diff --git "a/parse/train/k7aeAz4Vbb/k7aeAz4Vbb_middle.json" "b/parse/train/k7aeAz4Vbb/k7aeAz4Vbb_middle.json" new file mode 100644--- /dev/null +++ "b/parse/train/k7aeAz4Vbb/k7aeAz4Vbb_middle.json" @@ -0,0 +1,33586 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 128, + 97, + 483, + 137 + ], + "lines": [ + { + "bbox": [ + 126, + 97, + 485, + 117 + ], + "spans": [ + { + "bbox": [ + 126, + 97, + 485, + 117 + ], + "score": 1.0, + "content": "Instance-Conditional Knowledge Distillation for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 241, + 117, + 370, + 139 + ], + "spans": [ + { + "bbox": [ + 241, + 117, + 370, + 139 + ], + "score": 1.0, + "content": "Object Detection", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 179, + 230, + 213 + ], + "lines": [ + { + "bbox": [ + 144, + 178, + 203, + 193 + ], + "spans": [ + { + "bbox": [ + 144, + 178, + 203, + 193 + ], + "score": 1.0, + "content": "Zijian Kang∗", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 190, + 225, + 203 + ], + "spans": [ + { + "bbox": [ + 119, + 190, + 225, + 203 + ], + "score": 1.0, + "content": "Xi’an Jiaotong University", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 202, + 230, + 214 + ], + "spans": [ + { + "bbox": [ + 114, + 202, + 230, + 214 + ], + "score": 1.0, + "content": "kzj123@stu.xjtu.edu.cn", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 242, + 179, + 363, + 213 + ], + "lines": [ + { + "bbox": [ + 269, + 178, + 339, + 192 + ], + "spans": [ + { + "bbox": [ + 269, + 178, + 339, + 192 + ], + "score": 1.0, + "content": "Peizhen Zhang∗", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 258, + 189, + 347, + 204 + ], + "spans": [ + { + "bbox": [ + 258, + 189, + 347, + 204 + ], + "score": 1.0, + "content": "MEGVII Technology", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 242, + 202, + 365, + 214 + ], + "spans": [ + { + "bbox": [ + 242, + 202, + 365, + 214 + ], + "score": 1.0, + "content": "zhangpeizhen@megvii.com", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 375, + 179, + 497, + 213 + ], + "lines": [ + { + "bbox": [ + 401, + 178, + 472, + 193 + ], + "spans": [ + { + "bbox": [ + 401, + 178, + 472, + 193 + ], + "score": 1.0, + "content": "Xiangyu Zhang", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 392, + 189, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 392, + 189, + 480, + 204 + ], + "score": 1.0, + "content": "MEGVII Technology", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 375, + 202, + 498, + 215 + ], + "spans": [ + { + "bbox": [ + 375, + 202, + 498, + 215 + ], + "score": 1.0, + "content": "zhangxiangyu@megvii.com", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 164, + 230, + 259, + 263 + ], + "lines": [ + { + "bbox": [ + 191, + 229, + 232, + 241 + ], + "spans": [ + { + "bbox": [ + 191, + 229, + 232, + 241 + ], + "score": 1.0, + "content": "Jian Sun", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 167, + 238, + 255, + 254 + ], + "spans": [ + { + "bbox": [ + 167, + 238, + 255, + 254 + ], + "score": 1.0, + "content": "MEGVII Technology", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 163, + 252, + 260, + 264 + ], + "spans": [ + { + "bbox": [ + 163, + 252, + 260, + 264 + ], + "score": 1.0, + "content": "sunjian@megvii.com", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 320, + 229, + 447, + 263 + ], + "lines": [ + { + "bbox": [ + 349, + 228, + 419, + 243 + ], + "spans": [ + { + "bbox": [ + 349, + 228, + 419, + 243 + ], + "score": 1.0, + "content": "Nanning Zheng", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 331, + 240, + 436, + 253 + ], + "spans": [ + { + "bbox": [ + 331, + 240, + 436, + 253 + ], + "score": 1.0, + "content": "Xi’an Jiaotong University", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 252, + 447, + 264 + ], + "spans": [ + { + "bbox": [ + 320, + 252, + 447, + 264 + ], + "score": 1.0, + "content": "nnzheng@mail.xjtu.edu.cn", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 283, + 291, + 328, + 303 + ], + "lines": [ + { + "bbox": [ + 281, + 290, + 330, + 305 + ], + "spans": [ + { + "bbox": [ + 281, + 290, + 330, + 305 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 143, + 316, + 469, + 490 + ], + "lines": [ + { + "bbox": [ + 141, + 316, + 469, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 316, + 469, + 329 + ], + "score": 1.0, + "content": "Knowledge distillation has shown great success in classification, however, it is still", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 326, + 469, + 340 + ], + "spans": [ + { + "bbox": [ + 141, + 326, + 469, + 340 + ], + "score": 1.0, + "content": "challenging for detection. In a typical image for detection, representations from", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 337, + 469, + 351 + ], + "spans": [ + { + "bbox": [ + 141, + 337, + 469, + 351 + ], + "score": 1.0, + "content": "different locations may have different contributions to detection targets, making", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 348, + 469, + 361 + ], + "spans": [ + { + "bbox": [ + 141, + 348, + 469, + 361 + ], + "score": 1.0, + "content": "the distillation hard to balance. In this paper, we propose a conditional distillation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 360, + 469, + 372 + ], + "spans": [ + { + "bbox": [ + 141, + 360, + 469, + 372 + ], + "score": 1.0, + "content": "framework to distill the desired knowledge, namely knowledge that is beneficial", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 371, + 469, + 382 + ], + "spans": [ + { + "bbox": [ + 142, + 371, + 469, + 382 + ], + "score": 1.0, + "content": "in terms of both classification and localization for every instance. The framework", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 382, + 469, + 394 + ], + "spans": [ + { + "bbox": [ + 141, + 382, + 469, + 394 + ], + "score": 1.0, + "content": "introduces a learnable conditional decoding module, which retrieves information", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 393, + 470, + 405 + ], + "spans": [ + { + "bbox": [ + 141, + 393, + 470, + 405 + ], + "score": 1.0, + "content": "given each target instance as query. Specifically, we encode the condition informa-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 404, + 470, + 415 + ], + "spans": [ + { + "bbox": [ + 141, + 404, + 470, + 415 + ], + "score": 1.0, + "content": "tion as query and use the teacher’s representations as key. The attention between", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 414, + 470, + 428 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 470, + 428 + ], + "score": 1.0, + "content": "query and key is used to measure the contribution of different features, guided by", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 426, + 470, + 438 + ], + "spans": [ + { + "bbox": [ + 141, + 426, + 470, + 438 + ], + "score": 1.0, + "content": "a localization-recognition-sensitive auxiliary task. Extensive experiments demon-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 437, + 470, + 448 + ], + "spans": [ + { + "bbox": [ + 141, + 437, + 470, + 448 + ], + "score": 1.0, + "content": "strate the efficacy of our method: we observe impressive improvements under", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 447, + 470, + 459 + ], + "spans": [ + { + "bbox": [ + 141, + 447, + 470, + 459 + ], + "score": 1.0, + "content": "various settings. Notably, we boost RetinaNet with ResNet-50 backbone from", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 458, + 470, + 470 + ], + "spans": [ + { + "bbox": [ + 141, + 458, + 172, + 470 + ], + "score": 1.0, + "content": "37.4 to", + "type": "text" + }, + { + "bbox": [ + 173, + 458, + 214, + 469 + ], + "score": 0.35, + "content": "4 0 . 7 \\mathrm { m A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 458, + 243, + 469 + ], + "score": 0.82, + "content": "\\left( + 3 . 3 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 458, + 269, + 470 + ], + "score": 1.0, + "content": "under", + "type": "text" + }, + { + "bbox": [ + 270, + 459, + 284, + 469 + ], + "score": 0.86, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 458, + 470, + 470 + ], + "score": 1.0, + "content": "schedule, that even surpasses the teacher (40.4", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 142, + 469, + 469, + 480 + ], + "spans": [ + { + "bbox": [ + 142, + 469, + 304, + 480 + ], + "score": 1.0, + "content": "mAP) with ResNet-101 backbone under", + "type": "text" + }, + { + "bbox": [ + 305, + 469, + 319, + 479 + ], + "score": 0.86, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 469, + 469, + 480 + ], + "score": 1.0, + "content": "schedule. Code has been released on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 142, + 480, + 306, + 492 + ], + "spans": [ + { + "bbox": [ + 142, + 480, + 306, + 492 + ], + "score": 1.0, + "content": "https://github.com/megvii-research/ICD.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 25.5 + }, + { + "type": "title", + "bbox": [ + 108, + 511, + 190, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 192, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 192, + 527 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 505, + 624 + ], + "lines": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "Deep learning applications blossom in recent years with the breakthrough of Deep Neural Networks", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "(DNNs) [17, 24, 21]. In pursuit of high performance, advanced DNNs usually stack tons of blocks with", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 559, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 570 + ], + "score": 1.0, + "content": "millions of parameters, which are computation and memory consuming. The heavy design hinders", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "the deployment of many practical downstream applications like object detection in resource-limited", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 579, + 507, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 507, + 592 + ], + "score": 1.0, + "content": "devices. Plenty of techniques have been proposed to address this issue, like network pruning [15, 27,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "18], quantization [22, 23, 35], mobile architecture design [38, 39] and knowledge distillation (KD)", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "[19, 37, 43]. Among them, KD is one of the most popular choices, since it can boost a target network", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 613, + 371, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 371, + 625 + ], + "score": 1.0, + "content": "without introducing extra inference-time burden or modifications.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 107, + 629, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 630, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 505, + 641 + ], + "score": 1.0, + "content": "KD is popularized by Hinton et al. [19], where knowledge of a strong pretrained teacher network is", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "score": 1.0, + "content": "transferred to a small target student network in the classification scenario. Many good works emerge", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "following the classification track [50, 37, 32]. However, most methods for classification perform", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "badly in the detection: only slight improvements are observed [28, 51]. This can be attributed to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 673, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 505, + 686 + ], + "score": 1.0, + "content": "two reasons: (1) Other than category classification, another challenging goal to localize the object is", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 684, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 505, + 696 + ], + "score": 1.0, + "content": "seldomly considered. (2) Multiple target objects are presented in an image for detection, where objects", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 106, + 730, + 386, + 743 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 386, + 743 + ], + "score": 1.0, + "content": "35th Conference on Neural Information Processing Systems (NeurIPS 2021).", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 119, + 703, + 192, + 713 + ], + "lines": [ + { + "bbox": [ + 119, + 701, + 195, + 714 + ], + "spans": [ + { + "bbox": [ + 119, + 701, + 195, + 714 + ], + "score": 1.0, + "content": "∗Equal contribution.", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 128, + 97, + 483, + 137 + ], + "lines": [ + { + "bbox": [ + 126, + 97, + 485, + 117 + ], + "spans": [ + { + "bbox": [ + 126, + 97, + 485, + 117 + ], + "score": 1.0, + "content": "Instance-Conditional Knowledge Distillation for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 241, + 117, + 370, + 139 + ], + "spans": [ + { + "bbox": [ + 241, + 117, + 370, + 139 + ], + "score": 1.0, + "content": "Object Detection", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 179, + 230, + 213 + ], + "lines": [ + { + "bbox": [ + 144, + 178, + 203, + 193 + ], + "spans": [ + { + "bbox": [ + 144, + 178, + 203, + 193 + ], + "score": 1.0, + "content": "Zijian Kang∗", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 190, + 225, + 203 + ], + "spans": [ + { + "bbox": [ + 119, + 190, + 225, + 203 + ], + "score": 1.0, + "content": "Xi’an Jiaotong University", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 202, + 230, + 214 + ], + "spans": [ + { + "bbox": [ + 114, + 202, + 230, + 214 + ], + "score": 1.0, + "content": "kzj123@stu.xjtu.edu.cn", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 114, + 178, + 230, + 214 + ] + }, + { + "type": "text", + "bbox": [ + 242, + 179, + 363, + 213 + ], + "lines": [ + { + "bbox": [ + 269, + 178, + 339, + 192 + ], + "spans": [ + { + "bbox": [ + 269, + 178, + 339, + 192 + ], + "score": 1.0, + "content": "Peizhen Zhang∗", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 258, + 189, + 347, + 204 + ], + "spans": [ + { + "bbox": [ + 258, + 189, + 347, + 204 + ], + "score": 1.0, + "content": "MEGVII Technology", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 242, + 202, + 365, + 214 + ], + "spans": [ + { + "bbox": [ + 242, + 202, + 365, + 214 + ], + "score": 1.0, + "content": "zhangpeizhen@megvii.com", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6, + "bbox_fs": [ + 242, + 178, + 365, + 214 + ] + }, + { + "type": "text", + "bbox": [ + 375, + 179, + 497, + 213 + ], + "lines": [ + { + "bbox": [ + 401, + 178, + 472, + 193 + ], + "spans": [ + { + "bbox": [ + 401, + 178, + 472, + 193 + ], + "score": 1.0, + "content": "Xiangyu Zhang", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 392, + 189, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 392, + 189, + 480, + 204 + ], + "score": 1.0, + "content": "MEGVII Technology", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 375, + 202, + 498, + 215 + ], + "spans": [ + { + "bbox": [ + 375, + 202, + 498, + 215 + ], + "score": 1.0, + "content": "zhangxiangyu@megvii.com", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7, + "bbox_fs": [ + 375, + 178, + 498, + 215 + ] + }, + { + "type": "text", + "bbox": [ + 164, + 230, + 259, + 263 + ], + "lines": [ + { + "bbox": [ + 191, + 229, + 232, + 241 + ], + "spans": [ + { + "bbox": [ + 191, + 229, + 232, + 241 + ], + "score": 1.0, + "content": "Jian Sun", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 167, + 238, + 255, + 254 + ], + "spans": [ + { + "bbox": [ + 167, + 238, + 255, + 254 + ], + "score": 1.0, + "content": "MEGVII Technology", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 163, + 252, + 260, + 264 + ], + "spans": [ + { + "bbox": [ + 163, + 252, + 260, + 264 + ], + "score": 1.0, + "content": "sunjian@megvii.com", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 163, + 229, + 260, + 264 + ] + }, + { + "type": "text", + "bbox": [ + 320, + 229, + 447, + 263 + ], + "lines": [ + { + "bbox": [ + 349, + 228, + 419, + 243 + ], + "spans": [ + { + "bbox": [ + 349, + 228, + 419, + 243 + ], + "score": 1.0, + "content": "Nanning Zheng", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 331, + 240, + 436, + 253 + ], + "spans": [ + { + "bbox": [ + 331, + 240, + 436, + 253 + ], + "score": 1.0, + "content": "Xi’an Jiaotong University", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 252, + 447, + 264 + ], + "spans": [ + { + "bbox": [ + 320, + 252, + 447, + 264 + ], + "score": 1.0, + "content": "nnzheng@mail.xjtu.edu.cn", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 320, + 228, + 447, + 264 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 291, + 328, + 303 + ], + "lines": [ + { + "bbox": [ + 281, + 290, + 330, + 305 + ], + "spans": [ + { + "bbox": [ + 281, + 290, + 330, + 305 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 143, + 316, + 469, + 490 + ], + "lines": [ + { + "bbox": [ + 141, + 316, + 469, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 316, + 469, + 329 + ], + "score": 1.0, + "content": "Knowledge distillation has shown great success in classification, however, it is still", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 326, + 469, + 340 + ], + "spans": [ + { + "bbox": [ + 141, + 326, + 469, + 340 + ], + "score": 1.0, + "content": "challenging for detection. In a typical image for detection, representations from", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 337, + 469, + 351 + ], + "spans": [ + { + "bbox": [ + 141, + 337, + 469, + 351 + ], + "score": 1.0, + "content": "different locations may have different contributions to detection targets, making", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 348, + 469, + 361 + ], + "spans": [ + { + "bbox": [ + 141, + 348, + 469, + 361 + ], + "score": 1.0, + "content": "the distillation hard to balance. In this paper, we propose a conditional distillation", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 360, + 469, + 372 + ], + "spans": [ + { + "bbox": [ + 141, + 360, + 469, + 372 + ], + "score": 1.0, + "content": "framework to distill the desired knowledge, namely knowledge that is beneficial", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 142, + 371, + 469, + 382 + ], + "spans": [ + { + "bbox": [ + 142, + 371, + 469, + 382 + ], + "score": 1.0, + "content": "in terms of both classification and localization for every instance. The framework", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 382, + 469, + 394 + ], + "spans": [ + { + "bbox": [ + 141, + 382, + 469, + 394 + ], + "score": 1.0, + "content": "introduces a learnable conditional decoding module, which retrieves information", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 393, + 470, + 405 + ], + "spans": [ + { + "bbox": [ + 141, + 393, + 470, + 405 + ], + "score": 1.0, + "content": "given each target instance as query. Specifically, we encode the condition informa-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 404, + 470, + 415 + ], + "spans": [ + { + "bbox": [ + 141, + 404, + 470, + 415 + ], + "score": 1.0, + "content": "tion as query and use the teacher’s representations as key. The attention between", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 414, + 470, + 428 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 470, + 428 + ], + "score": 1.0, + "content": "query and key is used to measure the contribution of different features, guided by", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 426, + 470, + 438 + ], + "spans": [ + { + "bbox": [ + 141, + 426, + 470, + 438 + ], + "score": 1.0, + "content": "a localization-recognition-sensitive auxiliary task. Extensive experiments demon-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 437, + 470, + 448 + ], + "spans": [ + { + "bbox": [ + 141, + 437, + 470, + 448 + ], + "score": 1.0, + "content": "strate the efficacy of our method: we observe impressive improvements under", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 447, + 470, + 459 + ], + "spans": [ + { + "bbox": [ + 141, + 447, + 470, + 459 + ], + "score": 1.0, + "content": "various settings. Notably, we boost RetinaNet with ResNet-50 backbone from", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 458, + 470, + 470 + ], + "spans": [ + { + "bbox": [ + 141, + 458, + 172, + 470 + ], + "score": 1.0, + "content": "37.4 to", + "type": "text" + }, + { + "bbox": [ + 173, + 458, + 214, + 469 + ], + "score": 0.35, + "content": "4 0 . 7 \\mathrm { m A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 458, + 243, + 469 + ], + "score": 0.82, + "content": "\\left( + 3 . 3 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 458, + 269, + 470 + ], + "score": 1.0, + "content": "under", + "type": "text" + }, + { + "bbox": [ + 270, + 459, + 284, + 469 + ], + "score": 0.86, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 458, + 470, + 470 + ], + "score": 1.0, + "content": "schedule, that even surpasses the teacher (40.4", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 142, + 469, + 469, + 480 + ], + "spans": [ + { + "bbox": [ + 142, + 469, + 304, + 480 + ], + "score": 1.0, + "content": "mAP) with ResNet-101 backbone under", + "type": "text" + }, + { + "bbox": [ + 305, + 469, + 319, + 479 + ], + "score": 0.86, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 469, + 469, + 480 + ], + "score": 1.0, + "content": "schedule. Code has been released on", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 142, + 480, + 306, + 492 + ], + "spans": [ + { + "bbox": [ + 142, + 480, + 306, + 492 + ], + "score": 1.0, + "content": "https://github.com/megvii-research/ICD.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 25.5, + "bbox_fs": [ + 141, + 316, + 470, + 492 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 511, + 190, + 524 + ], + "lines": [ + { + "bbox": [ + 105, + 510, + 192, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 192, + 527 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 505, + 624 + ], + "lines": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "Deep learning applications blossom in recent years with the breakthrough of Deep Neural Networks", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "(DNNs) [17, 24, 21]. In pursuit of high performance, advanced DNNs usually stack tons of blocks with", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 559, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 570 + ], + "score": 1.0, + "content": "millions of parameters, which are computation and memory consuming. The heavy design hinders", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "the deployment of many practical downstream applications like object detection in resource-limited", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 579, + 507, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 507, + 592 + ], + "score": 1.0, + "content": "devices. Plenty of techniques have been proposed to address this issue, like network pruning [15, 27,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "18], quantization [22, 23, 35], mobile architecture design [38, 39] and knowledge distillation (KD)", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "[19, 37, 43]. Among them, KD is one of the most popular choices, since it can boost a target network", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 613, + 371, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 371, + 625 + ], + "score": 1.0, + "content": "without introducing extra inference-time burden or modifications.