbox": [ 173, 509, 222, 523 ], "score": 1.0, "content": "map into a", "type": "text" }, { "bbox": [ 222, 510, 247, 520 ], "score": 0.88, "content": "7 \\times 7", "type": "inline_equation" }, { "bbox": [ 248, 509, 505, 523 ], "score": 1.0, "content": "map, without antialiasing. Propagating unit gradients through", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 520, 505, 534 ], "spans": [ { "bbox": [ 105, 520, 505, 534 ], "score": 1.0, "content": "such an interpolation function (bicubic resize) leads to the following gradients, shown in Fig. 11.", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 532, 505, 545 ], "spans": [ { "bbox": [ 105, 532, 505, 545 ], "score": 1.0, "content": "In this work, when producing results with DINOv2 (especially for the results in Tables 2a,3), we", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 542, 506, 556 ], "spans": [ { "bbox": [ 105, 542, 506, 556 ], "score": 1.0, "content": "always apply antialiasing in the interpolation operator, removing the striping pattern, which gives an", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 554, 378, 567 ], "spans": [ { "bbox": [ 106, 554, 378, 567 ], "score": 1.0, "content": "updated distribution of outlier positions as shown in Fig. 10 (right).", "type": "text" } ], "index": 27 } ], "index": 24 }, { "type": "text", "bbox": [ 108, 570, 504, 615 ], "lines": [ { "bbox": [ 106, 571, 506, 583 ], "spans": [ { "bbox": [ 106, 571, 506, 583 ], "score": 1.0, "content": "Second, the outliers tend to appear in areas closer to the border of the feature map rather than in the", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 582, 506, 594 ], "spans": [ { "bbox": [ 106, 582, 506, 594 ], "score": 1.0, "content": "center. Our interpretation is that the base model tends to recycle tokens in low-informative areas to", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 594, 506, 604 ], "spans": [ { "bbox": [ 105, 594, 506, 604 ], "score": 1.0, "content": "use as registers: pictures produced by people tend to be object-centric, and in this case the border", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 604, 454, 615 ], "spans": [ { "bbox": [ 105, 604, 454, 615 ], "score": 1.0, "content": "areas often correspond to background, which contains less information than the center.", "type": "text" } ], "index": 31 } ], "index": 29.5 }, { "type": "title", "bbox": [ 108, 644, 251, 657 ], "lines": [ { "bbox": [ 105, 642, 252, 659 ], "spans": [ { "bbox": [ 105, 642, 252, 659 ], "score": 1.0, "content": "B COMPLEXITY ANALYSIS", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 108, 676, 504, 732 ], "lines": [ { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "Since our proposed fix introduces new tokens, it also increases the number of learnable parameters", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "and the FLOP count of the model. We show in Fig. 12 the relationship between number of registers", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "and increase in model FLOP count and parameter count. We observe that adding registers induces", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 709, 504, 723 ], "spans": [ { "bbox": [ 106, 709, 477, 723 ], "score": 1.0, "content": "a negligible change in number of parameters, and a slight change in FLOP count. Still, for", "type": "text" }, { "bbox": [ 477, 710, 504, 720 ], "score": 0.89, "content": "n = 4", "type": "inline_equation" } ], "index": 36 }, { "bbox": [ 106, 721, 303, 732 ], "spans": [ { "bbox": [ 106, 721, 285, 732 ], "score": 1.0, "content": "registers, the increase in FLOPs stays below", "type": "text" }, { "bbox": [ 285, 721, 299, 731 ], "score": 0.85, "content": "2 \\%", "type": "inline_equation" }, { "bbox": [ 299, 721, 303, 732 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 37 } ], "index": 35 } ], "page_idx": 11, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 106, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 294, 38 ], "spans": [ { "bbox": [ 106, 25, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 752, 311, 760 ], "lines": [] } ], "para_blocks": [ { "type": "image", "bbox": [ 163, 78, 445, 165 ], "blocks": [ { "type": "image_body", "bbox": [ 163, 78, 445, 165 ], "group_id": 0, "lines": [ { "bbox": [ 163, 78, 445, 165 ], "spans": [ { "bbox": [ 163, 78, 445, 165 ], "score": 0.963, "type": "image", "image_path": "75ec5aed27f6c58f47faea7645f37cca25a669603c12a1caa3c314687863319f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 163, 78, 445, 107.0 ], "spans": [], "index": 0 }, { "bbox": [ 163, 107.0, 445, 136.0 ], "spans": [], "index": 1 }, { "bbox": [ 163, 136.0, 445, 165.0 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 173, 505, 207 ], "group_id": 0, "lines": [ { "bbox": [ 106, 173, 506, 185 ], "spans": [ { "bbox": [ 106, 173, 506, 185 ], "score": 1.0, "content": "Figure 10: Feature norms along locations: proportion of tokens with norm larger than the cutoff", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 184, 505, 196 ], "spans": [ { "bbox": [ 106, 184, 505, 196 ], "score": 1.0, "content": "value at a given location. Left: official DINOv2 model (no antialiasing), right: our models (with", "type": "text" } ], "index": 4 }, { "bbox": [ 106, 195, 414, 208 ], "spans": [ { "bbox": [ 106, 195, 280, 208 ], "score": 1.0, "content": "antialiasing). At some positions, more than", "type": "text" }, { "bbox": [ 281, 195, 300, 206 ], "score": 0.88, "content": "20 \\%", "type": "inline_equation" }, { "bbox": [ 301, 195, 414, 208 ], "score": 1.0, "content": "of tokens have a high norm.", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 }, { "type": "image", "bbox": [ 225, 230, 387, 352 ], "blocks": [ { "type": "image_body", "bbox": [ 225, 230, 387, 352 ], "group_id": 1, "lines": [ { "bbox": [ 225, 230, 387, 352 ], "spans": [ { "bbox": [ 225, 230, 387, 352 ], "score": 0.966, "type": "image", "image_path": "512234d5b9cf40198f2dbec953145d2191a347a6c3dcc4e164c9eb81b2c6267e.jpg" } ] } ], "index": 10, "virtual_lines": [ { "bbox": [ 225, 230, 387, 243.55555555555554 ], "spans": [], "index": 6 }, { "bbox": [ 225, 243.55555555555554, 387, 257.1111111111111 ], "spans": [], "index": 7 }, { "bbox": [ 225, 257.1111111111111, 387, 270.66666666666663 ], "spans": [], "index": 8 }, { "bbox": [ 225, 270.66666666666663, 387, 284.2222222222222 ], "spans": [], "index": 9 }, { "bbox": [ 225, 284.2222222222222, 387, 297.7777777777777 ], "spans": [], "index": 10 }, { "bbox": [ 225, 297.7777777777777, 387, 311.33333333333326 ], "spans": [], "index": 11 }, { "bbox": [ 225, 311.33333333333326, 387, 324.8888888888888 ], "spans": [], "index": 12 }, { "bbox": [ 225, 324.8888888888888, 387, 338.44444444444434 ], "spans": [], "index": 13 }, { "bbox": [ 225, 338.44444444444434, 387, 351.9999999999999 ], "spans": [], "index": 14 } ] }, { "type": "image_caption", "bbox": [ 103, 362, 505, 385 ], "group_id": 1, "lines": [ { "bbox": [ 106, 361, 505, 374 ], "spans": [ { "bbox": [ 106, 361, 392, 374 ], "score": 1.0, "content": "Figure 11: Propagating unit gradients through a bicubic interpolation", "type": "text" }, { "bbox": [ 392, 362, 469, 373 ], "score": 0.87, "content": "( 1 6 \\times 1 6 \\to 7 \\times 7 )", "type": "inline_equation" }, { "bbox": [ 469, 361, 505, 374 ], "score": 1.0, "content": "without", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 372, 416, 385 ], "spans": [ { "bbox": [ 106, 372, 416, 385 ], "score": 1.0, "content": "antialiasing. We observe a striping pattern similar to the one of Fig. 10 (left).", "type": "text" } ], "index": 16 } ], "index": 15.5 } ], "index": 12.75 }, { "type": "title", "bbox": [ 108, 417, 472, 429 ], "lines": [ { "bbox": [ 105, 416, 474, 431 ], "spans": [ { "bbox": [ 105, 416, 474, 431 ], "score": 1.0, "content": "A INTERPOLATION ARTIFACTS AND OUTLIER POSITION DISTRIBUTION", "type": "text" } ], "index": 17 } ], "index": 17 }, { "type": "text", "bbox": [ 108, 448, 505, 482 ], "lines": [ { "bbox": [ 106, 448, 505, 462 ], "spans": [ { "bbox": [ 106, 448, 505, 462 ], "score": 1.0, "content": "We plot in Figure 10 (left) the proportion of outlier tokens, characterized by a norm larger than the", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 460, 504, 472 ], "spans": [ { "bbox": [ 106, 460, 504, 472 ], "score": 1.0, "content": "cutoff value defined manually, following the distribution of norms shown in Fig. 3 (main text). We", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 471, 204, 483 ], "spans": [ { "bbox": [ 106, 471, 204, 483 ], "score": 1.0, "content": "make two observations:", "type": "text" } ], "index": 20 } ], "index": 19, "bbox_fs": [ 106, 448, 505, 483 ] }, { "type": "text", "bbox": [ 107, 487, 505, 565 ], "lines": [ { "bbox": [ 105, 487, 506, 500 ], "spans": [ { "bbox": [ 105, 487, 506, 500 ], "score": 1.0, "content": "First, the distribution has a vertical-striped pattern. We investigate this phenomenon and notice that", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 498, 506, 513 ], "spans": [ { "bbox": [ 105, 498, 506, 513 ], "score": 1.0, "content": "in the original DINOv2 implementation, during training the position embeddings are interpolated", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 509, 505, 523 ], "spans": [ { "bbox": [ 105, 509, 137, 523 ], "score": 1.0, "content": "from a", "type": "text" }, { "bbox": [ 137, 510, 172, 521 ], "score": 0.87, "content": "1 6 \\times 1 6", "type": "inline_equation" }, { "bbox": [ 173, 509, 222, 523 ], "score": 1.0, "content": "map into a", "type": "text" }, { "bbox": [ 222, 510, 247, 520 ], "score": 0.88, "content": "7 \\times 7", "type": "inline_equation" }, { "bbox": [ 248, 509, 505, 523 ], "score": 1.0, "content": "map, without antialiasing. Propagating unit gradients through", "type": "text" } ], "index": 23 }, { "bbox": [ 105, 520, 505, 534 ], "spans": [ { "bbox": [ 105, 520, 505, 534 ], "score": 1.0, "content": "such an interpolation function (bicubic resize) leads to the following gradients, shown in Fig. 11.", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 532, 505, 545 ], "spans": [ { "bbox": [ 105, 532, 505, 545 ], "score": 1.0, "content": "In this work, when producing results with DINOv2 (especially for the results in Tables 2a,3), we", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 542, 506, 556 ], "spans": [ { "bbox": [ 105, 542, 506, 556 ], "score": 1.0, "content": "always apply antialiasing in the interpolation operator, removing the striping pattern, which gives an", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 554, 378, 567 ], "spans": [ { "bbox": [ 106, 554, 378, 567 ], "score": 1.0, "content": "updated distribution of outlier positions as shown in Fig. 10 (right).", "type": "text" } ], "index": 27 } ], "index": 24, "bbox_fs": [ 105, 487, 506, 567 ] }, { "type": "text", "bbox": [ 108, 570, 504, 615 ], "lines": [ { "bbox": [ 106, 571, 506, 583 ], "spans": [ { "bbox": [ 106, 571, 506, 583 ], "score": 1.0, "content": "Second, the outliers tend to appear in areas closer to the border of the feature map rather than in the", "type": "text" } ], "index": 28 }, { "bbox": [ 106, 582, 506, 594 ], "spans": [ { "bbox": [ 106, 582, 506, 594 ], "score": 1.0, "content": "center. Our interpretation is that the base model tends to recycle tokens in low-informative areas to", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 594, 506, 604 ], "spans": [ { "bbox": [ 105, 594, 506, 604 ], "score": 1.0, "content": "use as registers: pictures produced by people tend to be object-centric, and in this case the border", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 604, 454, 615 ], "spans": [ { "bbox": [ 105, 604, 454, 615 ], "score": 1.0, "content": "areas often correspond to background, which contains less information than the center.", "type": "text" } ], "index": 31 } ], "index": 29.5, "bbox_fs": [ 105, 571, 506, 615 ] }, { "type": "title", "bbox": [ 108, 644, 251, 657 ], "lines": [ { "bbox": [ 105, 642, 252, 659 ], "spans": [ { "bbox": [ 105, 642, 252, 659 ], "score": 1.0, "content": "B COMPLEXITY ANALYSIS", "type": "text" } ], "index": 32 } ], "index": 32 }, { "type": "text", "bbox": [ 108, 676, 504, 732 ], "lines": [ { "bbox": [ 106, 677, 505, 689 ], "spans": [ { "bbox": [ 106, 677, 505, 689 ], "score": 1.0, "content": "Since our proposed fix introduces new tokens, it also increases the number of learnable parameters", "type": "text" } ], "index": 33 }, { "bbox": [ 106, 688, 505, 700 ], "spans": [ { "bbox": [ 106, 688, 505, 700 ], "score": 1.0, "content": "and the FLOP count of the model. We show in Fig. 12 the relationship between number of registers", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 698, 506, 712 ], "spans": [ { "bbox": [ 105, 698, 506, 712 ], "score": 1.0, "content": "and increase in model FLOP count and parameter count. We observe that adding registers induces", "type": "text" } ], "index": 35 }, { "bbox": [ 106, 709, 504, 723 ], "spans": [ { "bbox": [ 106, 709, 477, 723 ], "score": 1.0, "content": "a negligible change in number of parameters, and a slight change in FLOP count. Still, for", "type": "text" }, { "bbox": [ 477, 710, 504, 720 ], "score": 0.89, "content": "n = 4", "type": "inline_equation" } ], "index": 36 }, { "bbox": [ 106, 721, 303, 732 ], "spans": [ { "bbox": [ 106, 721, 285, 732 ], "score": 1.0, "content": "registers, the increase in FLOPs stays below", "type": "text" }, { "bbox": [ 285, 721, 299, 731 ], "score": 0.85, "content": "2 \\%", "type": "inline_equation" }, { "bbox": [ 299, 721, 303, 732 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 37 } ], "index": 35, "bbox_fs": [ 105, 677, 506, 732 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 251, 85, 361, 172 ], "blocks": [ { "type": "image_body", "bbox": [ 251, 85, 361, 172 ], "group_id": 0, "lines": [ { "bbox": [ 251, 85, 361, 172 ], "spans": [ { "bbox": [ 251, 85, 361, 172 ], "score": 0.935, "type": "image", "image_path": "2222d5bdba5d04e7d0db7da11c5b233e7cd962723e2b8705fdc67ba53d52d124.jpg" } ] } ], "index": 0.5, "virtual_lines": [ { "bbox": [ 251, 85, 361, 128.5 ], "spans": [], "index": 0 }, { "bbox": [ 251, 128.5, 361, 172.0 ], "spans": [], "index": 1 } ] }, { "type": "image_caption", "bbox": [ 106, 185, 505, 231 ], "group_id": 0, "lines": [ { "bbox": [ 105, 185, 505, 199 ], "spans": [ { "bbox": [ 105, 185, 505, 199 ], "score": 1.0, "content": "Figure 12: Increase in model parameter and FLOP count when adding different numbers of registers.", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 196, 506, 210 ], "spans": [ { "bbox": [ 105, 196, 336, 210 ], "score": 1.0, "content": "Adding registers can increase model FLOP count by up to", "type": "text" }, { "bbox": [ 337, 197, 352, 208 ], "score": 0.86, "content": "6 \\%", "type": "inline_equation" }, { "bbox": [ 352, 196, 506, 210 ], "score": 1.0, "content": "for 16 registers. However, in the more", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 208, 506, 220 ], "spans": [ { "bbox": [ 105, 208, 486, 220 ], "score": 1.0, "content": "common case of using 4 registers, that we use in most of our experiments, this increase is below", "type": "text" }, { "bbox": [ 487, 208, 501, 218 ], "score": 0.85, "content": "2 \\%", "type": "inline_equation" }, { "bbox": [ 502, 208, 506, 220 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 218, 343, 231 ], "spans": [ { "bbox": [ 105, 218, 343, 231 ], "score": 1.0, "content": "In all cases, the increase in model parameters is negligible.", "type": "text" } ], "index": 5 } ], "index": 3.5 } ], "index": 2.0 }, { "type": "image", "bbox": [ 109, 248, 500, 379 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 248, 500, 379 ], "group_id": 1, "lines": [ { "bbox": [ 109, 248, 500, 379 ], "spans": [ { "bbox": [ 109, 248, 500, 379 ], "score": 0.972, "type": "image", "image_path": "02b7a0bdd50679a9421138f3523bee90c615663ce63d59e52a17f29cf4435b40.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 109, 248, 500, 291.6666666666667 ], "spans": [], "index": 6 }, { "bbox": [ 109, 291.6666666666667, 500, 335.33333333333337 ], "spans": [], "index": 7 }, { "bbox": [ 109, 335.33333333333337, 500, 379.00000000000006 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 106, 393, 504, 428 ], "group_id": 1, "lines": [ { "bbox": [ 106, 394, 504, 406 ], "spans": [ { "bbox": [ 106, 394, 504, 406 ], "score": 1.0, "content": "Figure 13: Illustration of the intermediate computations in the LOST algorithm for all models.", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 403, 505, 417 ], "spans": [ { "bbox": [ 105, 403, 505, 417 ], "score": 1.0, "content": "Adding registers drastically improves the look of all intermediate steps for DeiT-III and DINOv2.", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 415, 331, 429 ], "spans": [ { "bbox": [ 105, 415, 331, 429 ], "score": 1.0, "content": "The difference is less striking for the OpenCLIP model.", "type": "text" } ], "index": 11 } ], "index": 10 } ], "index": 8.5 }, { "type": "title", "bbox": [ 108, 451, 312, 464 ], "lines": [ { "bbox": [ 106, 450, 315, 466 ], "spans": [ { "bbox": [ 106, 450, 315, 466 ], "score": 1.0, "content": "C ANALYSIS OF LOST PERFORMANCE", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 478, 504, 534 ], "lines": [ { "bbox": [ 106, 478, 506, 492 ], "spans": [ { "bbox": [ 106, 478, 506, 492 ], "score": 1.0, "content": "The results presented in Sec. 3.3 show that adding registers allows us to obtain better object dis-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 489, 506, 502 ], "spans": [ { "bbox": [ 105, 489, 506, 502 ], "score": 1.0, "content": "covery performance with DINOv2 models. The conclusions for the two other models studied in", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 500, 505, 513 ], "spans": [ { "bbox": [ 105, 500, 505, 513 ], "score": 1.0, "content": "this work could be more crisp. In order to understand why this is so, we qualitatively study the", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 511, 506, 524 ], "spans": [ { "bbox": [ 105, 511, 506, 524 ], "score": 1.0, "content": "impact of removing artifacts on the intermediate computations in the LOST algorithm. We show the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 522, 424, 536 ], "spans": [ { "bbox": [ 105, 522, 424, 536 ], "score": 1.0, "content": "intermediate outputs of LOST for all models on a given input image in Fig. 13.", "type": "text" } ], "index": 17 } ], "index": 15 }, { "type": "text", "bbox": [ 107, 539, 504, 616 ], "lines": [ { "bbox": [ 106, 539, 505, 552 ], "spans": [ { "bbox": [ 106, 539, 505, 552 ], "score": 1.0, "content": "Adding registers improves the scores and the resulting seed expansion for DeiT-III and DINOv2.", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 550, 506, 563 ], "spans": [ { "bbox": [ 106, 550, 506, 563 ], "score": 1.0, "content": "This observation is coherent with the improved numbers reported in Table 3. For OpenCLIP, how-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 560, 506, 574 ], "spans": [ { "bbox": [ 105, 560, 506, 574 ], "score": 1.0, "content": "ever, the LOST algorithm seems robust to the type of outliers observed in the local features. Adding", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 572, 506, 585 ], "spans": [ { "bbox": [ 105, 572, 506, 585 ], "score": 1.0, "content": "registers does remove artifacts (as clearly shown in Fig. 20) but does not have much impact on the", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 583, 506, 596 ], "spans": [ { "bbox": [ 105, 583, 506, 596 ], "score": 1.0, "content": "LOST score. It is also worth noting that OpenCLIP, with or without registers, provides comparable", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 593, 506, 608 ], "spans": [ { "bbox": [ 105, 593, 506, 608 ], "score": 1.0, "content": "performance to DINOv2 without registers and DeiT-III with registers. The qualitative assessment is", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 605, 295, 618 ], "spans": [ { "bbox": [ 106, 605, 295, 618 ], "score": 1.0, "content": "coherent with the numbers reported in Table 3.", "type": "text" } ], "index": 24 } ], "index": 21 }, { "type": "text", "bbox": [ 107, 622, 504, 732 ], "lines": [ { "bbox": [ 105, 621, 505, 635 ], "spans": [ { "bbox": [ 105, 621, 505, 635 ], "score": 1.0, "content": "A surprising observation is that despite the existence of high-norm patches in the output of Open-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 632, 505, 645 ], "spans": [ { "bbox": [ 105, 632, 505, 645 ], "score": 1.0, "content": "CLIP models without registers (as seen in Fig. 7), the seed expansion score in Fig. 13 looks smooth.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 644, 506, 657 ], "spans": [ { "bbox": [ 105, 644, 506, 657 ], "score": 1.0, "content": "In the LOST experiment with OpenCLIP models, we do not use the features directly, but the values", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 654, 506, 667 ], "spans": [ { "bbox": [ 105, 654, 506, 667 ], "score": 1.0, "content": "from the computation of attention maps. In Fig. 14, we show the seed expansion score for Open-", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 665, 506, 679 ], "spans": [ { "bbox": [ 105, 665, 506, 679 ], "score": 1.0, "content": "CLIP models with and without registers for keys, queries and values. We see that artifacts are clearly", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 677, 506, 689 ], "spans": [ { "bbox": [ 106, 677, 506, 689 ], "score": 1.0, "content": "visible as spots in the background for keys and queries, for the model without registers. As soon as", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "registers are used, the LOST score is focusing on the object, with a smoother score for values. We", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 506, 712 ], "score": 1.0, "content": "qualitatively observe that for the OpenCLIP model, the value projection filters out the outliers even", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 710, 506, 723 ], "spans": [ { "bbox": [ 105, 710, 506, 723 ], "score": 1.0, "content": "without registers. This means that the outliers appear to live in the null space of the value projection", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 721, 375, 732 ], "spans": [ { "bbox": [ 105, 721, 375, 732 ], "score": 1.0, "content": "layer; the investigation for this phenomenon is left for future work.", "type": "text" } ], "index": 34 } ], "index": 29.5 } ], "page_idx": 12, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 25, 294, 38 ], "spans": [ { "bbox": [ 106, 25, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 752, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 251, 85, 361, 172 ], "blocks": [ { "type": "image_body", "bbox": [ 251, 85, 361, 172 ], "group_id": 0, "lines": [ { "bbox": [ 251, 85, 361, 172 ], "spans": [ { "bbox": [ 251, 85, 361, 172 ], "score": 0.935, "type": "image", "image_path": "2222d5bdba5d04e7d0db7da11c5b233e7cd962723e2b8705fdc67ba53d52d124.jpg" } ] } ], "index": 0.5, "virtual_lines": [ { "bbox": [ 251, 85, 361, 128.5 ], "spans": [], "index": 0 }, { "bbox": [ 251, 128.5, 361, 172.0 ], "spans": [], "index": 1 } ] }, { "type": "image_caption", "bbox": [ 106, 185, 505, 231 ], "group_id": 0, "lines": [ { "bbox": [ 105, 185, 505, 199 ], "spans": [ { "bbox": [ 105, 185, 505, 199 ], "score": 1.0, "content": "Figure 12: Increase in model parameter and FLOP count when adding different numbers of registers.", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 196, 506, 210 ], "spans": [ { "bbox": [ 105, 196, 336, 210 ], "score": 1.0, "content": "Adding registers can increase model FLOP count by up to", "type": "text" }, { "bbox": [ 337, 197, 352, 208 ], "score": 0.86, "content": "6 \\%", "type": "inline_equation" }, { "bbox": [ 352, 196, 506, 210 ], "score": 1.0, "content": "for 16 registers. However, in the more", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 208, 506, 220 ], "spans": [ { "bbox": [ 105, 208, 486, 220 ], "score": 1.0, "content": "common case of using 4 registers, that we use in most of our experiments, this increase is below", "type": "text" }, { "bbox": [ 487, 208, 501, 218 ], "score": 0.85, "content": "2 \\%", "type": "inline_equation" }, { "bbox": [ 502, 208, 506, 220 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 