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 536, + 507, + 625 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 629, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 630, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 630, + 505, + 641 + ], + "score": 1.0, + "content": "KD is popularized by Hinton et al. [19], where knowledge of a strong pretrained teacher network is", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 654 + ], + "score": 1.0, + "content": "transferred to a small target student network in the classification scenario. Many good works emerge", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "following the classification track [50, 37, 32]. However, most methods for classification perform", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "badly in the detection: only slight improvements are observed [28, 51]. This can be attributed to", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 673, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 505, + 686 + ], + "score": 1.0, + "content": "two reasons: (1) Other than category classification, another challenging goal to localize the object is", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 684, + 505, + 696 + ], + "spans": [ + { + "bbox": [ + 105, + 684, + 505, + 696 + ], + "score": 1.0, + "content": "seldomly considered. (2) Multiple target objects are presented in an image for detection, where objects", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "can distribute in different locations. Due to these reasons, the knowledge becomes rather ambiguous", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "and imbalance in detection: representations from different positions like foreground or background,", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 260, + 474, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 474, + 276 + ], + "score": 1.0, + "content": "borders or centers, could have different contributions, which makes distillation challenging.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 630, + 505, + 696 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 75, + 495, + 190 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 75, + 495, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 75, + 495, + 190 + ], + "spans": [ + { + "bbox": [ + 116, + 75, + 495, + 190 + ], + "score": 0.969, + "type": "image", + "image_path": "d42521b8ba9824717a7cf824913ac3671acb0f17bcca343180144f7444febb0a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 75, + 495, + 113.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 113.33333333333334, + 495, + 151.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 151.66666666666669, + 495, + 190.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 196, + 506, + 230 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "Figure 1: Compare with different methods for knowledge distillation. (a) KD [19] for classification", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "transfers logits. (b) Recent methods for detection KD distill intermediate features, different region-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 218, + 487, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 487, + 232 + ], + "score": 1.0, + "content": "based sampling methods are proposed. (c) Our method explicitly distill the desired knowledge.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 240, + 503, + 273 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 505, + 253 + ], + "score": 1.0, + "content": "can distribute in different locations. Due to these reasons, the knowledge becomes rather ambiguous", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "and imbalance in detection: representations from different positions like foreground or background,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 260, + 474, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 474, + 276 + ], + "score": 1.0, + "content": "borders or centers, could have different contributions, which makes distillation challenging.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 278, + 505, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "score": 1.0, + "content": "To handle the above challenge, two strategies are usually adopted by previous methods in detection.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 289, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 505, + 301 + ], + "score": 1.0, + "content": "First, the distillation is usually conducted among intermediate representations, which cover all", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 300, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 312 + ], + "score": 1.0, + "content": "necessary features for both classification and localization. Second, different feature selection methods", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "score": 1.0, + "content": "are proposed to overcome the imbalance issue. Existent works could be divided into three types", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "according to the feature selection paradigm: proposal-based, rule-based and attention-based. In", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 506, + 346 + ], + "score": 1.0, + "content": "proposal-based methods [28, 11, 6], proposal regions predicted by the RPN [36] or detector are", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "sampled for distillation. In rule-based methods [14, 45], regions selected by predesigned rules like", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 355, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 367 + ], + "score": 1.0, + "content": "foreground or label-assigned regions are sampled. Despite their improvements, limitations still exist", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 364, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 379 + ], + "score": 1.0, + "content": "due to the hand-crafted designs, e.g., many methods neglect the informative context regions or involve", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "score": 1.0, + "content": "meticulous decisions. Recently, Zhang et al. [51] propose to use attention [43], a type of intermediate", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "score": 1.0, + "content": "activation of the network, to guide the distillation. Although attention provides inherent hints for", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 399, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 410 + ], + "score": 1.0, + "content": "discriminative areas, the relation between activation and knowledge for detection is still unclear. To", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "score": 1.0, + "content": "further improve KD quality, we hope to provide an explicit solution to connect the desired knowledge", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 420, + 196, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 196, + 432 + ], + "score": 1.0, + "content": "with feature selection.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 505, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "Towards this goal, we present Instance-Conditional knowledge Distillation (ICD), which introduces a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 461 + ], + "score": 1.0, + "content": "new KD framework based-on conditional knowledge retrieval. In ICD, we propose to use a decoding", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 457, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 104, + 457, + 506, + 472 + ], + "score": 1.0, + "content": "network to find and distill knowledge associated with different instances, we deem such knowledge", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 469, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 482 + ], + "score": 1.0, + "content": "as instance-conditional knowledge. Fig. 1 shows the framework and compares it with former ones,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "score": 1.0, + "content": "ICD learns to find desired knowledge, which is much more flexible than previous methods, and is", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "more consistent with detection targets. In detail, we design a conditional decoding module to locate", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 501, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 515 + ], + "score": 1.0, + "content": "knowledge, the correlation between knowledge and each instance is measured by the instance-aware", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "attention via the transformer decoder [5, 43]. In which human observed instances are projected", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 524, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 536 + ], + "score": 1.0, + "content": "to query and the correlation is measured by scaled-product attention between query and teacher’s", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 535, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 505, + 547 + ], + "score": 1.0, + "content": "representations. Following this formulation, the distillation is conducted over features decomposed by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 544, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 559 + ], + "score": 1.0, + "content": "the decoder and weighted by the instance-aware attention. Last but not least, to optimize the decoding", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "module, we also introduce an auxiliary task, which teaches the decoder to find useful information", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "for identification and localization. The task defines the goal for knowledge retrieval, it facilitates the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 578, + 448, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 448, + 590 + ], + "score": 1.0, + "content": "decoder instead of the student. Overall, our contribution is summarized in three-fold:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 600, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "• We introduce a novel framework to locate useful knowledge in detection KD, we formulate the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 611, + 477, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 611, + 477, + 623 + ], + "score": 1.0, + "content": "knowledge retrieval explicitly by a decoding network and optimize it via an auxiliary task.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 626, + 507, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 507, + 641 + ], + "score": 1.0, + "content": "• We adopt the conditional modeling paradigm to facilitate instance-wise knowledge transferring.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 113, + 638, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 113, + 638, + 505, + 652 + ], + "score": 1.0, + "content": "We encode human observed instances as query and decompose teacher’s representations to key", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 650, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 505, + 662 + ], + "score": 1.0, + "content": "and value to locate fine-grained knowledge. To our knowledge, it is the first trial to explore", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 113, + 660, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 113, + 660, + 288, + 673 + ], + "score": 1.0, + "content": "instance-oriented knowledge for detection.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 507, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 507, + 691 + ], + "score": 1.0, + "content": "• We perform comprehensive experiments on challenging benchmarks. Results demonstrate impres-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 113, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 113, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "sive improvements over various detectors with up to 4 AP gain in MS-COCO, including recent", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 114, + 700, + 435, + 713 + ], + "score": 1.0, + "content": "detectors for instance segmentation [41, 46, 16]. In some cases, students with", + "type": "text" + }, + { + "bbox": [ + 436, + 700, + 450, + 711 + ], + "score": 0.86, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "schedule are", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 113, + 710, + 434, + 724 + ], + "spans": [ + { + "bbox": [ + 113, + 710, + 388, + 724 + ], + "score": 1.0, + "content": "even able to outperform their teachers with larger backbones trained", + "type": "text" + }, + { + "bbox": [ + 389, + 711, + 403, + 721 + ], + "score": 0.85, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 710, + 434, + 724 + ], + "score": 1.0, + "content": "longer.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 41.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 75, + 495, + 190 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 75, + 495, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 75, + 495, + 190 + ], + "spans": [ + { + "bbox": [ + 116, + 75, + 495, + 190 + ], + "score": 0.969, + "type": "image", + "image_path": "d42521b8ba9824717a7cf824913ac3671acb0f17bcca343180144f7444febb0a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 75, + 495, + 113.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 113.33333333333334, + 495, + 151.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 151.66666666666669, + 495, + 190.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 196, + 506, + 230 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "Figure 1: Compare with different methods for knowledge distillation. (a) KD [19] for classification", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 219 + ], + "score": 1.0, + "content": "transfers logits. (b) Recent methods for detection KD distill intermediate features, different region-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 218, + 487, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 487, + 232 + ], + "score": 1.0, + "content": "based sampling methods are proposed. (c) Our method explicitly distill the desired knowledge.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 240, + 503, + 273 + ], + "lines": [], + "index": 7, + "bbox_fs": [ + 105, + 239, + 505, + 276 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 278, + 505, + 431 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "score": 1.0, + "content": "To handle the above challenge, two strategies are usually adopted by previous methods in detection.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 289, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 505, + 301 + ], + "score": 1.0, + "content": "First, the distillation is usually conducted among intermediate representations, which cover all", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 300, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 506, + 312 + ], + "score": 1.0, + "content": "necessary features for both classification and localization. Second, different feature selection methods", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 324 + ], + "score": 1.0, + "content": "are proposed to overcome the imbalance issue. Existent works could be divided into three types", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 335 + ], + "score": 1.0, + "content": "according to the feature selection paradigm: proposal-based, rule-based and attention-based. In", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 333, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 506, + 346 + ], + "score": 1.0, + "content": "proposal-based methods [28, 11, 6], proposal regions predicted by the RPN [36] or detector are", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "sampled for distillation. In rule-based methods [14, 45], regions selected by predesigned rules like", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 355, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 367 + ], + "score": 1.0, + "content": "foreground or label-assigned regions are sampled. Despite their improvements, limitations still exist", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 364, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 379 + ], + "score": 1.0, + "content": "due to the hand-crafted designs, e.g., many methods neglect the informative context regions or involve", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 389 + ], + "score": 1.0, + "content": "meticulous decisions. Recently, Zhang et al. [51] propose to use attention [43], a type of intermediate", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 400 + ], + "score": 1.0, + "content": "activation of the network, to guide the distillation. Although attention provides inherent hints for", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 399, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 410 + ], + "score": 1.0, + "content": "discriminative areas, the relation between activation and knowledge for detection is still unclear. To", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "score": 1.0, + "content": "further improve KD quality, we hope to provide an explicit solution to connect the desired knowledge", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 420, + 196, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 196, + 432 + ], + "score": 1.0, + "content": "with feature selection.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 277, + 506, + 432 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 436, + 505, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "Towards this goal, we present Instance-Conditional knowledge Distillation (ICD), which introduces a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 446, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 461 + ], + "score": 1.0, + "content": "new KD framework based-on conditional knowledge retrieval. In ICD, we propose to use a decoding", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 457, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 104, + 457, + 506, + 472 + ], + "score": 1.0, + "content": "network to find and distill knowledge associated with different instances, we deem such knowledge", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 469, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 482 + ], + "score": 1.0, + "content": "as instance-conditional knowledge. Fig. 1 shows the framework and compares it with former ones,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 506, + 492 + ], + "score": 1.0, + "content": "ICD learns to find desired knowledge, which is much more flexible than previous methods, and is", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "more consistent with detection targets. In detail, we design a conditional decoding module to locate", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 501, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 506, + 515 + ], + "score": 1.0, + "content": "knowledge, the correlation between knowledge and each instance is measured by the instance-aware", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "attention via the transformer decoder [5, 43]. In which human observed instances are projected", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 524, + 506, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 536 + ], + "score": 1.0, + "content": "to query and the correlation is measured by scaled-product attention between query and teacher’s", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 535, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 505, + 547 + ], + "score": 1.0, + "content": "representations. Following this formulation, the distillation is conducted over features decomposed by", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 544, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 559 + ], + "score": 1.0, + "content": "the decoder and weighted by the instance-aware attention. Last but not least, to optimize the decoding", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "module, we also introduce an auxiliary task, which teaches the decoder to find useful information", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "for identification and localization. The task defines the goal for knowledge retrieval, it facilitates the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 578, + 448, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 448, + 590 + ], + "score": 1.0, + "content": "decoder instead of the student. Overall, our contribution is summarized in three-fold:", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 29.5, + "bbox_fs": [ + 104, + 435, + 506, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 600, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "• We introduce a novel framework to locate useful knowledge in detection KD, we formulate the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 611, + 477, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 611, + 477, + 623 + ], + "score": 1.0, + "content": "knowledge retrieval explicitly by a decoding network and optimize it via an auxiliary task.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 626, + 507, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 507, + 641 + ], + "score": 1.0, + "content": "• We adopt the conditional modeling paradigm to facilitate instance-wise knowledge transferring.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 113, + 638, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 113, + 638, + 505, + 652 + ], + "score": 1.0, + "content": "We encode human observed instances as query and decompose teacher’s representations to key", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 650, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 505, + 662 + ], + "score": 1.0, + "content": "and value to locate fine-grained knowledge. To our knowledge, it is the first trial to explore", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 113, + 660, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 113, + 660, + 288, + 673 + ], + "score": 1.0, + "content": "instance-oriented knowledge for detection.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 507, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 507, + 691 + ], + "score": 1.0, + "content": "• We perform comprehensive experiments on challenging benchmarks. Results demonstrate impres-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 113, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 113, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "sive improvements over various detectors with up to 4 AP gain in MS-COCO, including recent", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 114, + 700, + 435, + 713 + ], + "score": 1.0, + "content": "detectors for instance segmentation [41, 46, 16]. In some cases, students with", + "type": "text" + }, + { + "bbox": [ + 436, + 700, + 450, + 711 + ], + "score": 0.86, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "schedule are", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 113, + 710, + 434, + 724 + ], + "spans": [ + { + "bbox": [ + 113, + 710, + 388, + 724 + ], + "score": 1.0, + "content": "even able to outperform their teachers with larger backbones trained", + "type": "text" + }, + { + "bbox": [ + 389, + 711, + 403, + 721 + ], + "score": 0.85, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 710, + 434, + 724 + ], + "score": 1.0, + "content": "longer.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 599, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 202, + 84 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 203, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 203, + 86 + ], + "score": 1.0, + "content": "2 Related Works", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 102, + 227, + 114 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 229, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 229, + 117 + ], + "score": 1.0, + "content": "2.1 Knowledge Distillation", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 127, + 505, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "Knowledge distillation aims to transfer knowledge from a strong teacher to a weaker student network", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 138, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 151 + ], + "score": 1.0, + "content": "to facilitate supervised learning. The teacher is usually a large pretrained network, who provides", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "smoother supervision and more hints on visual concepts, that improves the training quality and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "convergence speed [49, 9]. KD for image classification has been studied for years, they are usually", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "categorized into three types [13]: response-based [19], feature-based [37, 43] and relation-based [32].", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 106, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "Among them, feature-based distillation over multi-scale features is adopted from most of detection", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "KD works, to deal with knowledge among multiple instances in different regions. Most of these works", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "can be formulated as region selection for distillation, where foreground-background unbalancing is", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 233 + ], + "score": 1.0, + "content": "considered as a key problem in some studies [45, 51, 14]. Under this paradigm, we divide them into", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 231, + 390, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 390, + 243 + ], + "score": 1.0, + "content": "three kinds: (1) proposal-based, (2) rule-based and (3) attention-based.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 247, + 505, + 280 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 506, + 259 + ], + "score": 1.0, + "content": "(1) Proposal-based methods rely on the prediction of the RPN or detection network to find foreground", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "regions, e.g., Chen et al. [6] and Li et al. [28] propose to distilling regions predicted by RPN [36],", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 268, + 443, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 443, + 282 + ], + "score": 1.0, + "content": "Dai et al. [11] proposes GI scores to locate controversial predictions for distillation.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 285, + 505, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 284, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 507, + 299 + ], + "score": 1.0, + "content": "(2) Rule-based methods rely on designed rules that can be inflexible and hyper-parameters inefficient,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 296, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 309 + ], + "score": 1.0, + "content": "e.g., Wang et al. [45] distill assigned regions where anchor and ground-truth have a large IoU, Guo et", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 307, + 440, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 440, + 320 + ], + "score": 1.0, + "content": "al. [14] distill foreground and background regions separately with different factors.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 323, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "score": 1.0, + "content": "(3) Attention-based methods rely on activations to locate discriminative areas, yet they do not direct", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "score": 1.0, + "content": "to knowledge that the student needs. Only a recent work from Zhang et al. [51] considers attention,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 345, + 384, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 384, + 358 + ], + "score": 1.0, + "content": "they build the spatial-channel-wise attention to weigh the distillation.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 375 + ], + "score": 1.0, + "content": "To overcome the above limitations, we explore the instance-conditional knowledge retrieval formu-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 371, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 385 + ], + "score": 1.0, + "content": "lated by a decoder to explicitly search for useful knowledge. Like other methods, ICD does not have", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 384, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 481, + 396 + ], + "score": 1.0, + "content": "extra cost during inference or use extra data (besides existing labels and a pretrained teacher).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 107, + 419, + 240, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 241, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 241, + 433 + ], + "score": 1.0, + "content": "2.2 Conditional Computation", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 564 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "Conditional computation is widely adopted to infer contents on a given condition. Our study mostly", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "focuses on how to identify visual contents given an instance as a condition. This is usually formulated", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 466, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 477 + ], + "score": 1.0, + "content": "as query an instance on the image, e.g., visual question answer [1, 2] and image-text matching [26]", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "queries information and regions specified by natural language. Besides query by language, other types", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 485, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 104, + 485, + 507, + 501 + ], + "score": 1.0, + "content": "of query are proposed in recent years. For example, DETR [5] queries on fixed proposal embeddings,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 497, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 511 + ], + "score": 1.0, + "content": "Chen et al. [8] encodes points as queries to facilitate weakly-supervised learning. These works adopt", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "score": 1.0, + "content": "transformer decoder to infer upon global receptive fields that cover all visual contents, yet they usually", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 520, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "rely on cascaded decoders that are costly for training. From another perspective, CondInst [41] and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "SOLOv2 [46] generate queries based on network predictions and achieves great performance on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "score": 1.0, + "content": "instance segmentation. Different from them, this work adopts the query-based approach to retrieve", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 553, + 336, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 336, + 564 + ], + "score": 1.0, + "content": "knowledge and build query based on annotated instances.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 108, + 587, + 202, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 204, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 204, + 602 + ], + "score": 1.0, + "content": "2.3 Object Detection", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 506, + 626 + ], + "score": 1.0, + "content": "Object detection has been developed rapidly. Modern detectors are roughly divided into two-stage or", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "one-stage detectors. Two-stage detectors usually adopt Region Proposal Network (RPN) to generate", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "initial rough predictions and refine them with detection heads, the typical example is Faster R-CNN", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "[36]. On the contrary, one-stage detectors directly predict on the feature map, which are usually", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "score": 1.0, + "content": "faster, they include RetinaNet [30], FCOS [42]. Besides of this rough division, many extensions are", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "score": 1.0, + "content": "introduced in recent years, e.g., extension to instance segmentation [41, 46, 16], anchor-free models", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "[42, 25] and end-to-end detection [5, 44, 20]. Among these works, multi-scale features are usually", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "adopted to enhance performance, e.g., FPN [29], which is considered as a typical case for our study.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "To generalize to various methods, the proposed method distills the intermediate features and does not", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 711, + 240, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 240, + 723 + ], + "score": 1.0, + "content": "rely on detector-specific designs.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 41.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 309, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 202, + 84 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 203, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 203, + 86 + ], + "score": 1.0, + "content": "2 Related Works", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 102, + 227, + 114 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 229, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 229, + 117 + ], + "score": 1.0, + "content": "2.1 Knowledge Distillation", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 127, + 505, + 182 + ], + "lines": [ + { + "bbox": [ + 106, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "Knowledge distillation aims to transfer knowledge from a strong teacher to a weaker student network", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 138, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 151 + ], + "score": 1.0, + "content": "to facilitate supervised learning. The teacher is usually a large pretrained network, who provides", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "smoother supervision and more hints on visual concepts, that improves the training quality and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 172 + ], + "score": 1.0, + "content": "convergence speed [49, 9]. KD for image classification has been studied for years, they are usually", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "categorized into three types [13]: response-based [19], feature-based [37, 43] and relation-based [32].", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 126, + 506, + 183 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 505, + 242 + ], + "lines": [ + { + "bbox": [ + 106, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "Among them, feature-based distillation over multi-scale features is adopted from most of detection", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "KD works, to deal with knowledge among multiple instances in different regions. Most of these works", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 505, + 222 + ], + "score": 1.0, + "content": "can be formulated as region selection for distillation, where foreground-background unbalancing is", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 233 + ], + "score": 1.0, + "content": "considered as a key problem in some studies [45, 51, 14]. Under this paradigm, we divide them into", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 231, + 390, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 390, + 243 + ], + "score": 1.0, + "content": "three kinds: (1) proposal-based, (2) rule-based and (3) attention-based.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 186, + 505, + 243 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 247, + 505, + 280 + ], + "lines": [ + { + "bbox": [ + 105, + 246, + 506, + 259 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 506, + 259 + ], + "score": 1.0, + "content": "(1) Proposal-based methods rely on the prediction of the RPN or detection network to find foreground", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 271 + ], + "score": 1.0, + "content": "regions, e.g., Chen et al. [6] and Li et al. [28] propose to distilling regions predicted by RPN [36],", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 268, + 443, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 443, + 282 + ], + "score": 1.0, + "content": "Dai et al. [11] proposes GI scores to locate controversial predictions for distillation.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 246, + 506, + 282 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 285, + 505, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 284, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 507, + 299 + ], + "score": 1.0, + "content": "(2) Rule-based methods rely on designed rules that can be inflexible and hyper-parameters inefficient,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 296, + 506, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 309 + ], + "score": 1.0, + "content": "e.g., Wang et al. [45] distill assigned regions where anchor and ground-truth have a large IoU, Guo et", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 307, + 440, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 440, + 320 + ], + "score": 1.0, + "content": "al. [14] distill foreground and background regions separately with different factors.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 284, + 507, + 320 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 323, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 506, + 336 + ], + "score": 1.0, + "content": "(3) Attention-based methods rely on activations to locate discriminative areas, yet they do not direct", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 507, + 347 + ], + "score": 1.0, + "content": "to knowledge that the student needs. Only a recent work from Zhang et al. [51] considers attention,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 345, + 384, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 384, + 358 + ], + "score": 1.0, + "content": "they build the spatial-channel-wise attention to weigh the distillation.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 322, + 507, + 358 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 506, + 375 + ], + "score": 1.0, + "content": "To overcome the above limitations, we explore the instance-conditional knowledge retrieval formu-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 371, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 385 + ], + "score": 1.0, + "content": "lated by a decoder to explicitly search for useful knowledge. Like other methods, ICD does not have", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 384, + 481, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 481, + 396 + ], + "score": 1.0, + "content": "extra cost during inference or use extra data (besides existing labels and a pretrained teacher).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 360, + 506, + 396 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 419, + 240, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 417, + 241, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 241, + 433 + ], + "score": 1.0, + "content": "2.2 Conditional Computation", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 505, + 564 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "Conditional computation is widely adopted to infer contents on a given condition. Our study mostly", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 467 + ], + "score": 1.0, + "content": "focuses on how to identify visual contents given an instance as a condition. This is usually formulated", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 466, + 505, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 477 + ], + "score": 1.0, + "content": "as query an instance on the image, e.g., visual question answer [1, 2] and image-text matching [26]", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "queries information and regions specified by natural language. Besides query by language, other types", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 485, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 104, + 485, + 507, + 501 + ], + "score": 1.0, + "content": "of query are proposed in recent years. For example, DETR [5] queries on fixed proposal embeddings,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 497, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 511 + ], + "score": 1.0, + "content": "Chen et al. [8] encodes points as queries to facilitate weakly-supervised learning. These works adopt", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 521 + ], + "score": 1.0, + "content": "transformer decoder to infer upon global receptive fields that cover all visual contents, yet they usually", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 520, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "rely on cascaded decoders that are costly for training. From another perspective, CondInst [41] and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "SOLOv2 [46] generate queries based on network predictions and achieves great performance on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 555 + ], + "score": 1.0, + "content": "instance segmentation. Different from them, this work adopts the query-based approach to retrieve", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 553, + 336, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 336, + 564 + ], + "score": 1.0, + "content": "knowledge and build query based on annotated instances.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30, + "bbox_fs": [ + 104, + 444, + 507, + 564 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 587, + 202, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 586, + 204, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 204, + 602 + ], + "score": 1.0, + "content": "2.3 Object Detection", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 611, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 506, + 626 + ], + "score": 1.0, + "content": "Object detection has been developed rapidly. Modern detectors are roughly divided into two-stage or", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 506, + 636 + ], + "score": 1.0, + "content": "one-stage detectors. Two-stage detectors usually adopt Region Proposal Network (RPN) to generate", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "initial rough predictions and refine them with detection heads, the typical example is Faster R-CNN", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "[36]. On the contrary, one-stage detectors directly predict on the feature map, which are usually", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "score": 1.0, + "content": "faster, they include RetinaNet [30], FCOS [42]. Besides of this rough division, many extensions are", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "score": 1.0, + "content": "introduced in recent years, e.g., extension to instance segmentation [41, 46, 16], anchor-free models", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "[42, 25] and end-to-end detection [5, 44, 20]. Among these works, multi-scale features are usually", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "adopted to enhance performance, e.g., FPN [29], which is considered as a typical case for our study.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "To generalize to various methods, the proposed method distills the intermediate features and does not", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 711, + 240, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 240, + 723 + ], + "score": 1.0, + "content": "rely on detector-specific designs.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 611, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 71, + 166, + 84 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 168, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 168, + 86 + ], + "score": 1.0, + "content": "3 Method", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 95, + 171, + 107 + ], + "lines": [ + { + "bbox": [ + 105, + 94, + 172, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 172, + 109 + ], + "score": 1.0, + "content": "3.1 Overview", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 115, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "As discussed in former studies [14, 51], useful knowledge for detection distributes unevenly in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "intermediate features. To facilitate KD, we propose to transfer instance-conditional knowledge", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 137, + 470, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 289, + 151 + ], + "score": 1.0, + "content": "between student and teacher network, termed", + "type": "text" + }, + { + "bbox": [ + 290, + 138, + 302, + 150 + ], + "score": 0.9, + "content": "\\kappa _ { i } ^ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 137, + 320, + 151 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 321, + 137, + 334, + 150 + ], + "score": 0.9, + "content": "\\kappa _ { i } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 137, + 419, + 151 + ], + "score": 1.0, + "content": "corresponding to the", + "type": "text" + }, + { + "bbox": [ + 420, + 138, + 430, + 149 + ], + "score": 0.88, + "content": "i _ { \\mathrm { t h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 137, + 470, + 151 + ], + "score": 1.0, + "content": "instance:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "interline_equation", + "bbox": [ + 251, + 155, + 360, + 190 + ], + "lines": [ + { + "bbox": [ + 251, + 155, + 360, + 190 + ], + "spans": [ + { + "bbox": [ + 251, + 155, + 360, + 190 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { d i s t i l l } = \\sum _ { i = 1 } ^ { N } \\mathcal { L } _ { d } ( \\kappa _ { i } ^ { S } , \\kappa _ { i } ^ { T } )", + "type": "interline_equation", + "image_path": "e013119ab86b21e33d2a565254641b211d57d7eefca001f262c88363fa9362eb.