218, 343, 231 ], "spans": [ { "bbox": [ 105, 218, 343, 231 ], "score": 1.0, "content": "In all cases, the increase in model parameters is negligible.", "type": "text" } ], "index": 5 } ], "index": 3.5 } ], "index": 2.0 }, { "type": "image", "bbox": [ 109, 248, 500, 379 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 248, 500, 379 ], "group_id": 1, "lines": [ { "bbox": [ 109, 248, 500, 379 ], "spans": [ { "bbox": [ 109, 248, 500, 379 ], "score": 0.972, "type": "image", "image_path": "02b7a0bdd50679a9421138f3523bee90c615663ce63d59e52a17f29cf4435b40.jpg" } ] } ], "index": 7, "virtual_lines": [ { "bbox": [ 109, 248, 500, 291.6666666666667 ], "spans": [], "index": 6 }, { "bbox": [ 109, 291.6666666666667, 500, 335.33333333333337 ], "spans": [], "index": 7 }, { "bbox": [ 109, 335.33333333333337, 500, 379.00000000000006 ], "spans": [], "index": 8 } ] }, { "type": "image_caption", "bbox": [ 106, 393, 504, 428 ], "group_id": 1, "lines": [ { "bbox": [ 106, 394, 504, 406 ], "spans": [ { "bbox": [ 106, 394, 504, 406 ], "score": 1.0, "content": "Figure 13: Illustration of the intermediate computations in the LOST algorithm for all models.", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 403, 505, 417 ], "spans": [ { "bbox": [ 105, 403, 505, 417 ], "score": 1.0, "content": "Adding registers drastically improves the look of all intermediate steps for DeiT-III and DINOv2.", "type": "text" } ], "index": 10 }, { "bbox": [ 105, 415, 331, 429 ], "spans": [ { "bbox": [ 105, 415, 331, 429 ], "score": 1.0, "content": "The difference is less striking for the OpenCLIP model.", "type": "text" } ], "index": 11 } ], "index": 10 } ], "index": 8.5 }, { "type": "title", "bbox": [ 108, 451, 312, 464 ], "lines": [ { "bbox": [ 106, 450, 315, 466 ], "spans": [ { "bbox": [ 106, 450, 315, 466 ], "score": 1.0, "content": "C ANALYSIS OF LOST PERFORMANCE", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 107, 478, 504, 534 ], "lines": [ { "bbox": [ 106, 478, 506, 492 ], "spans": [ { "bbox": [ 106, 478, 506, 492 ], "score": 1.0, "content": "The results presented in Sec. 3.3 show that adding registers allows us to obtain better object dis-", "type": "text" } ], "index": 13 }, { "bbox": [ 105, 489, 506, 502 ], "spans": [ { "bbox": [ 105, 489, 506, 502 ], "score": 1.0, "content": "covery performance with DINOv2 models. The conclusions for the two other models studied in", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 500, 505, 513 ], "spans": [ { "bbox": [ 105, 500, 505, 513 ], "score": 1.0, "content": "this work could be more crisp. In order to understand why this is so, we qualitatively study the", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 511, 506, 524 ], "spans": [ { "bbox": [ 105, 511, 506, 524 ], "score": 1.0, "content": "impact of removing artifacts on the intermediate computations in the LOST algorithm. We show the", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 522, 424, 536 ], "spans": [ { "bbox": [ 105, 522, 424, 536 ], "score": 1.0, "content": "intermediate outputs of LOST for all models on a given input image in Fig. 13.", "type": "text" } ], "index": 17 } ], "index": 15, "bbox_fs": [ 105, 478, 506, 536 ] }, { "type": "text", "bbox": [ 107, 539, 504, 616 ], "lines": [ { "bbox": [ 106, 539, 505, 552 ], "spans": [ { "bbox": [ 106, 539, 505, 552 ], "score": 1.0, "content": "Adding registers improves the scores and the resulting seed expansion for DeiT-III and DINOv2.", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 550, 506, 563 ], "spans": [ { "bbox": [ 106, 550, 506, 563 ], "score": 1.0, "content": "This observation is coherent with the improved numbers reported in Table 3. For OpenCLIP, how-", "type": "text" } ], "index": 19 }, { "bbox": [ 105, 560, 506, 574 ], "spans": [ { "bbox": [ 105, 560, 506, 574 ], "score": 1.0, "content": "ever, the LOST algorithm seems robust to the type of outliers observed in the local features. Adding", "type": "text" } ], "index": 20 }, { "bbox": [ 105, 572, 506, 585 ], "spans": [ { "bbox": [ 105, 572, 506, 585 ], "score": 1.0, "content": "registers does remove artifacts (as clearly shown in Fig. 20) but does not have much impact on the", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 583, 506, 596 ], "spans": [ { "bbox": [ 105, 583, 506, 596 ], "score": 1.0, "content": "LOST score. It is also worth noting that OpenCLIP, with or without registers, provides comparable", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 593, 506, 608 ], "spans": [ { "bbox": [ 105, 593, 506, 608 ], "score": 1.0, "content": "performance to DINOv2 without registers and DeiT-III with registers. The qualitative assessment is", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 605, 295, 618 ], "spans": [ { "bbox": [ 106, 605, 295, 618 ], "score": 1.0, "content": "coherent with the numbers reported in Table 3.", "type": "text" } ], "index": 24 } ], "index": 21, "bbox_fs": [ 105, 539, 506, 618 ] }, { "type": "text", "bbox": [ 107, 622, 504, 732 ], "lines": [ { "bbox": [ 105, 621, 505, 635 ], "spans": [ { "bbox": [ 105, 621, 505, 635 ], "score": 1.0, "content": "A surprising observation is that despite the existence of high-norm patches in the output of Open-", "type": "text" } ], "index": 25 }, { "bbox": [ 105, 632, 505, 645 ], "spans": [ { "bbox": [ 105, 632, 505, 645 ], "score": 1.0, "content": "CLIP models without registers (as seen in Fig. 7), the seed expansion score in Fig. 13 looks smooth.", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 644, 506, 657 ], "spans": [ { "bbox": [ 105, 644, 506, 657 ], "score": 1.0, "content": "In the LOST experiment with OpenCLIP models, we do not use the features directly, but the values", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 654, 506, 667 ], "spans": [ { "bbox": [ 105, 654, 506, 667 ], "score": 1.0, "content": "from the computation of attention maps. In Fig. 14, we show the seed expansion score for Open-", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 665, 506, 679 ], "spans": [ { "bbox": [ 105, 665, 506, 679 ], "score": 1.0, "content": "CLIP models with and without registers for keys, queries and values. We see that artifacts are clearly", "type": "text" } ], "index": 29 }, { "bbox": [ 106, 677, 506, 689 ], "spans": [ { "bbox": [ 106, 677, 506, 689 ], "score": 1.0, "content": "visible as spots in the background for keys and queries, for the model without registers. As soon as", "type": "text" } ], "index": 30 }, { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "registers are used, the LOST score is focusing on the object, with a smoother score for values. We", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 506, 712 ], "score": 1.0, "content": "qualitatively observe that for the OpenCLIP model, the value projection filters out the outliers even", "type": "text" } ], "index": 32 }, { "bbox": [ 105, 710, 506, 723 ], "spans": [ { "bbox": [ 105, 710, 506, 723 ], "score": 1.0, "content": "without registers. This means that the outliers appear to live in the null space of the value projection", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 721, 375, 732 ], "spans": [ { "bbox": [ 105, 721, 375, 732 ], "score": 1.0, "content": "layer; the investigation for this phenomenon is left for future work.", "type": "text" } ], "index": 34 } ], "index": 29.5, "bbox_fs": [ 105, 621, 506, 732 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 176, 86, 424, 201 ], "blocks": [ { "type": "image_body", "bbox": [ 176, 86, 424, 201 ], "group_id": 0, "lines": [ { "bbox": [ 176, 86, 424, 201 ], "spans": [ { "bbox": [ 176, 86, 424, 201 ], "score": 0.971, "type": "image", "image_path": "407c5844ed49ac0566563545f5eed24d6bde5e34ebdebeaf4112a7e95d8e45c2.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 176, 86, 424, 124.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 176, 124.33333333333334, 424, 162.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 176, 162.66666666666669, 424, 201.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 214, 505, 259 ], "group_id": 0, "lines": [ { "bbox": [ 106, 214, 506, 227 ], "spans": [ { "bbox": [ 106, 214, 506, 227 ], "score": 1.0, "content": "Figure 14: Illustration of the seed expansion score in LOST for an OpenCLIP model with and", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 225, 506, 238 ], "spans": [ { "bbox": [ 106, 225, 506, 238 ], "score": 1.0, "content": "without registers for the three types of features considered: keys, queries, and values. The score", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 236, 506, 249 ], "spans": [ { "bbox": [ 105, 236, 506, 249 ], "score": 1.0, "content": "is qualitatively improved across all features, with fewer artifacts appearing. Interestingly, the seed", "type": "text" } ], "index": 5 }, { "bbox": [ 104, 247, 471, 260 ], "spans": [ { "bbox": [ 104, 247, 471, 260 ], "score": 1.0, "content": "expansion map computed using values does not exhibit artifacts with nor without registers.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "title", "bbox": [ 107, 280, 385, 292 ], "lines": [ { "bbox": [ 105, 279, 386, 295 ], "spans": [ { "bbox": [ 105, 279, 386, 295 ], "score": 1.0, "content": "D BEHAVIOR OF MODELS TRAINED WITH REGISTERS", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 106, 305, 506, 327 ], "lines": [ { "bbox": [ 105, 304, 505, 317 ], "spans": [ { "bbox": [ 105, 304, 505, 317 ], "score": 1.0, "content": "In order to better understand the phenomenon at hand, we examine the question of to what extent", "type": "text" } ], "index": 8 }, { "bbox": [ 107, 316, 432, 327 ], "spans": [ { "bbox": [ 107, 316, 432, 327 ], "score": 1.0, "content": "did the register tokens ”replace” the high-norm tokens and took on the same role.", "type": "text" } ], "index": 9 } ], "index": 8.5 }, { "type": "image", "bbox": [ 115, 370, 500, 501 ], "blocks": [ { "type": "image_body", "bbox": [ 115, 370, 500, 501 ], "group_id": 1, "lines": [ { "bbox": [ 115, 370, 500, 501 ], "spans": [ { "bbox": [ 115, 370, 500, 501 ], "score": 0.961, "type": "image", "image_path": "802c3c84876b53f95537280862cacf3d102d0e5feb5e1d103c9b24e5c941af84.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 115, 370, 500, 413.6666666666667 ], "spans": [], "index": 10 }, { "bbox": [ 115, 413.6666666666667, 500, 457.33333333333337 ], "spans": [], "index": 11 }, { "bbox": [ 115, 457.33333333333337, 500, 501.00000000000006 ], "spans": [], "index": 12 } ] }, { "type": "image_caption", "bbox": [ 103, 510, 504, 533 ], "group_id": 1, "lines": [ { "bbox": [ 106, 510, 505, 523 ], "spans": [ { "bbox": [ 106, 510, 505, 523 ], "score": 1.0, "content": "Figure 15: Distribution of token norms for a DINOv2 model without (left) and with (right) 4 regis-", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 522, 468, 534 ], "spans": [ { "bbox": [ 106, 522, 468, 534 ], "score": 1.0, "content": "ters. Introducing registers entirely negates the high-norm outliers among the patch tokens.", "type": "text" } ], "index": 14 } ], "index": 13.5 } ], "index": 12.25 }, { "type": "text", "bbox": [ 107, 547, 505, 614 ], "lines": [ { "bbox": [ 105, 547, 506, 560 ], "spans": [ { "bbox": [ 105, 547, 506, 560 ], "score": 1.0, "content": "In Fig. 15 we compare the distribution of token norms for a model with or without registers. This", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 559, 506, 571 ], "spans": [ { "bbox": [ 105, 559, 506, 571 ], "score": 1.0, "content": "figure is similar to Fig. 7 but with a finer granularity, as we also plot the norm distribution of", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 569, 506, 582 ], "spans": [ { "bbox": [ 105, 569, 506, 582 ], "score": 1.0, "content": "individual register tokens and [CLS] tokens. We observe the following: with registers, the norms", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 581, 505, 594 ], "spans": [ { "bbox": [ 105, 581, 505, 594 ], "score": 1.0, "content": "of patch tokens do not contain outliers anymore, and the high-norm tokens are entirely contained in", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 592, 505, 604 ], "spans": [ { "bbox": [ 106, 592, 505, 604 ], "score": 