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 251, + 155, + 360, + 172.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 251, + 172.5, + 360, + 190.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 201, + 506, + 236 + ], + "lines": [ + { + "bbox": [ + 104, + 199, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 104, + 199, + 301, + 215 + ], + "score": 1.0, + "content": "The knowledge towards teacher’s representations", + "type": "text" + }, + { + "bbox": [ + 301, + 202, + 311, + 212 + ], + "score": 0.83, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 199, + 367, + 215 + ], + "score": 1.0, + "content": "and condition", + "type": "text" + }, + { + "bbox": [ + 367, + 203, + 376, + 213 + ], + "score": 0.84, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 199, + 443, + 215 + ], + "score": 1.0, + "content": "is formulated as", + "type": "text" + }, + { + "bbox": [ + 443, + 200, + 505, + 214 + ], + "score": 0.93, + "content": "\\kappa _ { i } ^ { \\mathcal { T } } = \\mathcal { G } ( \\mathcal { T } , y _ { i } )", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 109, + 211, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 109, + 213, + 122, + 226 + ], + "score": 0.87, + "content": "{ \\kappa } _ { i } ^ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 211, + 194, + 227 + ], + "score": 1.0, + "content": "similarly), where", + "type": "text" + }, + { + "bbox": [ + 194, + 214, + 201, + 224 + ], + "score": 0.84, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 211, + 506, + 227 + ], + "score": 1.0, + "content": "is the instance-conditional decoding module, optimized by an auxiliary loss", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 223, + 367, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 367, + 237 + ], + "score": 1.0, + "content": "illustrated in Sec. 3.3. The overall framework is shown in Fig. 2.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 105, + 240, + 504, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "In the following sections, we will introduce the instance-conditional knowledge (Sec. 3.2), describe", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 252, + 450, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 450, + 264 + ], + "score": 1.0, + "content": "the auxiliary task design (Sec. 3.3), and discuss the knowledge transferring (Sec. 3.4).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 107, + 276, + 267, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 268, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 268, + 290 + ], + "score": 1.0, + "content": "3.2 Instance-conditional Knowledge", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 296, + 297, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 298, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 298, + 307 + ], + "score": 1.0, + "content": "In this section, we elaborate the instance-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 307, + 298, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 232, + 319 + ], + "score": 1.0, + "content": "conditional decoding module", + "type": "text" + }, + { + "bbox": [ + 232, + 308, + 240, + 318 + ], + "score": 0.79, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 307, + 298, + 319 + ], + "score": 1.0, + "content": ", which com-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 317, + 298, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 259, + 330 + ], + "score": 1.0, + "content": "putes instance-conditional knowledge", + "type": "text" + }, + { + "bbox": [ + 260, + 317, + 273, + 330 + ], + "score": 0.9, + "content": "\\kappa _ { i } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 317, + 298, + 330 + ], + "score": 1.0, + "content": "from", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 329, + 297, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 297, + 341 + ], + "score": 1.0, + "content": "(1) unconditional knowledge given (2) instance", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 340, + 153, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 153, + 352 + ], + "score": 1.0, + "content": "conditions.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 297, + 490 + ], + "lines": [ + { + "bbox": [ + 106, + 356, + 297, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 237, + 369 + ], + "score": 1.0, + "content": "(1) The unconditional knowledge", + "type": "text" + }, + { + "bbox": [ + 238, + 357, + 247, + 367 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 356, + 297, + 369 + ], + "score": 1.0, + "content": ", symbolizes", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 367, + 298, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 298, + 379 + ], + "score": 1.0, + "content": "all available information from the teacher detec-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 378, + 297, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 297, + 390 + ], + "score": 1.0, + "content": "tor. Since modern detectors commonly involve", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 389, + 299, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 299, + 401 + ], + "score": 1.0, + "content": "a feature pyramid network (FPN) [29] to ex-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 399, + 298, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 298, + 412 + ], + "score": 1.0, + "content": "tract rich multi-scale representations, we present", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 410, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 235, + 424 + ], + "score": 1.0, + "content": "multi-scale representations as", + "type": "text" + }, + { + "bbox": [ + 235, + 411, + 296, + 423 + ], + "score": 0.92, + "content": "\\mathcal T ~ = ~ \\bar \\{ X _ { p } ~ \\in ~ ", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 419, + 299, + 439 + ], + "spans": [ + { + "bbox": [ + 107, + 422, + 178, + 436 + ], + "score": 0.91, + "content": "\\mathbb { R } ^ { D \\times H _ { p } \\times W _ { p } } \\} _ { p \\in \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 419, + 208, + 439 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 208, + 424, + 218, + 434 + ], + "score": 0.8, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 419, + 299, + 439 + ], + "score": 1.0, + "content": "signifies the spatial", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 434, + 299, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 180, + 446 + ], + "score": 1.0, + "content": "resolutions while", + "type": "text" + }, + { + "bbox": [ + 180, + 435, + 190, + 444 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 434, + 299, + 446 + ], + "score": 1.0, + "content": "is the channel dimension.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 445, + 298, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 298, + 457 + ], + "score": 1.0, + "content": "By concatenating representations at different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 456, + 297, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 297, + 468 + ], + "score": 1.0, + "content": "scales along the spatial dimension, we obtain", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 462, + 299, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 161, + 478 + ], + "score": 0.9, + "content": "\\mathbf { A } ^ { \\mathcal { T } } \\in \\mathbb { R } ^ { L \\times \\mathbf { \\smile } }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 462, + 193, + 484 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 193, + 466, + 271, + 481 + ], + "score": 0.92, + "content": "\\begin{array} { r } { L = \\sum _ { p \\in \\mathcal { P } } H _ { p } W _ { p } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 462, + 299, + 484 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 479, + 272, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 272, + 491 + ], + "score": 1.0, + "content": "sum of total pixels number across scales.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 23.5 + }, + { + "type": "image", + "bbox": [ + 307, + 282, + 504, + 419 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 307, + 282, + 504, + 419 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 504, + 419 + ], + "score": 0.972, + "type": "image", + "image_path": "25e2b995f25211ec29c9dcb5e6072f7c1761f226e166013e18fe0d98ea312278.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 307, + 282, + 504, + 295.7 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 307, + 295.7, + 504, + 309.4 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 307, + 309.4, + 504, + 323.09999999999997 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 307, + 323.09999999999997, + 504, + 336.79999999999995 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 307, + 336.79999999999995, + 504, + 350.49999999999994 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 307, + 350.49999999999994, + 504, + 364.19999999999993 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 307, + 364.19999999999993, + 504, + 377.8999999999999 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 307, + 377.8999999999999, + 504, + 391.5999999999999 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 307, + 391.5999999999999, + 504, + 405.2999999999999 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 307, + 405.2999999999999, + 504, + 418.9999999999999 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 429, + 505, + 496 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 506, + 442 + ], + "score": 1.0, + "content": "Figure 2: We propose a decoding module to re-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 441, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 505, + 452 + ], + "score": 1.0, + "content": "trieve knowledge via query-based attention, where", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "instance annotations are encoded as a query. An", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 462, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 506, + 474 + ], + "score": 1.0, + "content": "auxiliary task is proposed to optimize the decod-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 303, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 303, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "ing module and the feature-based distillation loss", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "weighted by the attention is used to update student.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 296, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 494, + 298, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 298, + 509 + ], + "score": 1.0, + "content": "(2) The instance condition, originally describing", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 506, + 297, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 275, + 519 + ], + "score": 1.0, + "content": "a human-observed object, is denoted by", + "type": "text" + }, + { + "bbox": [ + 275, + 506, + 297, + 517 + ], + "score": 0.87, + "content": "{ \\mathcal { V } } =", + "type": "inline_equation" + } + ], + "index": 47 + } + ], + "index": 46.5 + }, + { + "type": "text", + "bbox": [ + 108, + 518, + 503, + 550 + ], + "lines": [ + { + "bbox": [ + 107, + 514, + 507, + 533 + ], + "spans": [ + { + "bbox": [ + 107, + 516, + 139, + 529 + ], + "score": 0.92, + "content": "\\{ \\mathfrak { y } _ { i } \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 514, + 172, + 533 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 172, + 518, + 183, + 527 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 514, + 290, + 533 + ], + "score": 1.0, + "content": "is the object number and", + "type": "text" + }, + { + "bbox": [ + 290, + 517, + 345, + 529 + ], + "score": 0.9, + "content": "\\mathsf { y } _ { i } = \\left( c _ { i } , \\mathbf { b } _ { i } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 514, + 450, + 533 + ], + "score": 1.0, + "content": "is the annotation for the", + "type": "text" + }, + { + "bbox": [ + 450, + 518, + 455, + 527 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 514, + 507, + 533 + ], + "score": 1.0, + "content": "-th instance,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 527, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 182, + 542 + ], + "score": 1.0, + "content": "including category", + "type": "text" + }, + { + "bbox": [ + 182, + 530, + 191, + 539 + ], + "score": 0.83, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 527, + 261, + 542 + ], + "score": 1.0, + "content": "and box location", + "type": "text" + }, + { + "bbox": [ + 261, + 529, + 343, + 540 + ], + "score": 0.9, + "content": "\\mathbf { b } _ { i } = \\left( x _ { i } , y _ { i } , w _ { i } , h _ { i } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 527, + 506, + 542 + ], + "score": 1.0, + "content": "which specifies the localization and size", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 538, + 158, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 158, + 551 + ], + "score": 1.0, + "content": "information.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49 + }, + { + "type": "text", + "bbox": [ + 104, + 555, + 504, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "To produce learnable embeddings for each instance, the annotation is mapped to a query feature", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 566, + 456, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 133, + 579 + ], + "score": 1.0, + "content": "vector", + "type": "text" + }, + { + "bbox": [ + 133, + 568, + 144, + 578 + ], + "score": 0.85, + "content": "\\mathbf { q } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 566, + 456, + 579 + ], + "score": 1.0, + "content": "in the hidden space, which specifies a condition to collect desired knowledge:", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51.5 + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 583, + 361, + 599 + ], + "lines": [ + { + "bbox": [ + 249, + 583, + 361, + 599 + ], + "spans": [ + { + "bbox": [ + 249, + 583, + 361, + 599 + ], + "score": 0.92, + "content": "\\mathbf { q } _ { i } = \\mathcal { F } _ { q } ( \\mathcal { E } ( \\mathbf { y } _ { i } ) ) , \\ \\mathbf { q } _ { i } \\in \\mathbb { R } ^ { D }", + "type": "interline_equation", + "image_path": "c40dedb2df7d946a4b77f9cbd45eddc8f8cd4551e0e35190f0aa46660f3d884e.jpg" + } + ] + } + ], + "index": 53, + "virtual_lines": [ + { + "bbox": [ + 249, + 583, + 361, + 599 + ], + "spans": [], + "index": 53 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 603, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 602, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 132, + 617 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 604, + 150, + 615 + ], + "score": 0.9, + "content": "\\mathcal { E } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 602, + 380, + 617 + ], + "score": 1.0, + "content": "is a instance encoding function (detailed in in Sec. 3) and", + "type": "text" + }, + { + "bbox": [ + 380, + 604, + 393, + 616 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 602, + 505, + 617 + ], + "score": 1.0, + "content": "is a Multi-Layer Perception", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 615, + 173, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 173, + 627 + ], + "score": 1.0, + "content": "network (MLP).", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 54.5 + }, + { + "type": "text", + "bbox": [ + 106, + 630, + 505, + 666 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 219, + 644 + ], + "score": 1.0, + "content": "We retrieve knowledge from", + "type": "text" + }, + { + "bbox": [ + 219, + 631, + 229, + 641 + ], + "score": 0.83, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 630, + 253, + 644 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 253, + 632, + 264, + 642 + ], + "score": 0.86, + "content": "\\mathbf { q } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 630, + 505, + 644 + ], + "score": 1.0, + "content": "by measuring responses of correlation. This is formulated by", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 232, + 654 + ], + "score": 1.0, + "content": "dot-product attention [43] with", + "type": "text" + }, + { + "bbox": [ + 232, + 642, + 244, + 652 + ], + "score": 0.83, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "concurrent heads in a query-key attention manner. In which each", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 652, + 504, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 127, + 667 + ], + "score": 1.0, + "content": "head", + "type": "text" + }, + { + "bbox": [ + 128, + 654, + 134, + 664 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 652, + 270, + 667 + ], + "score": 1.0, + "content": "corresponds to three linear layers", + "type": "text" + }, + { + "bbox": [ + 270, + 652, + 346, + 667 + ], + "score": 0.93, + "content": "( \\mathcal { F } _ { j } ^ { k } , \\mathcal { F } _ { j } ^ { q } , \\mathcal { F } _ { j } ^ { v } ) \\mathop { w . r . t . }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 652, + 504, + 667 + ], + "score": 1.0, + "content": ". the key, query and value construction.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 57 + }, + { + "type": "text", + "bbox": [ + 107, + 671, + 505, + 724 + ], + "lines": [ + { + "bbox": [ + 105, + 670, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 174, + 684 + ], + "score": 1.0, + "content": "The key feature", + "type": "text" + }, + { + "bbox": [ + 174, + 671, + 190, + 685 + ], + "score": 0.91, + "content": "\\mathrm { K } _ { j } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 670, + 407, + 684 + ], + "score": 1.0, + "content": "is computed by projecting teacher’s representations", + "type": "text" + }, + { + "bbox": [ + 408, + 671, + 424, + 682 + ], + "score": 0.87, + "content": "\\mathrm { A } ^ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 670, + 505, + 684 + ], + "score": 1.0, + "content": "with the positional", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 682, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 189, + 699 + ], + "score": 1.0, + "content": "embeddings [43, 5]", + "type": "text" + }, + { + "bbox": [ + 189, + 684, + 236, + 696 + ], + "score": 0.9, + "content": "\\mathrm { P } \\in \\mathbb { R } ^ { L \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 682, + 317, + 699 + ], + "score": 1.0, + "content": "as in Eq. 3, where", + "type": "text" + }, + { + "bbox": [ + 317, + 685, + 334, + 698 + ], + "score": 0.91, + "content": "\\mathcal { F } _ { p e }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 682, + 506, + 699 + ], + "score": 1.0, + "content": "denotes a linear projection over position", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 695, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 236, + 711 + ], + "score": 1.0, + "content": "embeddings. The value feature", + "type": "text" + }, + { + "bbox": [ + 236, + 696, + 252, + 711 + ], + "score": 0.9, + "content": "\\mathrm { V } _ { j } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 695, + 297, + 711 + ], + "score": 1.0, + "content": "and query", + "type": "text" + }, + { + "bbox": [ + 297, + 699, + 311, + 710 + ], + "score": 0.88, + "content": "{ \\bf q } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 695, + 505, + 711 + ], + "score": 1.0, + "content": "is projected by linear mappings to a sub-space", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 708, + 505, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 127, + 725 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 710, + 172, + 723 + ], + "score": 0.93, + "content": "d = D / M", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 708, + 214, + 725 + ], + "score": 1.0, + "content": "channels,", + "type": "text" + }, + { + "bbox": [ + 215, + 711, + 228, + 724 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { j } ^ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 708, + 244, + 725 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 244, + 710, + 259, + 721 + ], + "score": 0.84, + "content": "\\mathrm { A } ^ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 708, + 278, + 725 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 279, + 711, + 292, + 725 + ], + "score": 0.91, + "content": "\\mathcal { F } _ { j } ^ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 708, + 314, + 725 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 314, + 712, + 325, + 722 + ], + "score": 0.85, + "content": "\\mathbf { q } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 708, + 505, + 725 + ], + "score": 1.0, + "content": "respectively, as shown in Eq. 4. At last, an", + "type": "text" + } + ], + "index": 62 + } + ], + "index": 60.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 11, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 71, + 166, + 84 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 168, + 86 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 168, + 86 + ], + "score": 1.0, + "content": "3 Method", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 95, + 171, + 107 + ], + "lines": [ + { + "bbox": [ + 105, + 94, + 172, + 109 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 172, + 109 + ], + "score": 1.0, + "content": "3.1 Overview", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 115, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "score": 1.0, + "content": "As discussed in former studies [14, 51], useful knowledge for detection distributes unevenly in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 140 + ], + "score": 1.0, + "content": "intermediate features. To facilitate KD, we propose to transfer instance-conditional knowledge", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 137, + 470, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 289, + 151 + ], + "score": 1.0, + "content": "between student and teacher network, termed", + "type": "text" + }, + { + "bbox": [ + 290, + 138, + 302, + 150 + ], + "score": 0.9, + "content": "\\kappa _ { i } ^ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 137, + 320, + 151 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 321, + 137, + 334, + 150 + ], + "score": 0.9, + "content": "\\kappa _ { i } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 137, + 419, + 151 + ], + "score": 1.0, + "content": "corresponding to the", + "type": "text" + }, + { + "bbox": [ + 420, + 138, + 430, + 149 + ], + "score": 0.88, + "content": "i _ { \\mathrm { t h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 137, + 470, + 151 + ], + "score": 1.0, + "content": "instance:", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 116, + 505, + 151 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 251, + 155, + 360, + 190 + ], + "lines": [ + { + "bbox": [ + 251, + 155, + 360, + 190 + ], + "spans": [ + { + "bbox": [ + 251, + 155, + 360, + 190 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { d i s t i l l } = \\sum _ { i = 1 } ^ { N } \\mathcal { L } _ { d } ( \\kappa _ { i } ^ { S } , \\kappa _ { i } ^ { T } )", + "type": "interline_equation", + "image_path": "e013119ab86b21e33d2a565254641b211d57d7eefca001f262c88363fa9362eb.jpg" + } + ] + } + ], + "index": 5.5, + "virtual_lines": [ + { + "bbox": [ + 251, + 155, + 360, + 172.5 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 251, + 172.5, + 360, + 190.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 201, + 506, + 236 + ], + "lines": [ + { + "bbox": [ + 104, + 199, + 505, + 215 + ], + "spans": [ + { + "bbox": [ + 104, + 199, + 301, + 215 + ], + "score": 1.0, + "content": "The knowledge towards teacher’s representations", + "type": "text" + }, + { + "bbox": [ + 301, + 202, + 311, + 212 + ], + "score": 0.83, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 199, + 367, + 215 + ], + "score": 1.0, + "content": "and condition", + "type": "text" + }, + { + "bbox": [ + 367, + 203, + 376, + 213 + ], + "score": 0.84, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 199, + 443, + 215 + ], + "score": 1.0, + "content": "is formulated as", + "type": "text" + }, + { + "bbox": [ + 443, + 200, + 505, + 214 + ], + "score": 0.93, + "content": "\\kappa _ { i } ^ { \\mathcal { T } } = \\mathcal { G } ( \\mathcal { T } , y _ { i } )", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 109, + 211, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 109, + 213, + 122, + 226 + ], + "score": 0.87, + "content": "{ \\kappa } _ { i } ^ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 123, + 211, + 194, + 227 + ], + "score": 1.0, + "content": "similarly), where", + "type": "text" + }, + { + "bbox": [ + 194, + 214, + 201, + 224 + ], + "score": 0.84, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 211, + 506, + 227 + ], + "score": 1.0, + "content": "is the instance-conditional decoding module, optimized by an auxiliary loss", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 223, + 367, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 367, + 237 + ], + "score": 1.0, + "content": "illustrated in Sec. 3.3. The overall framework is shown in Fig. 2.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 104, + 199, + 506, + 237 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 240, + 504, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "In the following sections, we will introduce the instance-conditional knowledge (Sec. 3.2), describe", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 252, + 450, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 450, + 264 + ], + "score": 1.0, + "content": "the auxiliary task design (Sec. 3.3), and discuss the knowledge transferring (Sec. 3.4).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 240, + 505, + 264 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 276, + 267, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 268, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 268, + 290 + ], + "score": 1.0, + "content": "3.2 Instance-conditional Knowledge", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 296, + 297, + 351 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 298, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 298, + 307 + ], + "score": 1.0, + "content": "In this section, we elaborate the instance-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 307, + 298, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 232, + 319 + ], + "score": 1.0, + "content": "conditional decoding module", + "type": "text" + }, + { + "bbox": [ + 232, + 308, + 240, + 318 + ], + "score": 0.79, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 307, + 298, + 319 + ], + "score": 1.0, + "content": ", which com-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 317, + 298, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 259, + 330 + ], + "score": 1.0, + "content": "putes instance-conditional knowledge", + "type": "text" + }, + { + "bbox": [ + 260, + 317, + 273, + 330 + ], + "score": 0.9, + "content": "\\kappa _ { i } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 317, + 298, + 330 + ], + "score": 1.0, + "content": "from", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 329, + 297, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 297, + 341 + ], + "score": 1.0, + "content": "(1) unconditional knowledge given (2) instance", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 340, + 153, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 153, + 352 + ], + "score": 1.0, + "content": "conditions.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 296, + 298, + 352 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 356, + 297, + 490 + ], + "lines": [ + { + "bbox": [ + 106, + 356, + 297, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 237, + 369 + ], + "score": 1.0, + "content": "(1) The unconditional knowledge", + "type": "text" + }, + { + "bbox": [ + 238, + 357, + 247, + 367 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 356, + 297, + 369 + ], + "score": 1.0, + "content": ", symbolizes", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 367, + 298, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 298, + 379 + ], + "score": 1.0, + "content": "all available information from the teacher detec-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 378, + 297, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 378, + 297, + 390 + ], + "score": 1.0, + "content": "tor. Since modern detectors commonly involve", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 389, + 299, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 299, + 401 + ], + "score": 1.0, + "content": "a feature pyramid network (FPN) [29] to ex-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 399, + 298, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 298, + 412 + ], + "score": 1.0, + "content": "tract rich multi-scale representations, we present", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 410, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 235, + 424 + ], + "score": 1.0, + "content": "multi-scale representations as", + "type": "text" + }, + { + "bbox": [ + 235, + 411, + 296, + 423 + ], + "score": 0.92, + "content": "\\mathcal T ~ = ~ \\bar \\{ X _ { p } ~ \\in ~ ", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 419, + 299, + 439 + ], + "spans": [ + { + "bbox": [ + 107, + 422, + 178, + 436 + ], + "score": 0.91, + "content": "\\mathbb { R } ^ { D \\times H _ { p } \\times W _ { p } } \\} _ { p \\in \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 419, + 208, + 439 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 208, + 424, + 218, + 434 + ], + "score": 0.8, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 419, + 299, + 439 + ], + "score": 1.0, + "content": "signifies the spatial", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 434, + 299, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 180, + 446 + ], + "score": 1.0, + "content": "resolutions while", + "type": "text" + }, + { + "bbox": [ + 180, + 435, + 190, + 444 + ], + "score": 0.82, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 434, + 299, + 446 + ], + "score": 1.0, + "content": "is the channel dimension.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 445, + 298, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 298, + 457 + ], + "score": 1.0, + "content": "By concatenating representations at different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 456, + 297, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 297, + 468 + ], + "score": 1.0, + "content": "scales along the spatial dimension, we obtain", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 462, + 299, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 161, + 478 + ], + "score": 0.9, + "content": "\\mathbf { A } ^ { \\mathcal { T } } \\in \\mathbb { R } ^ { L \\times \\mathbf { \\smile } }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 462, + 193, + 484 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 193, + 466, + 271, + 481 + ], + "score": 0.92, + "content": "\\begin{array} { r } { L = \\sum _ { p \\in \\mathcal { P } } H _ { p } W _ { p } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 462, + 299, + 484 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 479, + 272, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 272, + 491 + ], + "score": 1.0, + "content": "sum of total pixels number across scales.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 356, + 299, + 491 + ] + }, + { + "type": "image", + "bbox": [ + 307, + 282, + 504, + 419 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 307, + 282, + 504, + 419 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 504, + 419 + ], + "score": 0.972, + "type": "image", + "image_path": "25e2b995f25211ec29c9dcb5e6072f7c1761f226e166013e18fe0d98ea312278.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 307, + 282, + 504, + 295.7 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 307, + 295.7, + 504, + 309.4 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 307, + 309.4, + 504, + 323.09999999999997 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 307, + 323.09999999999997, + 504, + 336.79999999999995 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 307, + 336.79999999999995, + 504, + 350.49999999999994 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 307, + 350.49999999999994, + 504, + 364.19999999999993 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 307, + 364.19999999999993, + 504, + 377.8999999999999 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 307, + 377.8999999999999, + 504, + 391.5999999999999 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 307, + 391.5999999999999, + 504, + 405.2999999999999 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 307, + 405.2999999999999, + 504, + 418.9999999999999 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 429, + 505, + 496 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 506, + 442 + ], + "score": 1.0, + "content": "Figure 2: We propose a decoding module to re-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 441, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 505, + 452 + ], + "score": 1.0, + "content": "trieve knowledge via query-based attention, where", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "instance annotations are encoded as a query. An", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 462, + 506, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 506, + 474 + ], + "score": 1.0, + "content": "auxiliary task is proposed to optimize the decod-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 303, + 473, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 303, + 473, + 505, + 485 + ], + "score": 1.0, + "content": "ing module and the feature-based distillation loss", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 484, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 505, + 496 + ], + "score": 1.0, + "content": "weighted by the attention is used to update student.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 42.5 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 296, + 517 + ], + "lines": [ + { + "bbox": [ + 105, + 494, + 298, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 298, + 509 + ], + "score": 1.0, + "content": "(2) The instance condition, originally describing", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 506, + 297, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 275, + 519 + ], + "score": 1.0, + "content": "a human-observed object, is denoted by", + "type": "text" + }, + { + "bbox": [ + 275, + 506, + 297, + 517 + ], + "score": 0.87, + "content": "{ \\mathcal { V } } =", + "type": "inline_equation" + } + ], + "index": 47 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 494, + 298, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 518, + 503, + 550 + ], + "lines": [ + { + "bbox": [ + 107, + 514, + 507, + 533 + ], + "spans": [ + { + "bbox": [ + 107, + 516, + 139, + 529 + ], + "score": 0.92, + "content": "\\{ \\mathfrak { y } _ { i } \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 514, + 172, + 533 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 172, + 518, + 183, + 527 + ], + "score": 0.83, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 514, + 290, + 533 + ], + "score": 1.0, + "content": "is the object number and", + "type": "text" + }, + { + "bbox": [ + 290, + 517, + 345, + 529 + ], + "score": 0.9, + "content": "\\mathsf { y } _ { i } = \\left( c _ { i } , \\mathbf { b } _ { i } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 514, + 450, + 533 + ], + "score": 1.0, + "content": "is the annotation for the", + "type": "text" + }, + { + "bbox": [ + 450, + 518, + 455, + 527 + ], + "score": 0.76, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 514, + 507, + 533 + ], + "score": 1.0, + "content": "-th instance,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 527, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 182, + 542 + ], + "score": 1.0, + "content": "including category", + "type": "text" + }, + { + "bbox": [ + 182, + 530, + 191, + 539 + ], + "score": 0.83, + "content": "c _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 527, + 261, + 542 + ], + "score": 1.0, + "content": "and box location", + "type": "text" + }, + { + "bbox": [ + 261, + 529, + 343, + 540 + ], + "score": 0.9, + "content": "\\mathbf { b } _ { i } = \\left( x _ { i } , y _ { i } , w _ { i } , h _ { i } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 527, + 506, + 542 + ], + "score": 1.0, + "content": "which specifies the localization and size", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 538, + 158, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 158, + 551 + ], + "score": 1.0, + "content": "information.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49, + "bbox_fs": [ + 105, + 514, + 507, + 551 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 555, + 504, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "To produce learnable embeddings for each instance, the annotation is mapped to a query feature", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 566, + 456, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 133, + 579 + ], + "score": 1.0, + "content": "vector", + "type": "text" + }, + { + "bbox": [ + 133, + 568, + 144, + 578 + ], + "score": 0.85, + "content": "\\mathbf { q } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 566, + 456, + 579 + ], + "score": 1.0, + "content": "in the hidden space, which specifies a condition to collect desired knowledge:", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51.5, + "bbox_fs": [ + 105, + 555, + 505, + 579 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 249, + 583, + 361, + 599 + ], + "lines": [ + { + "bbox": [ + 249, + 583, + 361, + 599 + ], + "spans": [ + { + "bbox": [ + 249, + 583, + 361, + 599 + ], + "score": 0.92, + "content": "\\mathbf { q } _ { i } = \\mathcal { F } _ { q } ( \\mathcal { E } ( \\mathbf { y } _ { i } ) ) , \\ \\mathbf { q } _ { i } \\in \\mathbb { R } ^ { D }", + "type": "interline_equation", + "image_path": "c40dedb2df7d946a4b77f9cbd45eddc8f8cd4551e0e35190f0aa46660f3d884e.jpg" + } + ] + } + ], + "index": 53, + "virtual_lines": [ + { + "bbox": [ + 249, + 583, + 361, + 599 + ], + "spans": [], + "index": 53 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 603, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 105, + 602, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 132, + 617 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 604, + 150, + 615 + ], + "score": 0.9, + "content": "\\mathcal { E } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 602, + 380, + 617 + ], + "score": 1.0, + "content": "is a instance encoding function (detailed in in Sec. 3) and", + "type": "text" + }, + { + "bbox": [ + 380, + 604, + 393, + 616 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 602, + 505, + 617 + ], + "score": 1.0, + "content": "is a Multi-Layer Perception", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 106, + 615, + 173, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 173, + 627 + ], + "score": 1.0, + "content": "network (MLP).", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 54.5, + "bbox_fs": [ + 105, + 602, + 505, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 630, + 505, + 666 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 219, + 644 + ], + "score": 1.0, + "content": "We retrieve knowledge from", + "type": "text" + }, + { + "bbox": [ + 219, + 631, + 229, + 641 + ], + "score": 0.83, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 630, + 253, + 644 + ], + "score": 1.0, + "content": "given", + "type": "text" + }, + { + "bbox": [ + 253, + 632, + 264, + 642 + ], + "score": 0.86, + "content": "\\mathbf { q } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 630, + 505, + 644 + ], + "score": 1.0, + "content": "by measuring responses of correlation. This is formulated by", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 642, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 232, + 654 + ], + "score": 1.0, + "content": "dot-product attention [43] with", + "type": "text" + }, + { + "bbox": [ + 232, + 642, + 244, + 652 + ], + "score": 0.83, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 642, + 505, + 654 + ], + "score": 1.0, + "content": "concurrent heads in a query-key attention manner. In which each", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 652, + 504, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 127, + 667 + ], + "score": 1.0, + "content": "head", + "type": "text" + }, + { + "bbox": [ + 128, + 654, + 134, + 664 + ], + "score": 0.82, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 652, + 270, + 667 + ], + "score": 1.0, + "content": "corresponds to three linear layers", + "type": "text" + }, + { + "bbox": [ + 270, + 652, + 346, + 667 + ], + "score": 0.93, + "content": "( \\mathcal { F } _ { j } ^ { k } , \\mathcal { F } _ { j } ^ { q } , \\mathcal { F } _ { j } ^ { v } ) \\mathop { w . r . t . }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 652, + 504, + 667 + ], + "score": 1.0, + "content": ". the key, query and value construction.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 57, + "bbox_fs": [ + 105, + 630, + 505, + 667 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 671, + 505, + 724 + ], + "lines": [ + { + "bbox": [ + 105, + 670, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 174, + 684 + ], + "score": 1.0, + "content": "The key feature", + "type": "text" + }, + { + "bbox": [ + 174, + 671, + 190, + 685 + ], + "score": 0.91, + "content": "\\mathrm { K } _ { j } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 670, + 407, + 684 + ], + "score": 1.0, + "content": "is computed by projecting teacher’s representations", + "type": "text" + }, + { + "bbox": [ + 408, + 671, + 424, + 682 + ], + "score": 0.87, + "content": "\\mathrm { A } ^ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 670, + 505, + 684 + ], + "score": 1.0, + "content": "with the positional", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 682, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 189, + 699 + ], + "score": 1.0, + "content": "embeddings [43, 5]", + "type": "text" + }, + { + "bbox": [ + 189, + 684, + 236, + 696 + ], + "score": 0.9, + "content": "\\mathrm { P } \\in \\mathbb { R } ^ { L \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 682, + 317, + 699 + ], + "score": 1.0, + "content": "as in Eq. 3, where", + "type": "text" + }, + { + "bbox": [ + 317, + 685, + 334, + 698 + ], + "score": 0.91, + "content": "\\mathcal { F } _ { p e }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 682, + 506, + 699 + ], + "score": 1.0, + "content": "denotes a linear projection over position", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 695, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 236, + 711 + ], + "score": 1.0, + "content": "embeddings. The value feature", + "type": "text" + }, + { + "bbox": [ + 236, + 696, + 252, + 711 + ], + "score": 0.9, + "content": "\\mathrm { V } _ { j } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 695, + 297, + 711 + ], + "score": 1.0, + "content": "and query", + "type": "text" + }, + { + "bbox": [ + 297, + 699, + 311, + 710 + ], + "score": 0.88, + "content": "{ \\bf q } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 695, + 505, + 711 + ], + "score": 1.0, + "content": "is projected by linear mappings to a sub-space", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 708, + 505, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 127, + 725 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 710, + 172, + 723 + ], + "score": 0.93, + "content": "d = D / M", + "type": "inline_equation" + }, + { + "bbox": [ + 172, + 708, + 214, + 725 + ], + "score": 1.0, + "content": "channels,", + "type": "text" + }, + { + "bbox": [ + 215, + 711, + 228, + 724 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { j } ^ { v }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 708, + 244, + 725 + ], + "score": 1.0, + "content": "on", + "type": "text" + }, + { + "bbox": [ + 244, + 710, + 259, + 721 + ], + "score": 0.84, + "content": "\\mathrm { A } ^ { \\tau }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 708, + 278, + 725 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 279, + 711, + 292, + 725 + ], + "score": 0.91, + "content": "\\mathcal { F } _ { j } ^ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 708, + 314, + 725 + ], + "score": 1.0, + "content": "over", + "type": "text" + }, + { + "bbox": [ + 314, + 712, + 325, + 722 + ], + "score": 0.85, + "content": "\\mathbf { q } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 708, + 505, + 725 + ], + "score": 1.0, + "content": "respectively, as shown in Eq. 4. At last, an", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 106, + 249, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 231, + 263 + ], + "score": 1.0, + "content": "instance-aware attention mask", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 232, + 251, + 249, + 263 + ], + "score": 0.89, + "content": "\\mathbf { m } _ { i j }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 250, + 249, + 276, + 263 + ], + "score": 1.0, + "content": "of the", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 277, + 251, + 281, + 260 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 281, + 249, + 357, + 263 + ], + "score": 1.0, + "content": "-th instance by the", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 357, + 251, + 363, + 262 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 363, + 249, + 505, + 263 + ], + "score": 1.0, + "content": "-th head is obtained by normalized", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 259, + 244, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 191, + 277 + ], + "score": 1.0, + "content": "dot-product between", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 191, + 261, + 207, + 276 + ], + "score": 0.9, + "content": "\\mathrm { K } _ { j } ^ { \\mathcal { T } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 208, + 259, + 225, + 277 + ], + "score": 1.0, + "content": "and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 225, + 263, + 239, + 275 + ], + "score": 0.88, + "content": "\\mathbf { q } _ { i j }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 239, + 259, + 244, + 277 + ], + "score": 1.0, + "content": ":", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 60.5, + "bbox_fs": [ + 105, + 670, + 506, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 69, + 495, + 187 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 69, + 495, + 187 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 69, + 495, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 69, + 495, + 187 + ], + "score": 0.974, + "type": "image", + "image_path": "26f2f0d1133a571e224a34b9c1b4806a50feadae4f96310903a3434d50f769d5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 69, + 495, + 108.