1.0, "content": "the set of registers. As a result, we conclude that the behavior leading to high-norm outliers in the", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 603, 289, 615 ], "spans": [ { "bbox": [ 106, 603, 289, 615 ], "score": 1.0, "content": "model is effectively absorbed in the registers.", "type": "text" } ], "index": 20 } ], "index": 17.5 }, { "type": "text", "bbox": [ 107, 619, 505, 642 ], "lines": [ { "bbox": [ 106, 619, 505, 632 ], "spans": [ { "bbox": [ 106, 619, 505, 632 ], "score": 1.0, "content": "An additional interesting observation is that the norms of the registers appear to be quantized, com-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 630, 481, 643 ], "spans": [ { "bbox": [ 105, 630, 481, 643 ], "score": 1.0, "content": "pared to the previous outliers; we leave the investigation of this phenomenon for future work.", "type": "text" } ], "index": 22 } ], "index": 21.5 }, { "type": "title", "bbox": [ 109, 656, 272, 667 ], "lines": [ { "bbox": [ 106, 655, 273, 668 ], "spans": [ { "bbox": [ 106, 655, 273, 668 ], "score": 1.0, "content": "D.2 INFORMATION HELD BY TOKENS", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 108, 676, 504, 732 ], "lines": [ { "bbox": [ 107, 676, 505, 689 ], "spans": [ { "bbox": [ 107, 676, 505, 689 ], "score": 1.0, "content": "We report on table 4 the linear probing performance of models trained with and without registers,", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 687, 505, 699 ], "spans": [ { "bbox": [ 106, 687, 505, 699 ], "score": 1.0, "content": "when using different tokens as representations. We evaluate on the aircrafts dataset, as it showed", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 698, 505, 712 ], "spans": [ { "bbox": [ 106, 698, 505, 712 ], "score": 1.0, "content": "clear conclusions in the similar table 1. We observe that adding a register does not significantly", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 506, 723 ], "score": 1.0, "content": "modify the scores obtained with the [CLS] or patch tokens. However, the outlier patches are", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 720, 388, 734 ], "spans": [ { "bbox": [ 105, 720, 388, 734 ], "score": 1.0, "content": "removed, and their behavior is transferred to the newly added register.", "type": "text" } ], "index": 28 } ], "index": 26 } ], "page_idx": 13, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 752, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 313, 764 ], "spans": [ { "bbox": [ 299, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 14 } ] } ] }, { "type": "discarded", "bbox": [ 106, 341, 166, 352 ], "lines": [ { "bbox": [ 105, 339, 168, 354 ], "spans": [ { "bbox": [ 105, 339, 168, 354 ], "score": 1.0, "content": "D.1 NORMS", "type": "text" } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 176, 86, 424, 201 ], "blocks": [ { "type": "image_body", "bbox": [ 176, 86, 424, 201 ], "group_id": 0, "lines": [ { "bbox": [ 176, 86, 424, 201 ], "spans": [ { "bbox": [ 176, 86, 424, 201 ], "score": 0.971, "type": "image", "image_path": "407c5844ed49ac0566563545f5eed24d6bde5e34ebdebeaf4112a7e95d8e45c2.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 176, 86, 424, 124.33333333333334 ], "spans": [], "index": 0 }, { "bbox": [ 176, 124.33333333333334, 424, 162.66666666666669 ], "spans": [], "index": 1 }, { "bbox": [ 176, 162.66666666666669, 424, 201.00000000000003 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 214, 505, 259 ], "group_id": 0, "lines": [ { "bbox": [ 106, 214, 506, 227 ], "spans": [ { "bbox": [ 106, 214, 506, 227 ], "score": 1.0, "content": "Figure 14: Illustration of the seed expansion score in LOST for an OpenCLIP model with and", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 225, 506, 238 ], "spans": [ { "bbox": [ 106, 225, 506, 238 ], "score": 1.0, "content": "without registers for the three types of features considered: keys, queries, and values. The score", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 236, 506, 249 ], "spans": [ { "bbox": [ 105, 236, 506, 249 ], "score": 1.0, "content": "is qualitatively improved across all features, with fewer artifacts appearing. Interestingly, the seed", "type": "text" } ], "index": 5 }, { "bbox": [ 104, 247, 471, 260 ], "spans": [ { "bbox": [ 104, 247, 471, 260 ], "score": 1.0, "content": "expansion map computed using values does not exhibit artifacts with nor without registers.", "type": "text" } ], "index": 6 } ], "index": 4.5 } ], "index": 2.75 }, { "type": "title", "bbox": [ 107, 280, 385, 292 ], "lines": [ { "bbox": [ 105, 279, 386, 295 ], "spans": [ { "bbox": [ 105, 279, 386, 295 ], "score": 1.0, "content": "D BEHAVIOR OF MODELS TRAINED WITH REGISTERS", "type": "text" } ], "index": 7 } ], "index": 7 }, { "type": "text", "bbox": [ 106, 305, 506, 327 ], "lines": [ { "bbox": [ 105, 304, 505, 317 ], "spans": [ { "bbox": [ 105, 304, 505, 317 ], "score": 1.0, "content": "In order to better understand the phenomenon at hand, we examine the question of to what extent", "type": "text" } ], "index": 8 }, { "bbox": [ 107, 316, 432, 327 ], "spans": [ { "bbox": [ 107, 316, 432, 327 ], "score": 1.0, "content": "did the register tokens ”replace” the high-norm tokens and took on the same role.", "type": "text" } ], "index": 9 } ], "index": 8.5, "bbox_fs": [ 105, 304, 505, 327 ] }, { "type": "image", "bbox": [ 115, 370, 500, 501 ], "blocks": [ { "type": "image_body", "bbox": [ 115, 370, 500, 501 ], "group_id": 1, "lines": [ { "bbox": [ 115, 370, 500, 501 ], "spans": [ { "bbox": [ 115, 370, 500, 501 ], "score": 0.961, "type": "image", "image_path": "802c3c84876b53f95537280862cacf3d102d0e5feb5e1d103c9b24e5c941af84.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 115, 370, 500, 413.6666666666667 ], "spans": [], "index": 10 }, { "bbox": [ 115, 413.6666666666667, 500, 457.33333333333337 ], "spans": [], "index": 11 }, { "bbox": [ 115, 457.33333333333337, 500, 501.00000000000006 ], "spans": [], "index": 12 } ] }, { "type": "image_caption", "bbox": [ 103, 510, 504, 533 ], "group_id": 1, "lines": [ { "bbox": [ 106, 510, 505, 523 ], "spans": [ { "bbox": [ 106, 510, 505, 523 ], "score": 1.0, "content": "Figure 15: Distribution of token norms for a DINOv2 model without (left) and with (right) 4 regis-", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 522, 468, 534 ], "spans": [ { "bbox": [ 106, 522, 468, 534 ], "score": 1.0, "content": "ters. Introducing registers entirely negates the high-norm outliers among the patch tokens.", "type": "text" } ], "index": 14 } ], "index": 13.5 } ], "index": 12.25 }, { "type": "text", "bbox": [ 107, 547, 505, 614 ], "lines": [ { "bbox": [ 105, 547, 506, 560 ], "spans": [ { "bbox": [ 105, 547, 506, 560 ], "score": 1.0, "content": "In Fig. 15 we compare the distribution of token norms for a model with or without registers. This", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 559, 506, 571 ], "spans": [ { "bbox": [ 105, 559, 506, 571 ], "score": 1.0, "content": "figure is similar to Fig. 7 but with a finer granularity, as we also plot the norm distribution of", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 569, 506, 582 ], "spans": [ { "bbox": [ 105, 569, 506, 582 ], "score": 1.0, "content": "individual register tokens and [CLS] tokens. We observe the following: with registers, the norms", "type": "text" } ], "index": 17 }, { "bbox": [ 105, 581, 505, 594 ], "spans": [ { "bbox": [ 105, 581, 505, 594 ], "score": 1.0, "content": "of patch tokens do not contain outliers anymore, and the high-norm tokens are entirely contained in", "type": "text" } ], "index": 18 }, { "bbox": [ 106, 592, 505, 604 ], "spans": [ { "bbox": [ 106, 592, 505, 604 ], "score": 1.0, "content": "the set of registers. As a result, we conclude that the behavior leading to high-norm outliers in the", "type": "text" } ], "index": 19 }, { "bbox": [ 106, 603, 289, 615 ], "spans": [ { "bbox": [ 106, 603, 289, 615 ], "score": 1.0, "content": "model is effectively absorbed in the registers.", "type": "text" } ], "index": 20 } ], "index": 17.5, "bbox_fs": [ 105, 547, 506, 615 ] }, { "type": "text", "bbox": [ 107, 619, 505, 642 ], "lines": [ { "bbox": [ 106, 619, 505, 632 ], "spans": [ { "bbox": [ 106, 619, 505, 632 ], "score": 1.0, "content": "An additional interesting observation is that the norms of the registers appear to be quantized, com-", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 630, 481, 643 ], "spans": [ { "bbox": [ 105, 630, 481, 643 ], "score": 1.0, "content": "pared to the previous outliers; we leave the investigation of this phenomenon for future work.", "type": "text" } ], "index": 22 } ], "index": 21.5, "bbox_fs": [ 105, 619, 505, 643 ] }, { "type": "title", "bbox": [ 109, 656, 272, 667 ], "lines": [ { "bbox": [ 106, 655, 273, 668 ], "spans": [ { "bbox": [ 106, 655, 273, 668 ], "score": 1.0, "content": "D.2 INFORMATION HELD BY TOKENS", "type": "text" } ], "index": 23 } ], "index": 23 }, { "type": "text", "bbox": [ 108, 676, 504, 732 ], "lines": [ { "bbox": [ 107, 676, 505, 689 ], "spans": [ { "bbox": [ 107, 676, 505, 689 ], "score": 1.0, "content": "We report on table 4 the linear probing performance of models trained with and without registers,", "type": "text" } ], "index": 24 }, { "bbox": [ 106, 687, 505, 699 ], "spans": [ { "bbox": [ 106, 687, 505, 699 ], "score": 1.0, "content": "when using different tokens as representations. We evaluate on the aircrafts dataset, as it showed", "type": "text" } ], "index": 25 }, { "bbox": [ 106, 698, 505, 712 ], "spans": [ { "bbox": [ 106, 698, 505, 712 ], "score": 1.0, "content": "clear conclusions in the similar table 1. We observe that adding a register does not significantly", "type": "text" } ], "index": 26 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 506, 723 ], "score": 1.0, "content": "modify the scores obtained with the [CLS] or patch tokens. However, the outlier patches are", "type": "text" } ], "index": 27 }, { "bbox": [ 105, 720, 388, 734 ], "spans": [ { "bbox": [ 105, 720, 388, 734 ], "score": 1.0, "content": "removed, and their behavior is transferred to the newly added register.", "type": "text" } ], "index": 28 } ], "index": 26, "bbox_fs": [ 105, 676, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "table", "bbox": [ 173, 80, 437, 137 ], "blocks": [ { "type": "table_body", "bbox": [ 173, 80, 437, 137 ], "group_id": 0, "lines": [ { "bbox": [ 173, 80, 437, 137 ], "spans": [ { "bbox": [ 173, 80, 437, 137 ], "score": 0.973, "html": "
top-1 accuracy
#registers[CLS]normal patch outlier patchregister
084.615.573.3N/A
185.214.5N/A71.1
", "type": "table", "image_path": "ffecb4b93eeeca942cd65355beed88df1b6840652f8e255a92be0ab1474003a6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 173, 80, 437, 99.0 ], "spans": [], "index": 0 }, { "bbox": [ 173, 99.0, 437, 118.0 ], "spans": [], "index": 1 }, { "bbox": [ 173, 118.0, 437, 137.0 ], "spans": [], "index": 2 } ] } ], "index": 1 }, { "type": "text", "bbox": [ 109, 145, 503, 179 ], "lines": [ { "bbox": [ 107, 144, 505, 158 ], "spans": [ { "bbox": [ 107, 144, 505, 158 ], "score": 1.0, "content": "Table 4: Linear probing of models with and without registers on the Aircraft dataset, using various", "type": "text" } ], "index": 3 }, { "bbox": [ 107, 156, 505, 169 ], "spans": [ { "bbox": [ 107, 156, 505, 169 ], "score": 1.0, "content": "tokens as representation. We observe that the behavior of the outlier tokens, aggregating global", "type": "text" } ], "index": 4 }, { "bbox": [ 107, 167, 272, 179 ], "spans": [ { "bbox": [ 107, 167, 272, 179 ], "score": 1.0, "content": "information, is absorbed into the register.", "type": "text" } ], "index": 5 } ], "index": 4 }, { "type": "text", "bbox": [ 106, 204, 506, 249 ], "lines": [ { "bbox": [ 106, 205, 505, 215 ], "spans": [ { "bbox": [ 106, 205, 505, 215 ], "score": 1.0, "content": "We further conduct an evaluation of the local information contained in the patch tokens of a model", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 214, 505, 227 ], "spans": [ { "bbox": [ 105, 214, 505, 227 ], "score": 1.0, "content": "trained with and without registers (table 5). We observe that the non-outliers patches, in both cases,", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 226, 505, 238 ], "spans": [ { "bbox": [ 105, 226, 505, 238 ], "score": 1.0, "content": "hold similar local information, confirming that the registers only remove the outlier behavior, with-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 237, 387, 250 ], "spans": [ { "bbox": [ 105, 237, 387, 250 ], "score": 1.0, "content": "out significantly modifying the information held by the other patches.", "type": "text" } ], "index": 9 } ], "index": 7.5 }, { "type": "table", "bbox": [ 156, 262, 455, 320 ], "blocks": [ { "type": "table_body", "bbox": [ 156, 262, 455, 320 ], "group_id": 1, "lines": [ { "bbox": [ 156, 262, 455, 320 ], "spans": [ { "bbox": [ 156, 262, 455, 320 ], "score": 0.966, "html": "