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 108.33333333333334, + 495, + 147.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 147.66666666666669, + 495, + 187.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 193, + 505, + 237 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 193, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 506, + 206 + ], + "score": 1.0, + "content": "Figure 3: The illustration of the auxiliary task. (a) The instance encoding function encodes a instance", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "score": 1.0, + "content": "condition to a vector, it is then projected as query features. (b) The identification task learns to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "identify the existence the queried instance. (c) The localization task learns to predict the boundary", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 225, + 309, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 309, + 240 + ], + "score": 1.0, + "content": "given an uncertain position provided by the query.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 106, + 249, + 503, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 231, + 263 + ], + "score": 1.0, + "content": "instance-aware attention mask", + "type": "text" + }, + { + "bbox": [ + 232, + 251, + 249, + 263 + ], + "score": 0.89, + "content": "\\mathbf { m } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 249, + 276, + 263 + ], + "score": 1.0, + "content": "of the", + "type": "text" + }, + { + "bbox": [ + 277, + 251, + 281, + 260 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 249, + 357, + 263 + ], + "score": 1.0, + "content": "-th instance by the", + "type": "text" + }, + { + "bbox": [ + 357, + 251, + 363, + 262 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 249, + 505, + 263 + ], + "score": 1.0, + "content": "-th head is obtained by normalized", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 259, + 244, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 191, + 277 + ], + "score": 1.0, + "content": "dot-product between", + "type": "text" + }, + { + "bbox": [ + 191, + 261, + 207, + 276 + ], + "score": 0.9, + "content": "\\mathrm { K } _ { j } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 259, + 225, + 277 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 225, + 263, + 239, + 275 + ], + "score": 0.88, + "content": "\\mathbf { q } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 259, + 244, + 277 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 279, + 390, + 361 + ], + "lines": [ + { + "bbox": [ + 220, + 279, + 390, + 361 + ], + "spans": [ + { + "bbox": [ + 220, + 279, + 390, + 361 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\mathrm { K } _ { j } ^ { \\mathcal { T } } = \\mathcal { F } _ { j } ^ { k } ( \\mathrm { A } ^ { \\mathcal { T } } + \\mathcal { F } _ { p e } ( \\mathrm { P } ) ) , \\mathrm { K } _ { j } ^ { \\mathcal { T } } \\in \\mathbb { R } ^ { L \\times d } } \\\\ & { \\mathrm { V } _ { j } ^ { \\mathcal { T } } = \\mathcal { F } _ { j } ^ { v } ( \\mathrm { A } ^ { \\mathcal { T } } ) , \\mathrm { V } _ { j } ^ { \\mathcal { T } } \\in \\mathbb { R } ^ { L \\times d } } \\\\ & { \\mathbf { q } _ { i j } = \\mathcal { F } _ { j } ^ { q } ( \\mathbf { q } _ { i } ) , \\mathbf { q } _ { i j } \\in \\mathbb { R } ^ { d } } \\\\ & { \\mathbf { m } _ { i j } = s o f t m a x ( \\frac { \\mathrm { K } _ { j } ^ { \\mathcal { T } } \\mathbf { q } _ { i j } } { \\sqrt { d } } ) , \\mathbf { m } _ { i j } \\in \\mathbb { R } ^ { L } } \\end{array}", + "type": "interline_equation", + "image_path": "c688ddfb981ed8dc0544368a7f436331606b3ec56010e4d9e634d553fdee1423.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 220, + 279, + 390, + 295.4 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 220, + 295.4, + 390, + 311.79999999999995 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 220, + 311.79999999999995, + 390, + 328.19999999999993 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 220, + 328.19999999999993, + 390, + 344.5999999999999 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 220, + 344.5999999999999, + 390, + 360.9999999999999 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 506, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 381 + ], + "score": 1.0, + "content": "Intuitively, the querying along the key features and value features describes the correlation be-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 379, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 300, + 393 + ], + "score": 1.0, + "content": "tween representations and instances. We collect", + "type": "text" + }, + { + "bbox": [ + 300, + 379, + 394, + 393 + ], + "score": 0.93, + "content": "\\kappa _ { i } ^ { \\mathcal { T } } = \\{ ( \\mathbf { m } _ { i j } , \\mathrm { V } _ { j } ^ { \\mathcal { T } } ) \\} _ { j = 1 } ^ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 380, + 505, + 393 + ], + "score": 1.0, + "content": "as the instance-conditional", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 391, + 422, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 174, + 403 + ], + "score": 1.0, + "content": "knowledge from", + "type": "text" + }, + { + "bbox": [ + 174, + 392, + 183, + 401 + ], + "score": 0.82, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 391, + 422, + 403 + ], + "score": 1.0, + "content": ", which encodes knowledge corresponds to the ith instance.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 107, + 415, + 193, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 195, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 195, + 429 + ], + "score": 1.0, + "content": "3.3 Auxiliary Task", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 416, + 448 + ], + "score": 1.0, + "content": "In this section, we introduce the auxiliary task to optimize the decoding module", + "type": "text" + }, + { + "bbox": [ + 417, + 436, + 424, + 446 + ], + "score": 0.8, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 434, + 505, + 448 + ], + "score": 1.0, + "content": ". First, we aggregate", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 444, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 460 + ], + "score": 1.0, + "content": "instance-level information to identify and localize objects. This could be obtained by aggregating", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 314, + 471 + ], + "score": 1.0, + "content": "the instance-conditional knowledge by the function", + "type": "text" + }, + { + "bbox": [ + 315, + 457, + 335, + 470 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { a g g }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 456, + 506, + 471 + ], + "score": 1.0, + "content": ", which includes sum-product aggregation", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 468, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 468, + 164, + 482 + ], + "score": 1.0, + "content": "over attention", + "type": "text" + }, + { + "bbox": [ + 165, + 470, + 182, + 482 + ], + "score": 0.88, + "content": "\\mathbf { m } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 468, + 201, + 482 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 201, + 468, + 217, + 483 + ], + "score": 0.91, + "content": "\\mathrm { V } _ { j } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 468, + 506, + 482 + ], + "score": 1.0, + "content": ", concatenate features from each head, add residuals and project with a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 480, + 288, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 288, + 493 + ], + "score": 1.0, + "content": "feed-forward network as proposed in [43, 5]:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 496, + 371, + 512 + ], + "lines": [ + { + "bbox": [ + 239, + 496, + 371, + 512 + ], + "spans": [ + { + "bbox": [ + 239, + 496, + 371, + 512 + ], + "score": 0.92, + "content": "\\mathbf { g } _ { i } ^ { T } = \\mathcal { F } _ { a g g } ( \\kappa _ { i } ^ { T } , \\mathbf { q } _ { i } ) , \\ \\mathbf { g } _ { i } ^ { T } \\in \\mathbb { R } ^ { D }", + "type": "interline_equation", + "image_path": "335388df20d486987a5e9258482b0ffa53ddc06681685467d1f09359cf684128.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 239, + 496, + 371, + 512 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 522, + 504, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 298, + 536 + ], + "score": 1.0, + "content": "To let the instance-level aggregated information", + "type": "text" + }, + { + "bbox": [ + 298, + 522, + 312, + 535 + ], + "score": 0.91, + "content": "\\mathbf { g } _ { i } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 520, + 505, + 536 + ], + "score": 1.0, + "content": "retain sufficient instance cues, one could design", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 534, + 306, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 306, + 546 + ], + "score": 1.0, + "content": "an instance-sensitive task to optimize it as below:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "interline_equation", + "bbox": [ + 253, + 549, + 358, + 565 + ], + "lines": [ + { + "bbox": [ + 253, + 549, + 358, + 565 + ], + "spans": [ + { + "bbox": [ + 253, + 549, + 358, + 565 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { a u x } = \\mathcal { L } _ { i n s } ( \\mathbf { g } _ { i } ^ { T } , \\mathcal { H } ( \\mathbf { y } _ { i } ) )", + "type": "interline_equation", + "image_path": "5613b431532b5cffb452d4c282236811b525f6f89e613840b04655d54114cb47.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 253, + 549, + 358, + 565 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 569, + 505, + 624 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 133, + 581 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 570, + 144, + 579 + ], + "score": 0.78, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "encodes the instance information as targets. However, directly adopt Eq. 8 might lead to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 192, + 593 + ], + "score": 1.0, + "content": "trivial solution, since", + "type": "text" + }, + { + "bbox": [ + 192, + 582, + 202, + 592 + ], + "score": 0.84, + "content": "{ \\tt y } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 580, + 297, + 593 + ], + "score": 1.0, + "content": "is accessible from both", + "type": "text" + }, + { + "bbox": [ + 297, + 579, + 311, + 592 + ], + "score": 0.91, + "content": "\\mathbf { g } _ { i } ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 580, + 348, + 593 + ], + "score": 1.0, + "content": "(through", + "type": "text" + }, + { + "bbox": [ + 349, + 582, + 359, + 592 + ], + "score": 0.86, + "content": "\\mathbf { q } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 580, + 422, + 593 + ], + "score": 1.0, + "content": ", see Eq. 2) and", + "type": "text" + }, + { + "bbox": [ + 422, + 581, + 448, + 592 + ], + "score": 0.91, + "content": "\\mathcal { H } ( \\mathtt { y } _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 580, + 505, + 593 + ], + "score": 1.0, + "content": ". It is possible", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 590, + 507, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 407, + 604 + ], + "score": 1.0, + "content": "that parameters will learn a shortcut, that ignore the teacher representations", + "type": "text" + }, + { + "bbox": [ + 408, + 592, + 416, + 601 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 590, + 507, + 604 + ], + "score": 1.0, + "content": ". To resolve this issue,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 599, + 507, + 618 + ], + "spans": [ + { + "bbox": [ + 104, + 599, + 319, + 618 + ], + "score": 1.0, + "content": "we propose to drop information on encoding function", + "type": "text" + }, + { + "bbox": [ + 320, + 602, + 337, + 614 + ], + "score": 0.91, + "content": "\\mathcal { E } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 599, + 472, + 618 + ], + "score": 1.0, + "content": ", to force the aggregation function", + "type": "text" + }, + { + "bbox": [ + 473, + 603, + 493, + 615 + ], + "score": 0.91, + "content": "\\mathcal { F } _ { a g g }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 599, + 507, + 618 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 613, + 199, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 187, + 625 + ], + "score": 1.0, + "content": "excavate hints from", + "type": "text" + }, + { + "bbox": [ + 187, + 614, + 196, + 623 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 613, + 199, + 625 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 629, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "The information dropping is adopted by replacing the accurate annotation for instance conditions", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 640, + 504, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 504, + 653 + ], + "score": 1.0, + "content": "to uncertain ones. For bounding box annotations, we relieve them to rough box centers with rough", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 651, + 491, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 266, + 665 + ], + "score": 1.0, + "content": "scales indicators. The rough box center", + "type": "text" + }, + { + "bbox": [ + 266, + 651, + 296, + 663 + ], + "score": 0.93, + "content": "( x _ { i } ^ { \\prime } , y _ { i } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 651, + 491, + 665 + ], + "score": 1.0, + "content": "is obtained by random jittering as shown below:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "interline_equation", + "bbox": [ + 256, + 668, + 338, + 695 + ], + "lines": [ + { + "bbox": [ + 256, + 668, + 338, + 695 + ], + "spans": [ + { + "bbox": [ + 256, + 668, + 338, + 695 + ], + "score": 0.93, + "content": "\\left\\{ \\begin{array} { l l } { x _ { i } ^ { \\prime } = x _ { i } + \\phi _ { x } w _ { i } , } \\\\ { y _ { i } ^ { \\prime } = y _ { i } + \\phi _ { y } h _ { i } , } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "c3cfb8c9517b4dbe16cda6c2672bf94f2c864102100877e82ffb70dd7fabb35b.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 256, + 668, + 338, + 695 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 133, + 712 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 700, + 165, + 712 + ], + "score": 0.92, + "content": "( w _ { i } , h _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 699, + 362, + 712 + ], + "score": 1.0, + "content": "is the width and height of the bounding box and", + "type": "text" + }, + { + "bbox": [ + 362, + 700, + 389, + 712 + ], + "score": 0.94, + "content": "\\phi _ { x } , \\phi _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "are sampled from a uniform", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 156, + 723 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 156, + 711, + 214, + 723 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\dot { \\Phi } \\sim U [ - a , a ] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 710, + 274, + 723 + ], + "score": 1.0, + "content": ", where we set", + "type": "text" + }, + { + "bbox": [ + 274, + 711, + 299, + 721 + ], + "score": 0.77, + "content": "\\mathrm { a } { = } 0 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "empirically. The scales indicators are obtained by", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 740, + 309, + 753 + ], + "spans": [ + { + "bbox": [ + 302, + 740, + 309, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 69, + 495, + 187 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 69, + 495, + 187 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 69, + 495, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 69, + 495, + 187 + ], + "score": 0.974, + "type": "image", + "image_path": "26f2f0d1133a571e224a34b9c1b4806a50feadae4f96310903a3434d50f769d5.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 69, + 495, + 108.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 108.33333333333334, + 495, + 147.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 147.66666666666669, + 495, + 187.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 193, + 505, + 237 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 193, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 506, + 206 + ], + "score": 1.0, + "content": "Figure 3: The illustration of the auxiliary task. (a) The instance encoding function encodes a instance", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 506, + 217 + ], + "score": 1.0, + "content": "condition to a vector, it is then projected as query features. (b) The identification task learns to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 228 + ], + "score": 1.0, + "content": "identify the existence the queried instance. (c) The localization task learns to predict the boundary", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 225, + 309, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 309, + 240 + ], + "score": 1.0, + "content": "given an uncertain position provided by the query.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 106, + 249, + 503, + 275 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 105, + 249, + 505, + 277 + ], + "lines_deleted": true + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 279, + 390, + 361 + ], + "lines": [ + { + "bbox": [ + 220, + 279, + 390, + 361 + ], + "spans": [ + { + "bbox": [ + 220, + 279, + 390, + 361 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\mathrm { K } _ { j } ^ { \\mathcal { T } } = \\mathcal { F } _ { j } ^ { k } ( \\mathrm { A } ^ { \\mathcal { T } } + \\mathcal { F } _ { p e } ( \\mathrm { P } ) ) , \\mathrm { K } _ { j } ^ { \\mathcal { T } } \\in \\mathbb { R } ^ { L \\times d } } \\\\ & { \\mathrm { V } _ { j } ^ { \\mathcal { T } } = \\mathcal { F } _ { j } ^ { v } ( \\mathrm { A } ^ { \\mathcal { T } } ) , \\mathrm { V } _ { j } ^ { \\mathcal { T } } \\in \\mathbb { R } ^ { L \\times d } } \\\\ & { \\mathbf { q } _ { i j } = \\mathcal { F } _ { j } ^ { q } ( \\mathbf { q } _ { i } ) , \\mathbf { q } _ { i j } \\in \\mathbb { R } ^ { d } } \\\\ & { \\mathbf { m } _ { i j } = s o f t m a x ( \\frac { \\mathrm { K } _ { j } ^ { \\mathcal { T } } \\mathbf { q } _ { i j } } { \\sqrt { d } } ) , \\mathbf { m } _ { i j } \\in \\mathbb { R } ^ { L } } \\end{array}", + "type": "interline_equation", + "image_path": "c688ddfb981ed8dc0544368a7f436331606b3ec56010e4d9e634d553fdee1423.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 220, + 279, + 390, + 295.4 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 220, + 295.4, + 390, + 311.79999999999995 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 220, + 311.79999999999995, + 390, + 328.19999999999993 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 220, + 328.19999999999993, + 390, + 344.5999999999999 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 220, + 344.5999999999999, + 390, + 360.9999999999999 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 506, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 381 + ], + "score": 1.0, + "content": "Intuitively, the querying along the key features and value features describes the correlation be-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 379, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 300, + 393 + ], + "score": 1.0, + "content": "tween representations and instances. We collect", + "type": "text" + }, + { + "bbox": [ + 300, + 379, + 394, + 393 + ], + "score": 0.93, + "content": "\\kappa _ { i } ^ { \\mathcal { T } } = \\{ ( \\mathbf { m } _ { i j } , \\mathrm { V } _ { j } ^ { \\mathcal { T } } ) \\} _ { j = 1 } ^ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 380, + 505, + 393 + ], + "score": 1.0, + "content": "as the instance-conditional", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 391, + 422, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 174, + 403 + ], + "score": 1.0, + "content": "knowledge from", + "type": "text" + }, + { + "bbox": [ + 174, + 392, + 183, + 401 + ], + "score": 0.82, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 391, + 422, + 403 + ], + "score": 1.0, + "content": ", which encodes knowledge corresponds to the ith instance.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 369, + 506, + 403 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 415, + 193, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 195, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 195, + 429 + ], + "score": 1.0, + "content": "3.3 Auxiliary Task", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 434, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 416, + 448 + ], + "score": 1.0, + "content": "In this section, we introduce the auxiliary task to optimize the decoding module", + "type": "text" + }, + { + "bbox": [ + 417, + 436, + 424, + 446 + ], + "score": 0.8, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 434, + 505, + 448 + ], + "score": 1.0, + "content": ". First, we aggregate", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 444, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 460 + ], + "score": 1.0, + "content": "instance-level information to identify and localize objects. This could be obtained by aggregating", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 314, + 471 + ], + "score": 1.0, + "content": "the instance-conditional knowledge by the function", + "type": "text" + }, + { + "bbox": [ + 315, + 457, + 335, + 470 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { a g g }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 456, + 506, + 471 + ], + "score": 1.0, + "content": ", which includes sum-product aggregation", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 468, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 468, + 164, + 482 + ], + "score": 1.0, + "content": "over attention", + "type": "text" + }, + { + "bbox": [ + 165, + 470, + 182, + 482 + ], + "score": 0.88, + "content": "\\mathbf { m } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 468, + 201, + 482 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 201, + 468, + 217, + 483 + ], + "score": 0.91, + "content": "\\mathrm { V } _ { j } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 468, + 506, + 482 + ], + "score": 1.0, + "content": ", concatenate features from each head, add residuals and project with a", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 480, + 288, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 288, + 493 + ], + "score": 1.0, + "content": "feed-forward network as proposed in [43, 5]:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 104, + 434, + 506, + 493 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 496, + 371, + 512 + ], + "lines": [ + { + "bbox": [ + 239, + 496, + 371, + 512 + ], + "spans": [ + { + "bbox": [ + 239, + 496, + 371, + 512 + ], + "score": 0.92, + "content": "\\mathbf { g } _ { i } ^ { T } = \\mathcal { F } _ { a g g } ( \\kappa _ { i } ^ { T } , \\mathbf { q } _ { i } ) , \\ \\mathbf { g } _ { i } ^ { T } \\in \\mathbb { R } ^ { D }", + "type": "interline_equation", + "image_path": "335388df20d486987a5e9258482b0ffa53ddc06681685467d1f09359cf684128.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 239, + 496, + 371, + 512 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 522, + 504, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 520, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 298, + 536 + ], + "score": 1.0, + "content": "To let the instance-level aggregated information", + "type": "text" + }, + { + "bbox": [ + 298, + 522, + 312, + 535 + ], + "score": 0.91, + "content": "\\mathbf { g } _ { i } ^ { \\mathcal { T } }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 520, + 505, + 536 + ], + "score": 1.0, + "content": "retain sufficient instance cues, one could design", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 534, + 306, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 306, + 546 + ], + "score": 1.0, + "content": "an instance-sensitive task to optimize it as below:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 520, + 505, + 546 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 253, + 549, + 358, + 565 + ], + "lines": [ + { + "bbox": [ + 253, + 549, + 358, + 565 + ], + "spans": [ + { + "bbox": [ + 253, + 549, + 358, + 565 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { a u x } = \\mathcal { L } _ { i n s } ( \\mathbf { g } _ { i } ^ { T } , \\mathcal { H } ( \\mathbf { y } _ { i } ) )", + "type": "interline_equation", + "image_path": "5613b431532b5cffb452d4c282236811b525f6f89e613840b04655d54114cb47.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 253, + 549, + 358, + 565 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 569, + 505, + 624 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 133, + 581 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 570, + 144, + 579 + ], + "score": 0.78, + "content": "\\mathcal { H }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "encodes the instance information as targets. However, directly adopt Eq. 8 might lead to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 579, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 192, + 593 + ], + "score": 1.0, + "content": "trivial solution, since", + "type": "text" + }, + { + "bbox": [ + 192, + 582, + 202, + 592 + ], + "score": 0.84, + "content": "{ \\tt y } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 580, + 297, + 593 + ], + "score": 1.0, + "content": "is accessible from both", + "type": "text" + }, + { + "bbox": [ + 297, + 579, + 311, + 592 + ], + "score": 0.91, + "content": "\\mathbf { g } _ { i } ^ { \\top }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 580, + 348, + 593 + ], + "score": 1.0, + "content": "(through", + "type": "text" + }, + { + "bbox": [ + 349, + 582, + 359, + 592 + ], + "score": 0.86, + "content": "\\mathbf { q } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 580, + 422, + 593 + ], + "score": 1.0, + "content": ", see Eq. 2) and", + "type": "text" + }, + { + "bbox": [ + 422, + 581, + 448, + 592 + ], + "score": 0.91, + "content": "\\mathcal { H } ( \\mathtt { y } _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 580, + 505, + 593 + ], + "score": 1.0, + "content": ". It is possible", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 590, + 507, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 407, + 604 + ], + "score": 1.0, + "content": "that parameters will learn a shortcut, that ignore the teacher representations", + "type": "text" + }, + { + "bbox": [ + 408, + 592, + 416, + 601 + ], + "score": 0.8, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 590, + 507, + 604 + ], + "score": 1.0, + "content": ". To resolve this issue,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 599, + 507, + 618 + ], + "spans": [ + { + "bbox": [ + 104, + 599, + 319, + 618 + ], + "score": 1.0, + "content": "we propose to drop information on encoding function", + "type": "text" + }, + { + "bbox": [ + 320, + 602, + 337, + 614 + ], + "score": 0.91, + "content": "\\mathcal { E } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 599, + 472, + 618 + ], + "score": 1.0, + "content": ", to force the aggregation function", + "type": "text" + }, + { + "bbox": [ + 473, + 603, + 493, + 615 + ], + "score": 0.91, + "content": "\\mathcal { F } _ { a g g }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 599, + 507, + 618 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 613, + 199, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 187, + 625 + ], + "score": 1.0, + "content": "excavate hints from", + "type": "text" + }, + { + "bbox": [ + 187, + 614, + 196, + 623 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 613, + 199, + 625 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 104, + 569, + 507, + 625 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 629, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 642 + ], + "score": 1.0, + "content": "The information dropping is adopted by replacing the accurate annotation for instance conditions", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 640, + 504, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 504, + 653 + ], + "score": 1.0, + "content": "to uncertain ones. For bounding box annotations, we relieve them to rough box centers with rough", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 651, + 491, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 266, + 665 + ], + "score": 1.0, + "content": "scales indicators. The rough box center", + "type": "text" + }, + { + "bbox": [ + 266, + 651, + 296, + 663 + ], + "score": 0.93, + "content": "( x _ { i } ^ { \\prime } , y _ { i } ^ { \\prime } )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 651, + 491, + 665 + ], + "score": 1.0, + "content": "is obtained by random jittering as shown below:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 628, + 505, + 665 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 256, + 668, + 338, + 695 + ], + "lines": [ + { + "bbox": [ + 256, + 668, + 338, + 695 + ], + "spans": [ + { + "bbox": [ + 256, + 668, + 338, + 695 + ], + "score": 0.93, + "content": "\\left\\{ \\begin{array} { l l } { x _ { i } ^ { \\prime } = x _ { i } + \\phi _ { x } w _ { i } , } \\\\ { y _ { i } ^ { \\prime } = y _ { i } + \\phi _ { y } h _ { i } , } \\end{array} \\right.", + "type": "interline_equation", + "image_path": "c3cfb8c9517b4dbe16cda6c2672bf94f2c864102100877e82ffb70dd7fabb35b.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 256, + 668, + 338, + 695 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 699, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 133, + 712 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 700, + 165, + 712 + ], + "score": 0.92, + "content": "( w _ { i } , h _ { i } )", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 699, + 362, + 712 + ], + "score": 1.0, + "content": "is the width and height of the bounding box and", + "type": "text" + }, + { + "bbox": [ + 362, + 700, + 389, + 712 + ], + "score": 0.94, + "content": "\\phi _ { x } , \\phi _ { y }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "are sampled from a uniform", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 156, + 723 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 156, + 711, + 214, + 723 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\dot { \\Phi } \\sim U [ - a , a ] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 710, + 274, + 723 + ], + "score": 1.0, + "content": ", where we set", + "type": "text" + }, + { + "bbox": [ + 274, + 711, + 299, + 721 + ], + "score": 0.77, + "content": "\\mathrm { a } { = } 0 . 3", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "empirically. The scales indicators are obtained by", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 321, + 86 + ], + "score": 1.0, + "content": "rounding the box sizes in the logarithmic space, i.e.,", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 321, + 73, + 365, + 85 + ], + "score": 0.86, + "content": "\\textcircled { l o g _ { 2 } ( w _ { i } ) } ", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 365, + 72, + 370, + 86 + ], + "score": 1.0, + "content": ",", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 370, + 73, + 413, + 85 + ], + "score": 0.85, + "content": "\\lfloor l o g _ { 2 } ( h _ { i } ) \\rfloor", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 413, + 72, + 505, + 86 + ], + "score": 1.0, + "content": ". In addition, to let the", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "decoder learn to identify instances and be aware of the uncertainty, we generate fake instances for the", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "identification task according to dataset distributions, this is detailed in Appendix A. As a result, it", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 504, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 360, + 119 + ], + "score": 1.0, + "content": "collect coarse information and obtain instance encoding through", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 360, + 105, + 378, + 118 + ], + "score": 0.91, + "content": "\\mathcal { E } ( \\cdot )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 379, + 104, + 495, + 119 + ], + "score": 1.0, + "content": "as depicted in Fig. 3a. where", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 495, + 107, + 504, + 117 + ], + "score": 0.8, + "content": "c _ { i }", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 115, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 168, + 129 + ], + "score": 1.0, + "content": "is the category,", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 169, + 117, + 186, + 128 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { o h }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 186, + 115, + 463, + 129 + ], + "score": 1.0, + "content": "is the one hot vectorization, concat is the concatenation operator and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 464, + 117, + 480, + 129 + ], + "score": 0.91, + "content": "\\mathcal { F } _ { p e }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 480, + 115, + 505, + 129 + ], + "score": 1.0, + "content": "is the", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 226, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 226, + 139 + ], + "score": 1.0, + "content": "position embedding function.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 699, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 139 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 321, + 86 + ], + "score": 1.0, + "content": "rounding the box sizes in the logarithmic space, i.e.,", + "type": "text" + }, + { + "bbox": [ + 321, + 73, + 365, + 85 + ], + "score": 0.86, + "content": "\\textcircled { l o g _ { 2 } ( w _ { i } ) } ", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 72, + 370, + 86 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 370, + 73, + 413, + 85 + ], + "score": 0.85, + "content": "\\lfloor l o g _ { 2 } ( h _ { i } ) \\rfloor", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 72, + 505, + 86 + ], + "score": 1.0, + "content": ". In addition, to let the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "decoder learn to identify instances and be aware of the uncertainty, we generate fake instances for the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "identification task according to dataset distributions, this is detailed in Appendix A. As a result, it", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 504, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 360, + 119 + ], + "score": 1.0, + "content": "collect coarse information and obtain instance encoding through", + "type": "text" + }, + { + "bbox": [ + 360, + 105, + 378, + 118 + ], + "score": 0.91, + "content": "\\mathcal { E } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 104, + 495, + 119 + ], + "score": 1.0, + "content": "as depicted in Fig. 3a. where", + "type": "text" + }, + { + "bbox": [ + 495, + 107, + 504, + 117 + ], + "score": 0.8, + "content": "c _ { i }", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 115, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 168, + 129 + ], + "score": 1.0, + "content": "is the category,", + "type": "text" + }, + { + "bbox": [ + 169, + 117, + 186, + 128 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { o h }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 115, + 463, + 129 + ], + "score": 1.0, + "content": "is the one hot vectorization, concat is the concatenation operator and", + "type": "text" + }, + { + "bbox": [ + 464, + 117, + 480, + 129 + ], + "score": 0.91, + "content": "\\mathcal { F } _ { p e }", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 115, + 505, + 129 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 226, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 226, + 139 + ], + "score": 1.0, + "content": "position embedding function.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 142, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 262, + 157 + ], + "score": 1.0, + "content": "Finally, the aggregated representation", + "type": "text" + }, + { + "bbox": [ + 262, + 143, + 276, + 156 + ], + "score": 0.91, + "content": "\\mathbf { g } _ { i } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 143, + 506, + 157 + ], + "score": 1.0, + "content": "are optimized by the auxiliary task. We introduces two", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 193, + 167 + ], + "score": 1.0, + "content": "predictors denoted by", + "type": "text" + }, + { + "bbox": [ + 193, + 155, + 213, + 167 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { o b j }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 154, + 231, + 167 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 231, + 155, + 251, + 166 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { r e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "respectively to predict identification and localization results. We", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 426, + 178 + ], + "score": 1.0, + "content": "adopt binary cross entropy loss (BCE) to optimize the real-fake identification and", + "type": "text" + }, + { + "bbox": [ + 427, + 166, + 439, + 175 + ], + "score": 0.8, + "content": "L 1", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "loss to optimize", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 176, + 167, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 167, + 190 + ], + "score": 1.0, + "content": "the regression.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 201, + 433, + 218 + ], + "lines": [ + { + "bbox": [ + 177, + 201, + 433, + 218 + ], + "spans": [ + { + "bbox": [ + 177, + 201, + 433, + 218 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { a u x } = \\mathcal { L } _ { B C E } ( \\mathcal { F } _ { o b j } ( \\mathbf { g } _ { i } ^ { T } ) , \\delta _ { o b j } ( \\mathbf { y } _ { i } ) ) + \\mathcal { L } _ { 1 } ( \\mathcal { F } _ { r e g } ( \\mathbf { g } _ { i } ^ { T } ) , \\delta _ { r e g } ( \\mathbf { y } _ { i } ) )", + "type": "interline_equation", + "image_path": "af7bb8afe0a58cd75764a645ba03e94b48dc610de26922d5f3d1ffb40796af89.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 177, + 201, + 433, + 218 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 225, + 506, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 132, + 238 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 225, + 160, + 238 + ], + "score": 0.92, + "content": "\\delta _ { o b j } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 225, + 271, + 238 + ], + "score": 1.0, + "content": "is an indicator, it yields 1 if", + "type": "text" + }, + { + "bbox": [ + 271, + 227, + 281, + 237 + ], + "score": 0.86, + "content": "\\mathtt { y } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "is real and 0 otherwise. Following common practice, the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 235, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 295, + 250 + ], + "score": 1.0, + "content": "localization loss for fake examples is ignored.", + "type": "text" + }, + { + "bbox": [ + 295, + 237, + 313, + 249 + ], + "score": 0.91, + "content": "\\delta _ { r e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 235, + 506, + 250 + ], + "score": 1.0, + "content": "is the preparing function for regression targets,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 248, + 377, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 377, + 260 + ], + "score": 1.0, + "content": "following [42]. Appendix A provides more implementation details.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "title", + "bbox": [ + 108, + 272, + 270, + 284 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 271, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 271, + 286 + ], + "score": 1.0, + "content": "3.4 Instance-Conditional Distillation", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 292, + 506, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 293, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 505, + 304 + ], + "score": 1.0, + "content": "Lastly, we present the formulation for conditional knowledge distillation. We obtain the projected", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 301, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 104, + 301, + 163, + 317 + ], + "score": 1.0, + "content": "value features", + "type": "text" + }, + { + "bbox": [ + 163, + 303, + 213, + 317 + ], + "score": 0.92, + "content": "\\mathrm { V } _ { j } ^ { S } \\in \\mathbb { R } ^ { L \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 301, + 506, + 317 + ], + "score": 1.0, + "content": "of the student representations analogous to Eq. 4 in Sec. 3.2. By adopting", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 313, + 507, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 507, + 327 + ], + "score": 1.0, + "content": "the instance-aware attention mask as a measurement of correlations between feature and each instance,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 325, + 450, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 450, + 338 + ], + "score": 1.0, + "content": "we formulate the distillation loss as value features mimicking guided by the attention:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 342, + 431, + 378 + ], + "lines": [ + { + "bbox": [ + 179, + 342, + 431, + 378 + ], + "spans": [ + { + "bbox": [ + 179, + 342, + 431, + 378 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { d i s t i l l } = \\frac { 1 } { M N _ { r } } \\sum _ { j = 1 } ^ { M } \\sum _ { i = 1 } ^ { N } \\delta _ { o b j } ( \\mathrm { y } _ { i } ) \\cdot < \\mathbf { m } _ { i j } , \\mathcal { L } _ { M S E } ( \\mathrm { V } _ { j } ^ { S } , \\mathrm { V } _ { j } ^ { \\mathcal { T } } ) >", + "type": "interline_equation", + "image_path": "6f7bad57fe54e67c244dfc48335122d2e93a1afcb349baf96fb24135fe4a7b95.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 179, + 342, + 431, + 354.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 179, + 354.0, + 431, + 366.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 179, + 366.0, + 431, + 378.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 101, + 379, + 509, + 405 + ], + "spans": [ + { + "bbox": [ + 101, + 379, + 133, + 405 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 384, + 218, + 399 + ], + "score": 0.91, + "content": "\\begin{array} { r } { N _ { r } = \\sum _ { i = 1 } ^ { N } \\delta _ { o b j } ( \\mathbf { y } _ { i } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 379, + 223, + 405 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 224, + 386, + 266, + 399 + ], + "score": 0.81, + "content": "( N _ { r } \\leq N )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 379, + 385, + 405 + ], + "score": 1.0, + "content": "is the real instances number,", + "type": "text" + }, + { + "bbox": [ + 386, + 385, + 479, + 400 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { M S E } ( \\mathrm { V } _ { j } ^ { S } , \\mathrm { V } _ { j } ^ { T } ) \\in \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 379, + 509, + 405 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 399, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 354, + 412 + ], + "score": 1.0, + "content": "pixel-wise mean-square error along the hidden dimension2 and", + "type": "text" + }, + { + "bbox": [ + 354, + 401, + 386, + 411 + ], + "score": 0.87, + "content": "< \\cdot , \\cdot >", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 399, + 506, + 412 + ], + "score": 1.0, + "content": "is the Dirac notation for inner", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 283, + 423 + ], + "score": 1.0, + "content": "product. For stability, the learnable variable", + "type": "text" + }, + { + "bbox": [ + 283, + 412, + 301, + 423 + ], + "score": 0.89, + "content": "\\mathbf { m } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 410, + 319, + 423 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 320, + 410, + 335, + 424 + ], + "score": 0.9, + "content": "\\mathrm { V } _ { j } ^ { \\mathcal { T } } .", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "are detached during distillation. Combine", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 421, + 495, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 239, + 434 + ], + "score": 1.0, + "content": "with the supervised learning loss", + "type": "text" + }, + { + "bbox": [ + 239, + 422, + 258, + 433 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { d e t }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 421, + 396, + 434 + ], + "score": 1.0, + "content": ", the overall loss with a coefficient", + "type": "text" + }, + { + "bbox": [ + 397, + 422, + 404, + 432 + ], + "score": 0.82, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 421, + 495, + 434 + ], + "score": 1.0, + "content": "is summarized below:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 449, + 375, + 462 + ], + "lines": [ + { + "bbox": [ + 236, + 449, + 375, + 462 + ], + "spans": [ + { + "bbox": [ + 236, + 449, + 375, + 462 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { t o t a l } = \\mathcal { L } _ { d e t } + \\mathcal { L } _ { a u x } + \\lambda \\mathcal { L } _ { d i s t i l l }", + "type": "interline_equation", + "image_path": "766645ca6dd38a224b6b0ab543eda7bfcc9f1bea20b912e061f6f3ffac7de111.