#registerspatches consideredposition prediction top-1 accreconstruction L2 error↓
0non-outliers66.315.9
4non-outliers (ie all)65.816.0
", "type": "table", "image_path": "7af5c7bb21371d59a0e4c96933527f9f27e25e7b362b3f35b2cae0035f5893e4.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 156, 262, 455, 281.3333333333333 ], "spans": [], "index": 10 }, { "bbox": [ 156, 281.3333333333333, 455, 300.66666666666663 ], "spans": [], "index": 11 }, { "bbox": [ 156, 300.66666666666663, 455, 319.99999999999994 ], "spans": [], "index": 12 } ] } ], "index": 11 }, { "type": "text", "bbox": [ 108, 327, 504, 361 ], "lines": [ { "bbox": [ 106, 326, 506, 340 ], "spans": [ { "bbox": [ 106, 326, 506, 340 ], "score": 1.0, "content": "Table 5: Linear probing for local information on the patch tokens of models trained without or", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 338, 505, 351 ], "spans": [ { "bbox": [ 106, 338, 505, 351 ], "score": 1.0, "content": "with registers. We only consider patches considered ”normal”, i.e. not the high-norm outliers. We", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 349, 453, 362 ], "spans": [ { "bbox": [ 106, 349, 453, 362 ], "score": 1.0, "content": "observe that adding registers does not significantly modify the scores of these patches.", "type": "text" } ], "index": 15 } ], "index": 14 }, { "type": "title", "bbox": [ 107, 389, 217, 401 ], "lines": [ { "bbox": [ 106, 389, 218, 402 ], "spans": [ { "bbox": [ 106, 389, 218, 402 ], "score": 1.0, "content": "D.3 POSITIONAL FOCUS", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "image", "bbox": [ 107, 418, 502, 495 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 418, 502, 495 ], "group_id": 0, "lines": [ { "bbox": [ 107, 418, 502, 495 ], "spans": [ { "bbox": [ 107, 418, 502, 495 ], "score": 0.967, "type": "image", "image_path": "fe2c0b900feaa480797388324abd6b7cac85160fdde2328db2482583641c8e32.jpg" } ] } ], "index": 18, "virtual_lines": [ { "bbox": [ 107, 418, 502, 443.6666666666667 ], "spans": [], "index": 17 }, { "bbox": [ 107, 443.6666666666667, 502, 469.33333333333337 ], "spans": [], "index": 18 }, { "bbox": [ 107, 469.33333333333337, 502, 495.00000000000006 ], "spans": [], "index": 19 } ] }, { "type": "image_caption", "bbox": [ 106, 503, 505, 537 ], "group_id": 0, "lines": [ { "bbox": [ 106, 504, 505, 516 ], "spans": [ { "bbox": [ 106, 504, 505, 516 ], "score": 1.0, "content": "Figure 16: Average attention map of registers and [CLS] token. There is a variability observed,", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 515, 505, 527 ], "spans": [ { "bbox": [ 106, 515, 505, 527 ], "score": 1.0, "content": "with register 3 of this model focusing more on border areas. We also include the average attention", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 525, 452, 537 ], "spans": [ { "bbox": [ 105, 525, 452, 537 ], "score": 1.0, "content": "map of a patch for comparison. The patch has a much more focused average attention.", "type": "text" } ], "index": 22 } ], "index": 21 } ], "index": 19.5 }, { "type": "text", "bbox": [ 107, 555, 505, 610 ], "lines": [ { "bbox": [ 105, 554, 505, 569 ], "spans": [ { "bbox": [ 105, 554, 505, 569 ], "score": 1.0, "content": "In Fig. 16 we display the positional focus for the class token and the 4 registers of a DINOv2+reg", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 565, 504, 579 ], "spans": [ { "bbox": [ 106, 565, 504, 579 ], "score": 1.0, "content": "model. We produce these plots by running the model on a random subset of ImageNet-22k, and av-", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 577, 504, 590 ], "spans": [ { "bbox": [ 105, 577, 487, 590 ], "score": 1.0, "content": "eraging the attention maps for the corresponding tokens at the last layer. We note that ImageNet-", "type": "text" }, { "bbox": [ 488, 577, 504, 588 ], "score": 0.36, "content": "2 2 \\mathrm { k }", "type": "inline_equation" } ], "index": 25 }, { "bbox": [ 105, 588, 505, 602 ], "spans": [ { "bbox": [ 105, 588, 505, 602 ], "score": 1.0, "content": "contains mostly object-centric images rather than scenes, which explains why the average attention", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 600, 249, 612 ], "spans": [ { "bbox": [ 106, 600, 249, 612 ], "score": 1.0, "content": "maps correspond to centered blobs.", "type": "text" } ], "index": 27 } ], "index": 25 }, { "type": "text", "bbox": [ 108, 615, 504, 671 ], "lines": [ { "bbox": [ 107, 616, 506, 628 ], "spans": [ { "bbox": [ 107, 616, 506, 628 ], "score": 1.0, "content": "We make several observations. First, the attention maps for registers can be different of each other;", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 626, 506, 640 ], "spans": [ { "bbox": [ 105, 626, 506, 640 ], "score": 1.0, "content": "for example, register 3 tends to focus on border areas, while the other registers tend to focus on more", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 638, 506, 650 ], "spans": [ { "bbox": [ 105, 638, 506, 650 ], "score": 1.0, "content": "centered areas. Register 2 tends to focus slightly more on the upper areas of images that others. This", "type": "text" } ], "index": 30 }, { "bbox": [ 104, 647, 506, 663 ], "spans": [ { "bbox": [ 104, 647, 506, 663 ], "score": 1.0, "content": "is consistent with Fig. 9, where we show registers focusing on different large areas of the image,", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 660, 267, 672 ], "spans": [ { "bbox": [ 105, 660, 267, 672 ], "score": 1.0, "content": "suggesting some level of specialization.", "type": "text" } ], "index": 32 } ], "index": 30 }, { "type": "text", "bbox": [ 107, 677, 505, 732 ], "lines": [ { "bbox": [ 106, 676, 505, 690 ], "spans": [ { "bbox": [ 106, 676, 505, 690 ], "score": 1.0, "content": "Second, by comparing the register maps to the [CLS] token map and to a patch token map, we", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 687, 505, 700 ], "spans": [ { "bbox": [ 105, 687, 505, 700 ], "score": 1.0, "content": "observe that registers produce maps with a large support area, very similarly to the [CLS] token,", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 506, 712 ], "score": 1.0, "content": "and very different of a typical patch token which is more localized. As the [CLS] token is known to", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 506, 723 ], "score": 1.0, "content": "carry global information (as proven by the linear probing classification performance): this suggests", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 721, 280, 733 ], "spans": [ { "bbox": [ 106, 721, 280, 733 ], "score": 1.0, "content": "that registers also carry global information.", "type": "text" } ], "index": 37 } ], "index": 35 } ], "page_idx": 14, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 752, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 765 ], "spans": [ { "bbox": [ 299, 750, 312, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "table", "bbox": [ 173, 80, 437, 137 ], "blocks": [ { "type": "table_body", "bbox": [ 173, 80, 437, 137 ], "group_id": 0, "lines": [ { "bbox": [ 173, 80, 437, 137 ], "spans": [ { "bbox": [ 173, 80, 437, 137 ], "score": 0.973, "html": "
top-1 accuracy
#registers[CLS]normal patch outlier patchregister
084.615.573.3N/A
185.214.5N/A71.1
", "type": "table", "image_path": "ffecb4b93eeeca942cd65355beed88df1b6840652f8e255a92be0ab1474003a6.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 173, 80, 437, 99.0 ], "spans": [], "index": 0 }, { "bbox": [ 173, 99.0, 437, 118.0 ], "spans": [], "index": 1 }, { "bbox": [ 173, 118.0, 437, 137.0 ], "spans": [], "index": 2 } ] } ], "index": 1 }, { "type": "text", "bbox": [ 109, 145, 503, 179 ], "lines": [ { "bbox": [ 107, 144, 505, 158 ], "spans": [ { "bbox": [ 107, 144, 505, 158 ], "score": 1.0, "content": "Table 4: Linear probing of models with and without registers on the Aircraft dataset, using various", "type": "text" } ], "index": 3 }, { "bbox": [ 107, 156, 505, 169 ], "spans": [ { "bbox": [ 107, 156, 505, 169 ], "score": 1.0, "content": "tokens as representation. We observe that the behavior of the outlier tokens, aggregating global", "type": "text" } ], "index": 4 }, { "bbox": [ 107, 167, 272, 179 ], "spans": [ { "bbox": [ 107, 167, 272, 179 ], "score": 1.0, "content": "information, is absorbed into the register.", "type": "text" } ], "index": 5 } ], "index": 4, "bbox_fs": [ 107, 144, 505, 179 ] }, { "type": "text", "bbox": [ 106, 204, 506, 249 ], "lines": [ { "bbox": [ 106, 205, 505, 215 ], "spans": [ { "bbox": [ 106, 205, 505, 215 ], "score": 1.0, "content": "We further conduct an evaluation of the local information contained in the patch tokens of a model", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 214, 505, 227 ], "spans": [ { "bbox": [ 105, 214, 505, 227 ], "score": 1.0, "content": "trained with and without registers (table 5). We observe that the non-outliers patches, in both cases,", "type": "text" } ], "index": 7 }, { "bbox": [ 105, 226, 505, 238 ], "spans": [ { "bbox": [ 105, 226, 505, 238 ], "score": 1.0, "content": "hold similar local information, confirming that the registers only remove the outlier behavior, with-", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 237, 387, 250 ], "spans": [ { "bbox": [ 105, 237, 387, 250 ], "score": 1.0, "content": "out significantly modifying the information held by the other patches.", "type": "text" } ], "index": 9 } ], "index": 7.5, "bbox_fs": [ 105, 205, 505, 250 ] }, { "type": "table", "bbox": [ 156, 262, 455, 320 ], "blocks": [ { "type": "table_body", "bbox": [ 156, 262, 455, 320 ], "group_id": 1, "lines": [ { "bbox": [ 156, 262, 455, 320 ], "spans": [ { "bbox": [ 156, 262, 455, 320 ], "score": 0.966, "html": "
#registerspatches consideredposition prediction top-1 accreconstruction L2 error↓
0non-outliers66.315.9
4non-outliers (ie all)65.816.0
", "type": "table", "image_path": "7af5c7bb21371d59a0e4c96933527f9f27e25e7b362b3f35b2cae0035f5893e4.jpg" } ] } ], "index": 11, "virtual_lines": [ { "bbox": [ 156, 262, 455, 281.3333333333333 ], "spans": [], "index": 10 }, { "bbox": [ 156, 281.3333333333333, 455, 300.66666666666663 ], "spans": [], "index": 11 }, { "bbox": [ 156, 300.66666666666663, 455, 319.99999999999994 ], "spans": [], "index": 12 } ] } ], "index": 11 }, { "type": "text", "bbox": [ 108, 327, 504, 361 ], "lines": [ { "bbox": [ 106, 326, 506, 340 ], "spans": [ { "bbox": [ 106, 326, 506, 340 ], "score": 1.0, "content": "Table 5: Linear probing for local information on the patch tokens of models trained without or", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 338, 505, 351 ], "spans": [ { "bbox": [ 106, 338, 505, 351 ], "score": 1.0, "content": "with registers. We only consider patches considered ”normal”, i.e. not the high-norm outliers. We", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 349, 453, 362 ], "spans": [ { "bbox": [ 106, 349, 453, 362 ], "score": 1.0, "content": "observe that adding registers does not significantly modify the scores of these patches.", "type": "text" } ], "index": 15 } ], "index": 14, "bbox_fs": [ 106, 326, 506, 362 ] }, { "type": "title", "bbox": [ 107, 389, 217, 401 ], "lines": [ { "bbox": [ 106, 389, 218, 402 ], "spans": [ { "bbox": [ 106, 389, 218, 402 ], "score": 1.0, "content": "D.3 POSITIONAL FOCUS", "type": "text" } ], "index": 16 } ], "index": 16 }, { "type": "image", "bbox": [ 107, 418, 502, 495 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 418, 502, 495 ], "group_id": 0, "lines": [ { "bbox": [ 107, 418, 502, 495 ], "spans": [ { "bbox": [ 107, 418, 502, 495 ], "score": 0.967, "type": "image", "image_path": "fe2c0b900feaa480797388324abd6b7cac85160fdde2328db2482583641c8e32.jpg" } ] } ], "index": 18, "virtual_lines": [ { "bbox": [ 107, 418, 502, 443.6666666666667 ], "spans": [], "index": 17 }, { "bbox": [ 107, 443.6666666666667, 502, 469.33333333333337 ], "spans": [], "index": 18 }, { "bbox": [ 107, 469.33333333333337, 502, 495.00000000000006 ], "spans": [], "index": 19 } ] }, { "type": "image_caption", "bbox": [ 106, 503, 505, 537 ], "group_id": 0, "lines": [ { "bbox": [ 106, 504, 505, 516 ], "spans": [ { "bbox": [ 106, 504, 505, 516 ], "score": 1.0, "content": "Figure 16: Average attention map of registers and [CLS] token. There is a variability observed,", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 515, 505, 527 ], "spans": [ { "bbox": [ 106, 515, 505, 527 ], "score": 1.0, "content": "with register 3 of this model focusing more on border areas. We also include the average attention", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 525, 452, 537 ], "spans": [ { "bbox": [ 105, 525, 452, 537 ], "score": 1.0, "content": "map of a patch for comparison. The patch has a much more focused average attention.", "type": "text" } ], "index": 22 } ], "index": 21 } ], "index": 19.5 }, { "type": "text", "bbox": [ 107, 555, 505, 610 ], "lines": [ { "bbox": [ 105, 554, 505, 569 ], "spans": [ { "bbox": [ 105, 554, 505, 569 ], "score": 1.0, "content": "In Fig. 16 we display the positional focus for the class token and the 4 registers of a DINOv2+reg", "type": "text" } ], "index": 23 }, { "bbox": [ 106, 565, 504, 579 ], "spans": [ { "bbox": [ 106, 565, 504, 579 ], "score": 1.0, "content": "model. We produce these plots by running the model on a random subset of ImageNet-22k, and av-", "type": "text" } ], "index": 24 }, { "bbox": [ 105, 577, 504, 590 ], "spans": [ { "bbox": [ 105, 577, 487, 590 ], "score": 1.0, "content": "eraging the attention maps for the corresponding tokens at the last layer. We note that ImageNet-", "type": "text" }, { "bbox": [ 488, 577, 504, 588 ], "score": 0.36, "content": "2 2 \\mathrm { k }", "type": "inline_equation" } ], "index": 25 }, { "bbox": [ 105, 588, 505, 602 ], "spans": [ { "bbox": [ 105, 588, 505, 602 ], "score": 1.0, "content": "contains mostly object-centric images rather than scenes, which explains why the average attention", "type": "text" } ], "index": 26 }, { "bbox": [ 106, 600, 249, 612 ], "spans": [ { "bbox": [ 106, 600, 249, 612 ], "score": 1.0, "content": "maps correspond to centered blobs.", "type": "text" } ], "index": 27 } ], "index": 25, "bbox_fs": [ 105, 554, 505, 612 ] }, { "type": "text", "bbox": [ 108, 615, 504, 671 ], "lines": [ { "bbox": [ 107, 616, 506, 628 ], "spans": [ { "bbox": [ 107, 616, 506, 628 ], "score": 1.0, "content": "We make several observations. First, the attention maps for registers can be different of each other;", "type": "text" } ], "index": 28 }, { "bbox": [ 105, 626, 506, 640 ], "spans": [ { "bbox": [ 105, 626, 506, 640 ], "score": 1.0, "content": "for example, register 3 tends to focus on border areas, while the other registers tend to focus on more", "type": "text" } ], "index": 29 }, { "bbox": [ 105, 638, 506, 650 ], "spans": [ { "bbox": [ 105, 638, 506, 650 ], "score": 1.0, "content": "centered areas. Register 2 tends to focus slightly more on the upper areas of images that others. This", "type": "text" } ], "index": 30 }, { "bbox": [ 104, 647, 506, 663 ], "spans": [ { "bbox": [ 104, 647, 506, 663 ], "score": 1.0, "content": "is consistent with Fig. 9, where we show registers focusing on different large areas of the image,", "type": "text" } ], "index": 31 }, { "bbox": [ 105, 660, 267, 672 ], "spans": [ { "bbox": [ 105, 660, 267, 672 ], "score": 1.0, "content": "suggesting some level of specialization.", "type": "text" } ], "index": 32 } ], "index": 30, "bbox_fs": [ 104, 616, 506, 672 ] }, { "type": "text", "bbox": [ 107, 677, 505, 732 ], "lines": [ { "bbox": [ 106, 676, 505, 690 ], "spans": [ { "bbox": [ 106, 676, 505, 690 ], "score": 1.0, "content": "Second, by comparing the register maps to the [CLS] token map and to a patch token map, we", "type": "text" } ], "index": 33 }, { "bbox": [ 105, 687, 505, 700 ], "spans": [ { "bbox": [ 105, 687, 505, 700 ], "score": 1.0, "content": "observe that registers produce maps with a large support area, very similarly to the [CLS] token,", "type": "text" } ], "index": 34 }, { "bbox": [ 105, 699, 506, 712 ], "spans": [ { "bbox": [ 105, 699, 506, 712 ], "score": 1.0, "content": "and very different of a typical patch token which is more localized. As the [CLS] token is known to", "type": "text" } ], "index": 35 }, { "bbox": [ 105, 709, 506, 723 ], "spans": [ { "bbox": [ 105, 709, 506, 723 ], "score": 1.0, "content": "carry global information (as proven by the linear probing classification performance): this suggests", "type": "text" } ], "index": 36 }, { "bbox": [ 106, 721, 280, 733 ], "spans": [ { "bbox": [ 106, 721, 280, 733 ], "score": 1.0, "content": "that registers also carry global information.", "type": "text" } ], "index": 37 } ], "index": 35, "bbox_fs": [ 105, 676, 506, 733 ] } ] }, { "preproc_blocks": [ { "type": "title", "bbox": [ 108, 81, 260, 93 ], "lines": [ { "bbox": [ 105, 80, 262, 96 ], "spans": [ { "bbox": [ 105, 80, 262, 96 ], "score": 1.0, "content": "E MASKED AUTOENCODERS", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 109, 504, 186 ], "lines": [ { "bbox": [ 105, 109, 506, 122 ], "spans": [ { "bbox": [ 105, 109, 506, 122 ], "score": 1.0, "content": "Masked Autoencoding (He et al., 2022) is another common way of pretraining self-supervised mod-", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 119, 506, 133 ], "spans": [ { "bbox": [ 105, 119, 506, 133 ], "score": 1.0, "content": "els. We observe in Fig. 17 that there are no artifacts in the maps produced by MAE: our hypothesis", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 130, 506, 144 ], "spans": [ { "bbox": [ 105, 130, 506, 144 ], "score": 1.0, "content": "is that the absence of artifacts is due to the training procedure using only a local loss on the patch", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 142, 506, 155 ], "spans": [ { "bbox": [ 105, 142, 506, 155 ], "score": 1.0, "content": "tokens, rather than an objective involving global aggregation of information. However, we also note", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 153, 504, 165 ], "spans": [ { "bbox": [ 105, 153, 484, 165 ], "score": 1.0, "content": "that the performance of MAE models is very low for self-supervised representation learning", "type": "text" }, { "bbox": [ 484, 153, 504, 164 ], "score": 0.83, "content": "7 5 \\%", "type": "inline_equation" } ], "index": 5 }, { "bbox": [ 105, 164, 506, 177 ], "spans": [ { "bbox": [ 105, 164, 506, 177 ], "score": 1.0, "content": "linear probing performance on ImageNet classification for ViT-Large), preventing it from being used", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 175, 284, 188 ], "spans": [ { "bbox": [ 105, 175, 284, 188 ], "score": 1.0, "content": "as is, and making fine-tuning a requirement.", "type": "text" } ], "index": 7 } ], "index": 4 }, { "type": "image", "bbox": [ 107, 202, 504, 491 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 202, 504, 491 ], "group_id": 0, "lines": [ { "bbox": [ 107, 202, 504, 491 ], "spans": [ { "bbox": [ 107, 202, 504, 491 ], "score": 0.976, "type": "image", "image_path": "11b4c5ecd47d8eff24f3af3ce8a4f58e410c5af79cc174f3052e0a3b40af9965.jpg" } ] } ], "index": 9, "virtual_lines": [ { "bbox": [ 107, 202, 504, 298.3333333333333 ], "spans": [], "index": 8 }, { "bbox": [ 107, 298.3333333333333, 504, 394.66666666666663 ], "spans": [], "index": 9 }, { "bbox": [ 107, 394.66666666666663, 504, 490.99999999999994 ], "spans": [], "index": 10 } ] }, { "type": "image_caption", "bbox": [ 103, 499, 504, 522 ], "group_id": 0, "lines": [ { "bbox": [ 105, 498, 505, 513 ], "spans": [ { "bbox": [ 105, 498, 505, 513 ], "score": 1.0, "content": "Figure 17: First three principal components of the output feature map of a ViT-Large Masked Au-", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 510, 151, 522 ], "spans": [ { "bbox": [ 105, 510, 151, 522 ], "score": 1.0, "content": "toencoder.", "type": "text" } ], "index": 12 } ], "index": 11.5 } ], "index": 10.25 }, { "type": "title", "bbox": [ 108, 555, 297, 567 ], "lines": [ { "bbox": [ 105, 554, 298, 569 ], "spans": [ { "bbox": [ 105, 554, 298, 569 ], "score": 1.0, "content": "F BEHAVIOR PER ATTENTION HEAD", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 583, 505, 638 ], "lines": [ { "bbox": [ 105, 583, 505, 595 ], "spans": [ { "bbox": [ 105, 583, 505, 595 ], "score": 1.0, "content": "In this section, we investigate whether the artifacts appear only on the attention maps for specific", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 592, 505, 608 ], "spans": [ { "bbox": [ 105, 592, 505, 608 ], "score": 1.0, "content": "heads of the last vision transformer block, or for all of them. We show in Fig. 18 the input image", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 604, 505, 618 ], "spans": [ { "bbox": [ 105, 604, 505, 618 ], "score": 1.0, "content": "along with the attention maps for different heads. We observe that the artifacts appear for all atten-", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 616, 505, 628 ], "spans": [ { "bbox": [ 106, 616, 505, 628 ], "score": 1.0, "content": "tion heads, despite heads focusing on different areas of the object. We still observe that some heads", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 627, 249, 639 ], "spans": [ { "bbox": [ 106, 627, 249, 639 ], "score": 1.0, "content": "focus more on artifacts than others.", "type": "text" } ], "index": 18 } ], "index": 16 }, { "type": "title", "bbox": [ 107, 660, 364, 672 ], "lines": [ { "bbox": [ 105, 658, 365, 676 ], "spans": [ { "bbox": [ 105, 658, 365, 676 ], "score": 1.0, "content": "G VARIANCE ON TOKEN INFORMATION PROBING", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 107, 687, 505, 732 ], "lines": [ { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "The results presented in table 1 are obtained by taking a random patch token, either normal or", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 700, 505, 710 ], "spans": [ { "bbox": [ 106, 700, 505, 710 ], "score": 1.0, "content": "outlier. However, the choice of this token adds a significant source of variance in the evaluation.", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "For thoroughness, we report in table 6 the standard deviation of the scores obtained relative to this", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 720, 138, 734 ], "spans": [ { "bbox": [ 105, 720, 138, 734 ], "score": 1.0, "content": "choice.", "type": "text" } ], "index": 23 } ], "index": 21.5 } ], "page_idx": 15, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 752, 311, 760 ], "lines": [] } ], "para_blocks": [ { "type": "title", "bbox": [ 108, 81, 260, 93 ], "lines": [ { "bbox": [ 105, 80, 262, 96 ], "spans": [ { "bbox": [ 105, 80, 262, 96 ], "score": 1.0, "content": "E MASKED AUTOENCODERS", "type": "text" } ], "index": 0 } ], "index": 0 }, { "type": "text", "bbox": [ 107, 109, 504, 186 ], "lines": [ { "bbox": [ 105, 109, 506, 122 ], "spans": [ { "bbox": [ 105, 109, 506, 122 ], "score": 1.0, "content": "Masked Autoencoding (He et al., 2022) is another common way of pretraining self-supervised mod-", "type": "text" } ], "index": 1 }, { "bbox": [ 105, 119, 506, 133 ], "spans": [ { "bbox": [ 105, 119, 506, 133 ], "score": 1.0, "content": "els. We observe in Fig. 17 that there are no artifacts in the maps produced by MAE: our hypothesis", "type": "text" } ], "index": 2 }, { "bbox": [ 105, 130, 506, 144 ], "spans": [ { "bbox": [ 105, 130, 506, 144 ], "score": 1.0, "content": "is that the absence of artifacts is due to the training procedure using only a local loss on the patch", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 142, 506, 155 ], "spans": [ { "bbox": [ 105, 142, 506, 155 ], "score": 1.0, "content": "tokens, rather than an objective involving global aggregation of information. However, we also note", "type": "text" } ], "index": 4 }, { "bbox": [ 105, 153, 504, 165 ], "spans": [ { "bbox": [ 105, 153, 484, 165 ], "score": 1.0, "content": "that the performance of MAE models is very low for self-supervised representation learning", "type": "text" }, { "bbox": [ 484, 153, 504, 164 ], "score": 0.83, "content": "7 5 \\%", "type": "inline_equation" } ], "index": 5 }, { "bbox": [ 105, 164, 506, 177 ], "spans": [ { "bbox": [ 105, 164, 506, 177 ], "score": 1.0, "content": "linear probing performance on ImageNet classification for ViT-Large), preventing it from being used", "type": "text" } ], "index": 6 }, { "bbox": [ 105, 175, 284, 188 ], "spans": [ { "bbox": [ 105, 175, 284, 188 ], "score": 1.0, "content": "as is, and making fine-tuning a requirement.", "type": "text" } ], "index": 7 } ], "index": 4, "bbox_fs": [ 105, 109, 506, 188 ] }, { "type": "image", "bbox": [ 107, 202, 504, 491 ], "blocks": [ { "type": "image_body", "bbox": [ 107, 202, 504, 491 ], "group_id": 0, "lines": [ { "bbox": [ 107, 202, 504, 491 ], "spans": [ { "bbox": [ 107, 202, 504, 491 ], "score": 0.976, "type": "image", "image_path": "11b4c5ecd47d8eff24f3af3ce8a4f58e410c5af79cc174f3052e0a3b40af9965.jpg" } ] } ], "index": 9, "virtual_lines": [ { "bbox": [ 107, 202, 504, 298.3333333333333 ], "spans": [], "index": 8 }, { "bbox": [ 107, 298.3333333333333, 504, 394.66666666666663 ], "spans": [], "index": 9 }, { "bbox": [ 107, 394.66666666666663, 504, 490.99999999999994 ], "spans": [], "index": 10 } ] }, { "type": "image_caption", "bbox": [ 103, 499, 504, 522 ], "group_id": 0, "lines": [ { "bbox": [ 105, 498, 505, 513 ], "spans": [ { "bbox": [ 105, 498, 505, 513 ], "score": 1.0, "content": "Figure 17: First three principal components of the output feature map of a ViT-Large Masked Au-", "type": "text" } ], "index": 11 }, { "bbox": [ 105, 510, 151, 522 ], "spans": [ { "bbox": [ 105, 510, 151, 522 ], "score": 1.0, "content": "toencoder.", "type": "text" } ], "index": 12 } ], "index": 11.5 } ], "index": 10.25 }, { "type": "title", "bbox": [ 108, 555, 297, 567 ], "lines": [ { "bbox": [ 105, 554, 298, 569 ], "spans": [ { "bbox": [ 105, 554, 298, 569 ], "score": 1.0, "content": "F BEHAVIOR PER ATTENTION HEAD", "type": "text" } ], "index": 13 } ], "index": 13 }, { "type": "text", "bbox": [ 107, 583, 505, 638 ], "lines": [ { "bbox": [ 105, 583, 505, 595 ], "spans": [ { "bbox": [ 105, 583, 505, 595 ], "score": 1.0, "content": "In this section, we investigate whether the artifacts appear only on the attention maps for specific", "type": "text" } ], "index": 14 }, { "bbox": [ 105, 592, 505, 608 ], "spans": [ { "bbox": [ 105, 592, 505, 608 ], "score": 1.0, "content": "heads of the last vision transformer block, or for all of them. We show in Fig. 18 the input image", "type": "text" } ], "index": 15 }, { "bbox": [ 105, 604, 505, 618 ], "spans": [ { "bbox": [ 105, 604, 505, 618 ], "score": 1.0, "content": "along with the attention maps for different heads. We observe that the artifacts appear for all atten-", "type": "text" } ], "index": 16 }, { "bbox": [ 106, 616, 505, 628 ], "spans": [ { "bbox": [ 106, 616, 505, 628 ], "score": 1.0, "content": "tion heads, despite heads focusing on different areas of the object. We still observe that some heads", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 627, 249, 639 ], "spans": [ { "bbox": [ 106, 627, 249, 639 ], "score": 1.0, "content": "focus more on artifacts than others.", "type": "text" } ], "index": 18 } ], "index": 16, "bbox_fs": [ 105, 583, 505, 639 ] }, { "type": "title", "bbox": [ 107, 660, 364, 672 ], "lines": [ { "bbox": [ 105, 658, 365, 676 ], "spans": [ { "bbox": [ 105, 658, 365, 676 ], "score": 1.0, "content": "G VARIANCE ON TOKEN INFORMATION PROBING", "type": "text" } ], "index": 19 } ], "index": 19 }, { "type": "text", "bbox": [ 107, 687, 505, 732 ], "lines": [ { "bbox": [ 105, 687, 506, 700 ], "spans": [ { "bbox": [ 105, 687, 506, 700 ], "score": 1.0, "content": "The results presented in table 1 are obtained by taking a random patch token, either normal or", "type": "text" } ], "index": 20 }, { "bbox": [ 106, 700, 505, 710 ], "spans": [ { "bbox": [ 106, 700, 505, 710 ], "score": 1.0, "content": "outlier. However, the choice of this token adds a significant source of variance in the evaluation.", "type": "text" } ], "index": 21 }, { "bbox": [ 105, 709, 505, 722 ], "spans": [ { "bbox": [ 105, 709, 505, 722 ], "score": 1.0, "content": "For thoroughness, we report in table 6 the standard deviation of the scores obtained relative to this", "type": "text" } ], "index": 22 }, { "bbox": [ 105, 720, 138, 734 ], "spans": [ { "bbox": [ 105, 720, 138, 734 ], "score": 1.0, "content": "choice.", "type": "text" } ], "index": 23 } ], "index": 21.5, "bbox_fs": [ 105, 687, 506, 734 ] } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 114, 80, 505, 391 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 80, 505, 391 ], "group_id": 0, "lines": [ { "bbox": [ 114, 80, 505, 391 ], "spans": [ { "bbox": [ 114, 80, 505, 391 ], "score": 0.963, "type": "image", "image_path": "974451da8bebe1ebfdfc59f3b4ad6e0e08789bd3881702dd120ac596c5986c07.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 114, 80, 505, 183.66666666666669 ], "spans": [], "index": 0 }, { "bbox": [ 114, 183.66666666666669, 505, 287.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 114, 287.33333333333337, 505, 391.00000000000006 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 105, 399, 506, 423 ], "group_id": 0, "lines": [ { "bbox": [ 106, 399, 505, 413 ], "spans": [ { "bbox": [ 106, 399, 505, 413 ], "score": 1.0, "content": "Figure 18: Attention maps of the [CLS] token to the patch tokens, shown here separately per", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 411, 468, 423 ], "spans": [ { "bbox": [ 106, 411, 468, 423 ], "score": 1.0, "content": "attention head. We produce these maps with a DINOv2-L model trained without registers.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "table", "bbox": [ 141, 436, 470, 573 ], "blocks": [ { "type": "table_body", "bbox": [ 141, 436, 470, 573 ], "group_id": 0, "lines": [ { "bbox": [ 141, 436, 470, 573 ], "spans": [ { "bbox": [ 141, 436, 470, 573 ], "score": 0.972, "html": "
dataset tokenAirc.CF10CF100CUBCal101CarsDTD
normal outlier17.1±0.5 79.1±0.597.1±0.1 99.3±0.081.3±0.3 93.7±0.318.6±0.6 84.9±2.173.2±1.3 97.6±0.710.8±0.3 85.2±0.963.1±0.8 84.9±0.9
[CLS] dataset87.3 Flow.99.4 Food94.5 IN1k91.3 P20596.9 Pets91.5 SUN85.2 VOC
token normal47.8±0.5
outlier59.5±1.2 99.6±0.074.2±0.3 93.5±0.265.8±0.1 69.0±0.753.1±0.3 55.1±1.094.1±0.237.7±0.3 78.5±0.270.8±0.5 89.7±0.1
[CLS]99.794.786.066.496.978.689.1
", "type": "table", "image_path": "e995635cd0131453b4e3ced98d3bdd8a331e7b3eab66af7019673e1fab8175cf.jpg" } ] } ], "index": 6, "virtual_lines": [ { "bbox": [ 141, 436, 470, 481.6666666666667 ], "spans": [], "index": 5 }, { "bbox": [ 141, 481.6666666666667, 470, 527.3333333333334 ], "spans": [], "index": 6 }, { "bbox": [ 141, 527.3333333333334, 470, 573.0 ], "spans": [], "index": 7 } ] } ], "index": 6 }, { "type": "text", "bbox": [ 108, 581, 504, 625 ], "lines": [ { "bbox": [ 106, 581, 505, 593 ], "spans": [ { "bbox": [ 106, 581, 505, 593 ], "score": 1.0, "content": "Table 6: Image classification via linear probing on normal and outlier patch tokens. As we select the", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 592, 506, 605 ], "spans": [ { "bbox": [ 105, 592, 506, 605 ], "score": 1.0, "content": "patch tokens randomly among the set of eligible tokens, this adds a source of variability. We report", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 603, 506, 616 ], "spans": [ { "bbox": [ 105, 603, 506, 616 ], "score": 1.0, "content": "the standard deviation of this variability in grey along with the scores. This table is a detailed view", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 614, 150, 625 ], "spans": [ { "bbox": [ 106, 614, 150, 625 ], "score": 1.0, "content": "of table 1.", "type": "text" } ], "index": 11 } ], "index": 9.5 }, { "type": "title", "bbox": [ 107, 649, 249, 663 ], "lines": [ { "bbox": [ 105, 649, 250, 665 ], "spans": [ { "bbox": [ 105, 649, 250, 665 ], "score": 1.0, "content": "H QUALITATIVE RESULTS", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 108, 676, 504, 732 ], "lines": [ { "bbox": [ 106, 676, 506, 690 ], "spans": [ { "bbox": [ 106, 676, 506, 690 ], "score": 1.0, "content": "We trained three popular models: DeiT-III, OpenCLIP, DINOv2 with and without the introduction", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 687, 505, 701 ], "spans": [ { "bbox": [ 106, 687, 505, 701 ], "score": 1.0, "content": "of register tokens. We observe in Fig. 19 the attention maps in the last layer of the Vision Trans-", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 505, 711 ], "score": 1.0, "content": "former, for all three cases. We see that our approach provides much cleaner attention maps, with", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 709, 506, 723 ], "spans": [ { "bbox": [ 106, 709, 506, 723 ], "score": 1.0, "content": "considerably fewer artifacts, explaining the improvement on the downstream object discovery task", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 719, 506, 735 ], "spans": [ { "bbox": [ 105, 719, 506, 735 ], "score": 1.0, "content": "mentioned in Sec. 3.3. The feature maps are also visibly improved, as shown in Fig. 20. Finally,", "type": "text" } ], "index": 17 } ], "index": 15 } ], "page_idx": 16, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 752, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 765 ], "spans": [ { "bbox": [ 299, 750, 312, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 114, 80, 505, 391 ], "blocks": [ { "type": "image_body", "bbox": [ 114, 80, 505, 391 ], "group_id": 0, "lines": [ { "bbox": [ 114, 80, 505, 391 ], "spans": [ { "bbox": [ 114, 80, 505, 391 ], "score": 0.963, "type": "image", "image_path": "974451da8bebe1ebfdfc59f3b4ad6e0e08789bd3881702dd120ac596c5986c07.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 114, 80, 505, 183.66666666666669 ], "spans": [], "index": 0 }, { "bbox": [ 114, 183.66666666666669, 505, 287.33333333333337 ], "spans": [], "index": 1 }, { "bbox": [ 114, 287.33333333333337, 505, 391.00000000000006 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 105, 399, 506, 423 ], "group_id": 0, "lines": [ { "bbox": [ 106, 399, 505, 413 ], "spans": [ { "bbox": [ 106, 399, 505, 413 ], "score": 1.0, "content": "Figure 18: Attention maps of the [CLS] token to the patch tokens, shown here separately per", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 411, 468, 423 ], "spans": [ { "bbox": [ 106, 411, 468, 423 ], "score": 1.0, "content": "attention head. We produce these maps with a DINOv2-L model trained without registers.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 }, { "type": "table", "bbox": [ 141, 436, 470, 573 ], "blocks": [ { "type": "table_body", "bbox": [ 141, 436, 470, 573 ], "group_id": 0, "lines": [ { "bbox": [ 141, 436, 470, 573 ], "spans": [ { "bbox": [ 141, 436, 470, 573 ], "score": 0.972, "html": "
dataset tokenAirc.CF10CF100CUBCal101CarsDTD
normal outlier17.1±0.5 79.1±0.597.1±0.1 99.3±0.081.3±0.3 93.7±0.318.6±0.6 84.9±2.173.2±1.3 97.6±0.710.8±0.3 85.2±0.963.1±0.8 84.9±0.9
[CLS] dataset87.3 Flow.99.4 Food94.5 IN1k91.3 P20596.9 Pets91.5 SUN85.2 VOC
token normal47.8±0.5
outlier59.5±1.2 99.6±0.074.2±0.3 93.5±0.265.8±0.1 69.0±0.753.1±0.3 55.1±1.094.1±0.237.7±0.3 78.5±0.270.8±0.5 89.7±0.1
[CLS]99.794.786.066.496.978.689.1
", "type": "table", "image_path": "e995635cd0131453b4e3ced98d3bdd8a331e7b3eab66af7019673e1fab8175cf.jpg" } ] } ], "index": 6, "virtual_lines": [ { "bbox": [ 141, 436, 470, 481.6666666666667 ], "spans": [], "index": 5 }, { "bbox": [ 141, 481.6666666666667, 470, 527.3333333333334 ], "spans": [], "index": 6 }, { "bbox": [ 141, 527.3333333333334, 470, 573.0 ], "spans": [], "index": 7 } ] } ], "index": 6 }, { "type": "text", "bbox": [ 108, 581, 504, 625 ], "lines": [ { "bbox": [ 106, 581, 505, 593 ], "spans": [ { "bbox": [ 106, 581, 505, 593 ], "score": 1.0, "content": "Table 6: Image classification via linear probing on normal and outlier patch tokens. As we select the", "type": "text" } ], "index": 8 }, { "bbox": [ 105, 592, 506, 605 ], "spans": [ { "bbox": [ 105, 592, 506, 605 ], "score": 1.0, "content": "patch tokens randomly among the set of eligible tokens, this adds a source of variability. We report", "type": "text" } ], "index": 9 }, { "bbox": [ 105, 603, 506, 616 ], "spans": [ { "bbox": [ 105, 603, 506, 616 ], "score": 1.0, "content": "the standard deviation of this variability in grey along with the scores. This table is a detailed view", "type": "text" } ], "index": 10 }, { "bbox": [ 106, 614, 150, 625 ], "spans": [ { "bbox": [ 106, 614, 150, 625 ], "score": 1.0, "content": "of table 1.", "type": "text" } ], "index": 11 } ], "index": 9.5, "bbox_fs": [ 105, 581, 506, 625 ] }, { "type": "title", "bbox": [ 107, 649, 249, 663 ], "lines": [ { "bbox": [ 105, 649, 250, 665 ], "spans": [ { "bbox": [ 105, 649, 250, 665 ], "score": 1.0, "content": "H QUALITATIVE RESULTS", "type": "text" } ], "index": 12 } ], "index": 12 }, { "type": "text", "bbox": [ 108, 676, 504, 732 ], "lines": [ { "bbox": [ 106, 676, 506, 690 ], "spans": [ { "bbox": [ 106, 676, 506, 690 ], "score": 1.0, "content": "We trained three popular models: DeiT-III, OpenCLIP, DINOv2 with and without the introduction", "type": "text" } ], "index": 13 }, { "bbox": [ 106, 687, 505, 701 ], "spans": [ { "bbox": [ 106, 687, 505, 701 ], "score": 1.0, "content": "of register tokens. We observe in Fig. 19 the attention maps in the last layer of the Vision Trans-", "type": "text" } ], "index": 14 }, { "bbox": [ 106, 699, 505, 711 ], "spans": [ { "bbox": [ 106, 699, 505, 711 ], "score": 1.0, "content": "former, for all three cases. We see that our approach provides much cleaner attention maps, with", "type": "text" } ], "index": 15 }, { "bbox": [ 106, 709, 506, 723 ], "spans": [ { "bbox": [ 106, 709, 506, 723 ], "score": 1.0, "content": "considerably fewer artifacts, explaining the improvement on the downstream object discovery task", "type": "text" } ], "index": 16 }, { "bbox": [ 105, 719, 506, 735 ], "spans": [ { "bbox": [ 105, 719, 506, 735 ], "score": 1.0, "content": "mentioned in Sec. 3.3. The feature maps are also visibly improved, as shown in Fig. 20. Finally,", "type": "text" } ], "index": 17 }, { "bbox": [ 106, 81, 505, 94 ], "spans": [ { "bbox": [ 106, 81, 505, 94 ], "score": 1.0, "content": "we also show the norm of the patch tokens in Fig. 21, and confirm that in all three models, artifact", "type": "text", "cross_page": true } ], "index": 0 }, { "bbox": [ 105, 94, 253, 106 ], "spans": [ { "bbox": [ 105, 94, 253, 106 ], "score": 1.0, "content": "patches correspond to norm outliers.", "type": "text", "cross_page": true } ], "index": 1 } ], "index": 15, "bbox_fs": [ 105, 676, 506, 735 ] } ] }, { "preproc_blocks": [ { "type": "text", "bbox": [ 106, 82, 505, 105 ], "lines": [ { "bbox": [ 106, 81, 505, 94 ], "spans": [ { "bbox": [ 106, 81, 505, 94 ], "score": 1.0, "content": "we also show the norm of the patch tokens in Fig. 21, and confirm that in all three models, artifact", "type": "text" } ], "index": 0 }, { "bbox": [ 105, 94, 253, 106 ], "spans": [ { "bbox": [ 105, 94, 253, 106 ], "score": 1.0, "content": "patches correspond to norm outliers.", "type": "text" } ], "index": 1 } ], "index": 0.5 } ], "page_idx": 17, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 752, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 764 ], "spans": [ { "bbox": [ 299, 750, 312, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "text", "bbox": [ 106, 82, 505, 105 ], "lines": [], "index": 0.5, "bbox_fs": [ 105, 81, 505, 106 ], "lines_deleted": true } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 109, 92, 508, 694 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 92, 508, 694 ], "group_id": 0, "lines": [ { "bbox": [ 109, 92, 508, 694 ], "spans": [ { "bbox": [ 109, 92, 508, 694 ], "score": 0.955, "type": "image", "image_path": "9a36cf1a6809e8330cc35ef235e9f4512df473aae621315af36529bbb424832b.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 109, 92, 508, 292.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 109, 292.66666666666663, 508, 493.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 109, 493.33333333333326, 508, 693.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 122, 703, 486, 715 ], "group_id": 0, "lines": [ { "bbox": [ 122, 702, 488, 717 ], "spans": [ { "bbox": [ 122, 702, 488, 717 ], "score": 1.0, "content": "Figure 19: Attention maps of models trained without and with registers on various images.", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ], "page_idx": 18, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 301, 752, 311, 760 ], "lines": [ { "bbox": [ 299, 750, 312, 765 ], "spans": [ { "bbox": [ 299, 750, 312, 765 ], "score": 1.0, "content": "", "type": "text", "height": 15, "width": 13 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 109, 92, 508, 694 ], "blocks": [ { "type": "image_body", "bbox": [ 109, 92, 508, 694 ], "group_id": 0, "lines": [ { "bbox": [ 109, 92, 508, 694 ], "spans": [ { "bbox": [ 109, 92, 508, 694 ], "score": 0.955, "type": "image", "image_path": "9a36cf1a6809e8330cc35ef235e9f4512df473aae621315af36529bbb424832b.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 109, 92, 508, 292.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 109, 292.66666666666663, 508, 493.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 109, 493.33333333333326, 508, 693.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 122, 703, 486, 715 ], "group_id": 0, "lines": [ { "bbox": [ 122, 702, 488, 717 ], "spans": [ { "bbox": [ 122, 702, 488, 717 ], "score": 1.0, "content": "Figure 19: Attention maps of models trained without and with registers on various images.", "type": "text" } ], "index": 3 } ], "index": 3 } ], "index": 2.0 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 110, 75, 507, 686 ], "blocks": [ { "type": "image_body", "bbox": [ 110, 75, 507, 686 ], "group_id": 0, "lines": [ { "bbox": [ 110, 75, 507, 686 ], "spans": [ { "bbox": [ 110, 75, 507, 686 ], "score": 0.962, "type": "image", "image_path": "d7e4fc126cf9376fc5002ba64b2c9748d5f652d55390ac7b3b075cfee6c2614f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 110, 75, 507, 278.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 110, 278.66666666666663, 507, 482.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 110, 482.33333333333326, 507, 685.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 691, 505, 726 ], "group_id": 0, "lines": [ { "bbox": [ 106, 692, 505, 704 ], "spans": [ { "bbox": [ 106, 692, 505, 704 ], "score": 1.0, "content": "Figure 20: First principal component of the feature maps output by models trained without and", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 702, 505, 716 ], "spans": [ { "bbox": [ 105, 702, 505, 716 ], "score": 1.0, "content": "with registers on various images. The components are whitened and the colormap covers the range", "type": "text" } ], "index": 4 }, { "bbox": [ 107, 713, 160, 727 ], "spans": [ { "bbox": [ 107, 714, 154, 726 ], "score": 0.91, "content": "[ - 3 \\sigma , \\bar { + } 3 \\sigma ]", "type": "inline_equation" }, { "bbox": [ 155, 713, 160, 727 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ], "page_idx": 19, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 752, 311, 760 ], "lines": [ { "bbox": [ 298, 750, 313, 764 ], "spans": [ { "bbox": [ 298, 750, 313, 764 ], "score": 1.0, "content": "", "type": "text", "height": 14, "width": 15 } ] } ] } ], "para_blocks": [ { "type": "image", "bbox": [ 110, 75, 507, 686 ], "blocks": [ { "type": "image_body", "bbox": [ 110, 75, 507, 686 ], "group_id": 0, "lines": [ { "bbox": [ 110, 75, 507, 686 ], "spans": [ { "bbox": [ 110, 75, 507, 686 ], "score": 0.962, "type": "image", "image_path": "d7e4fc126cf9376fc5002ba64b2c9748d5f652d55390ac7b3b075cfee6c2614f.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 110, 75, 507, 278.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 110, 278.66666666666663, 507, 482.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 110, 482.33333333333326, 507, 685.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 106, 691, 505, 726 ], "group_id": 0, "lines": [ { "bbox": [ 106, 692, 505, 704 ], "spans": [ { "bbox": [ 106, 692, 505, 704 ], "score": 1.0, "content": "Figure 20: First principal component of the feature maps output by models trained without and", "type": "text" } ], "index": 3 }, { "bbox": [ 105, 702, 505, 716 ], "spans": [ { "bbox": [ 105, 702, 505, 716 ], "score": 1.0, "content": "with registers on various images. The components are whitened and the colormap covers the range", "type": "text" } ], "index": 4 }, { "bbox": [ 107, 713, 160, 727 ], "spans": [ { "bbox": [ 107, 714, 154, 726 ], "score": 0.91, "content": "[ - 3 \\sigma , \\bar { + } 3 \\sigma ]", "type": "inline_equation" }, { "bbox": [ 155, 713, 160, 727 ], "score": 1.0, "content": ".", "type": "text" } ], "index": 5 } ], "index": 4 } ], "index": 2.5 } ] }, { "preproc_blocks": [ { "type": "image", "bbox": [ 110, 81, 507, 689 ], "blocks": [ { "type": "image_body", "bbox": [ 110, 81, 507, 689 ], "group_id": 0, "lines": [ { "bbox": [ 110, 81, 507, 689 ], "spans": [ { "bbox": [ 110, 81, 507, 689 ], "score": 0.966, "type": "image", "image_path": "a81d0c0f1c9d7a70a87440ea06c2aef1237a9845cd4ce99c226a6d525d47e1d8.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 110, 81, 507, 283.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 110, 283.66666666666663, 507, 486.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 110, 486.33333333333326, 507, 688.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 104, 697, 504, 720 ], "group_id": 0, "lines": [ { "bbox": [ 105, 696, 505, 711 ], "spans": [ { "bbox": [ 105, 696, 505, 711 ], "score": 1.0, "content": "Figure 21: Maps of token norms for models trained without and with registers on various images.", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 708, 390, 721 ], "spans": [ { "bbox": [ 106, 708, 390, 721 ], "score": 1.0, "content": "The norm outliers are very visible for models trained without registers.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 } ], "page_idx": 20, "page_size": [ 595, 841 ], "discarded_blocks": [ { "type": "discarded", "bbox": [ 107, 27, 293, 37 ], "lines": [ { "bbox": [ 106, 26, 294, 38 ], "spans": [ { "bbox": [ 106, 26, 294, 38 ], "score": 1.0, "content": "Published as a conference paper at ICLR 2024", "type": "text" } ] } ] }, { "type": "discarded", "bbox": [ 300, 752, 310, 760 ], "lines": [] } ], "para_blocks": [ { "type": "image", "bbox": [ 110, 81, 507, 689 ], "blocks": [ { "type": "image_body", "bbox": [ 110, 81, 507, 689 ], "group_id": 0, "lines": [ { "bbox": [ 110, 81, 507, 689 ], "spans": [ { "bbox": [ 110, 81, 507, 689 ], "score": 0.966, "type": "image", "image_path": "a81d0c0f1c9d7a70a87440ea06c2aef1237a9845cd4ce99c226a6d525d47e1d8.jpg" } ] } ], "index": 1, "virtual_lines": [ { "bbox": [ 110, 81, 507, 283.66666666666663 ], "spans": [], "index": 0 }, { "bbox": [ 110, 283.66666666666663, 507, 486.33333333333326 ], "spans": [], "index": 1 }, { "bbox": [ 110, 486.33333333333326, 507, 688.9999999999999 ], "spans": [], "index": 2 } ] }, { "type": "image_caption", "bbox": [ 104, 697, 504, 720 ], "group_id": 0, "lines": [ { "bbox": [ 105, 696, 505, 711 ], "spans": [ { "bbox": [ 105, 696, 505, 711 ], "score": 1.0, "content": "Figure 21: Maps of token norms for models trained without and with registers on various images.", "type": "text" } ], "index": 3 }, { "bbox": [ 106, 708, 390, 721 ], "spans": [ { "bbox": [ 106, 708, 390, 721 ], "score": 1.0, "content": "The norm outliers are very visible for models trained without registers.", "type": "text" } ], "index": 4 } ], "index": 3.5 } ], "index": 2.25 } ] } ], "_backend": "pipeline", "_version_name": "2.1.11" }