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 236, + 449, + 375, + 462 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 282, + 484 + ], + "score": 1.0, + "content": "It is worth noticing, only the gradients w.r.t.", + "type": "text" + }, + { + "bbox": [ + 282, + 471, + 312, + 482 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { d i s t i l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 470, + 330, + 484 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 331, + 471, + 350, + 482 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { d e t }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "back-propagate to the student network", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 480, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 197, + 495 + ], + "score": 1.0, + "content": "(from representations", + "type": "text" + }, + { + "bbox": [ + 198, + 483, + 206, + 492 + ], + "score": 0.74, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 480, + 287, + 495 + ], + "score": 1.0, + "content": "). The gradients of", + "type": "text" + }, + { + "bbox": [ + 288, + 482, + 310, + 493 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 480, + 505, + 495 + ], + "score": 1.0, + "content": "only update the instance-conditional decoding", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 493, + 293, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 142, + 505 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 142, + 494, + 150, + 504 + ], + "score": 0.84, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 493, + 293, + 505 + ], + "score": 1.0, + "content": "and auxiliary task related modules.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "title", + "bbox": [ + 107, + 520, + 192, + 533 + ], + "lines": [ + { + "bbox": [ + 104, + 518, + 193, + 537 + ], + "spans": [ + { + "bbox": [ + 104, + 518, + 193, + 537 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 107, + 545, + 217, + 557 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 218, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 218, + 560 + ], + "score": 1.0, + "content": "4.1 Experiment Settings", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 565, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "We conduct experiments on Pytorch [34] with the widely used Detectron2 library [47] and AdelaiDet", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 575, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 589 + ], + "score": 1.0, + "content": "library 3 [40]. All experiments are running on eight 2080ti GPUs with 2 images in each. We adopt the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 120, + 598 + ], + "score": 0.86, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "scheduler, which denotes 9k iterations of training, following the standard protocols in Detectron2", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 597, + 475, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 475, + 611 + ], + "score": 1.0, + "content": "unless otherwise specified. Scale jittering with random flip is adopted as data augmentation.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 614, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 614, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 250, + 627 + ], + "score": 1.0, + "content": "For distillation, the hyper-parameter", + "type": "text" + }, + { + "bbox": [ + 250, + 615, + 257, + 624 + ], + "score": 0.77, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 614, + 505, + 627 + ], + "score": 1.0, + "content": "is set to 8 for one-stage detectors and 3 for two-stage detectors", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "respectively. To optimize the transformer decoder, we adopt AdamW optimizer [33] for the decoder", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 635, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 649 + ], + "score": 1.0, + "content": "and MLPs following common settings for transformer [43, 5]. Corresponding hyper-parameters", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "follows [5], where the initial learning rate and weight decay are set to 1e-4. We adopt the 256 hidden", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 658, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 506, + 671 + ], + "score": 1.0, + "content": "dimension for our decoder and all MLPs, the decoder has 8 heads in parallel. The projection layer", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 668, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 119, + 681 + ], + "score": 0.88, + "content": "\\mathcal { F } _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 668, + 192, + 682 + ], + "score": 1.0, + "content": "is a 3 layer MLP,", + "type": "text" + }, + { + "bbox": [ + 192, + 669, + 212, + 681 + ], + "score": 0.91, + "content": "F _ { r e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 668, + 230, + 682 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 231, + 669, + 250, + 681 + ], + "score": 0.91, + "content": "F _ { o b j }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 668, + 506, + 682 + ], + "score": 1.0, + "content": "share another 3 layer MLP. In addition, we notice some newly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 679, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 507, + 693 + ], + "score": 1.0, + "content": "initialized modules of the student share the same size of the teacher, e.g., the detection head, FPN.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 117, + 700, + 488, + 722 + ], + "lines": [ + { + "bbox": [ + 121, + 698, + 489, + 713 + ], + "spans": [ + { + "bbox": [ + 121, + 700, + 158, + 712 + ], + "score": 0.81, + "content": "^ 2 { \\cal L } _ { M S E } ( . )", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 698, + 489, + 713 + ], + "score": 1.0, + "content": "is conducted over normalized features with parameter-free LayerNorm [3] for stabilization.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 709, + 455, + 724 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 455, + 724 + ], + "score": 1.0, + "content": "3All libraries are open-sourced and public available, please refer to citations for more details.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 139 + ], + "lines": [], + "index": 2.5, + "bbox_fs": [ + 105, + 72, + 506, + 139 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 142, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 143, + 506, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 262, + 157 + ], + "score": 1.0, + "content": "Finally, the aggregated representation", + "type": "text" + }, + { + "bbox": [ + 262, + 143, + 276, + 156 + ], + "score": 0.91, + "content": "\\mathbf { g } _ { i } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 143, + 506, + 157 + ], + "score": 1.0, + "content": "are optimized by the auxiliary task. We introduces two", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 193, + 167 + ], + "score": 1.0, + "content": "predictors denoted by", + "type": "text" + }, + { + "bbox": [ + 193, + 155, + 213, + 167 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { o b j }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 154, + 231, + 167 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 231, + 155, + 251, + 166 + ], + "score": 0.9, + "content": "\\mathcal { F } _ { r e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 154, + 505, + 167 + ], + "score": 1.0, + "content": "respectively to predict identification and localization results. We", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 426, + 178 + ], + "score": 1.0, + "content": "adopt binary cross entropy loss (BCE) to optimize the real-fake identification and", + "type": "text" + }, + { + "bbox": [ + 427, + 166, + 439, + 175 + ], + "score": 0.8, + "content": "L 1", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "loss to optimize", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 176, + 167, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 167, + 190 + ], + "score": 1.0, + "content": "the regression.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 143, + 506, + 190 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 177, + 201, + 433, + 218 + ], + "lines": [ + { + "bbox": [ + 177, + 201, + 433, + 218 + ], + "spans": [ + { + "bbox": [ + 177, + 201, + 433, + 218 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { a u x } = \\mathcal { L } _ { B C E } ( \\mathcal { F } _ { o b j } ( \\mathbf { g } _ { i } ^ { T } ) , \\delta _ { o b j } ( \\mathbf { y } _ { i } ) ) + \\mathcal { L } _ { 1 } ( \\mathcal { F } _ { r e g } ( \\mathbf { g } _ { i } ^ { T } ) , \\delta _ { r e g } ( \\mathbf { y } _ { i } ) )", + "type": "interline_equation", + "image_path": "af7bb8afe0a58cd75764a645ba03e94b48dc610de26922d5f3d1ffb40796af89.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 177, + 201, + 433, + 218 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 225, + 506, + 259 + ], + "lines": [ + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 132, + 238 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 225, + 160, + 238 + ], + "score": 0.92, + "content": "\\delta _ { o b j } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 225, + 271, + 238 + ], + "score": 1.0, + "content": "is an indicator, it yields 1 if", + "type": "text" + }, + { + "bbox": [ + 271, + 227, + 281, + 237 + ], + "score": 0.86, + "content": "\\mathtt { y } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "is real and 0 otherwise. Following common practice, the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 235, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 295, + 250 + ], + "score": 1.0, + "content": "localization loss for fake examples is ignored.", + "type": "text" + }, + { + "bbox": [ + 295, + 237, + 313, + 249 + ], + "score": 0.91, + "content": "\\delta _ { r e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 235, + 506, + 250 + ], + "score": 1.0, + "content": "is the preparing function for regression targets,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 248, + 377, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 377, + 260 + ], + "score": 1.0, + "content": "following [42]. Appendix A provides more implementation details.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 225, + 506, + 260 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 272, + 270, + 284 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 271, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 271, + 286 + ], + "score": 1.0, + "content": "3.4 Instance-Conditional Distillation", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 292, + 506, + 337 + ], + "lines": [ + { + "bbox": [ + 106, + 293, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 505, + 304 + ], + "score": 1.0, + "content": "Lastly, we present the formulation for conditional knowledge distillation. We obtain the projected", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 301, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 104, + 301, + 163, + 317 + ], + "score": 1.0, + "content": "value features", + "type": "text" + }, + { + "bbox": [ + 163, + 303, + 213, + 317 + ], + "score": 0.92, + "content": "\\mathrm { V } _ { j } ^ { S } \\in \\mathbb { R } ^ { L \\times d }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 301, + 506, + 317 + ], + "score": 1.0, + "content": "of the student representations analogous to Eq. 4 in Sec. 3.2. By adopting", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 313, + 507, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 507, + 327 + ], + "score": 1.0, + "content": "the instance-aware attention mask as a measurement of correlations between feature and each instance,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 325, + 450, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 450, + 338 + ], + "score": 1.0, + "content": "we formulate the distillation loss as value features mimicking guided by the attention:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 104, + 293, + 507, + 338 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 179, + 342, + 431, + 378 + ], + "lines": [ + { + "bbox": [ + 179, + 342, + 431, + 378 + ], + "spans": [ + { + "bbox": [ + 179, + 342, + 431, + 378 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { d i s t i l l } = \\frac { 1 } { M N _ { r } } \\sum _ { j = 1 } ^ { M } \\sum _ { i = 1 } ^ { N } \\delta _ { o b j } ( \\mathrm { y } _ { i } ) \\cdot < \\mathbf { m } _ { i j } , \\mathcal { L } _ { M S E } ( \\mathrm { V } _ { j } ^ { S } , \\mathrm { V } _ { j } ^ { \\mathcal { T } } ) >", + "type": "interline_equation", + "image_path": "6f7bad57fe54e67c244dfc48335122d2e93a1afcb349baf96fb24135fe4a7b95.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 179, + 342, + 431, + 354.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 179, + 354.0, + 431, + 366.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 179, + 366.0, + 431, + 378.0 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 384, + 505, + 433 + ], + "lines": [ + { + "bbox": [ + 101, + 379, + 509, + 405 + ], + "spans": [ + { + "bbox": [ + 101, + 379, + 133, + 405 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 384, + 218, + 399 + ], + "score": 0.91, + "content": "\\begin{array} { r } { N _ { r } = \\sum _ { i = 1 } ^ { N } \\delta _ { o b j } ( \\mathbf { y } _ { i } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 379, + 223, + 405 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 224, + 386, + 266, + 399 + ], + "score": 0.81, + "content": "( N _ { r } \\leq N )", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 379, + 385, + 405 + ], + "score": 1.0, + "content": "is the real instances number,", + "type": "text" + }, + { + "bbox": [ + 386, + 385, + 479, + 400 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { M S E } ( \\mathrm { V } _ { j } ^ { S } , \\mathrm { V } _ { j } ^ { T } ) \\in \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 379, + 509, + 405 + ], + "score": 1.0, + "content": "is the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 399, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 354, + 412 + ], + "score": 1.0, + "content": "pixel-wise mean-square error along the hidden dimension2 and", + "type": "text" + }, + { + "bbox": [ + 354, + 401, + 386, + 411 + ], + "score": 0.87, + "content": "< \\cdot , \\cdot >", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 399, + 506, + 412 + ], + "score": 1.0, + "content": "is the Dirac notation for inner", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 410, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 283, + 423 + ], + "score": 1.0, + "content": "product. For stability, the learnable variable", + "type": "text" + }, + { + "bbox": [ + 283, + 412, + 301, + 423 + ], + "score": 0.89, + "content": "\\mathbf { m } _ { i j }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 410, + 319, + 423 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 320, + 410, + 335, + 424 + ], + "score": 0.9, + "content": "\\mathrm { V } _ { j } ^ { \\mathcal { T } } .", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "are detached during distillation. Combine", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 421, + 495, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 239, + 434 + ], + "score": 1.0, + "content": "with the supervised learning loss", + "type": "text" + }, + { + "bbox": [ + 239, + 422, + 258, + 433 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { d e t }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 421, + 396, + 434 + ], + "score": 1.0, + "content": ", the overall loss with a coefficient", + "type": "text" + }, + { + "bbox": [ + 397, + 422, + 404, + 432 + ], + "score": 0.82, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 421, + 495, + 434 + ], + "score": 1.0, + "content": "is summarized below:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 101, + 379, + 509, + 434 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 449, + 375, + 462 + ], + "lines": [ + { + "bbox": [ + 236, + 449, + 375, + 462 + ], + "spans": [ + { + "bbox": [ + 236, + 449, + 375, + 462 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { t o t a l } = \\mathcal { L } _ { d e t } + \\mathcal { L } _ { a u x } + \\lambda \\mathcal { L } _ { d i s t i l l }", + "type": "interline_equation", + "image_path": "766645ca6dd38a224b6b0ab543eda7bfcc9f1bea20b912e061f6f3ffac7de111.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 236, + 449, + 375, + 462 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 105, + 470, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 282, + 484 + ], + "score": 1.0, + "content": "It is worth noticing, only the gradients w.r.t.", + "type": "text" + }, + { + "bbox": [ + 282, + 471, + 312, + 482 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { d i s t i l l }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 470, + 330, + 484 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 331, + 471, + 350, + 482 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { d e t }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 470, + 506, + 484 + ], + "score": 1.0, + "content": "back-propagate to the student network", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 480, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 197, + 495 + ], + "score": 1.0, + "content": "(from representations", + "type": "text" + }, + { + "bbox": [ + 198, + 483, + 206, + 492 + ], + "score": 0.74, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 480, + 287, + 495 + ], + "score": 1.0, + "content": "). The gradients of", + "type": "text" + }, + { + "bbox": [ + 288, + 482, + 310, + 493 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { a u x }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 480, + 505, + 495 + ], + "score": 1.0, + "content": "only update the instance-conditional decoding", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 493, + 293, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 142, + 505 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 142, + 494, + 150, + 504 + ], + "score": 0.84, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 493, + 293, + 505 + ], + "score": 1.0, + "content": "and auxiliary task related modules.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 470, + 506, + 505 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 520, + 192, + 533 + ], + "lines": [ + { + "bbox": [ + 104, + 518, + 193, + 537 + ], + "spans": [ + { + "bbox": [ + 104, + 518, + 193, + 537 + ], + "score": 1.0, + "content": "4 Experiments", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "title", + "bbox": [ + 107, + 545, + 217, + 557 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 218, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 218, + 560 + ], + "score": 1.0, + "content": "4.1 Experiment Settings", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 565, + 505, + 609 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "We conduct experiments on Pytorch [34] with the widely used Detectron2 library [47] and AdelaiDet", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 575, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 505, + 589 + ], + "score": 1.0, + "content": "library 3 [40]. All experiments are running on eight 2080ti GPUs with 2 images in each. We adopt the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 120, + 598 + ], + "score": 0.86, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "scheduler, which denotes 9k iterations of training, following the standard protocols in Detectron2", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 597, + 475, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 475, + 611 + ], + "score": 1.0, + "content": "unless otherwise specified. Scale jittering with random flip is adopted as data augmentation.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 565, + 506, + 611 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 614, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 614, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 250, + 627 + ], + "score": 1.0, + "content": "For distillation, the hyper-parameter", + "type": "text" + }, + { + "bbox": [ + 250, + 615, + 257, + 624 + ], + "score": 0.77, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 614, + 505, + 627 + ], + "score": 1.0, + "content": "is set to 8 for one-stage detectors and 3 for two-stage detectors", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "respectively. To optimize the transformer decoder, we adopt AdamW optimizer [33] for the decoder", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 635, + 505, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 505, + 649 + ], + "score": 1.0, + "content": "and MLPs following common settings for transformer [43, 5]. Corresponding hyper-parameters", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 660 + ], + "score": 1.0, + "content": "follows [5], where the initial learning rate and weight decay are set to 1e-4. We adopt the 256 hidden", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 658, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 506, + 671 + ], + "score": 1.0, + "content": "dimension for our decoder and all MLPs, the decoder has 8 heads in parallel. The projection layer", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 668, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 119, + 681 + ], + "score": 0.88, + "content": "\\mathcal { F } _ { q }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 668, + 192, + 682 + ], + "score": 1.0, + "content": "is a 3 layer MLP,", + "type": "text" + }, + { + "bbox": [ + 192, + 669, + 212, + 681 + ], + "score": 0.91, + "content": "F _ { r e g }", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 668, + 230, + 682 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 231, + 669, + 250, + 681 + ], + "score": 0.91, + "content": "F _ { o b j }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 668, + 506, + 682 + ], + "score": 1.0, + "content": "share another 3 layer MLP. In addition, we notice some newly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 679, + 507, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 507, + 693 + ], + "score": 1.0, + "content": "initialized modules of the student share the same size of the teacher, e.g., the detection head, FPN.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 614, + 507, + 693 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 104, + 505, + 209 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 504, + 93 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 505, + 82 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 505, + 82 + ], + "score": 1.0, + "content": "Table 1: Comparison with previous methods on challenging benchmark MS-COCO. The method", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 80, + 459, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 327, + 96 + ], + "score": 1.0, + "content": "proposed by Li et al. [28] does not apply to RetinaNet.", + "type": "text" + }, + { + "bbox": [ + 327, + 82, + 334, + 93 + ], + "score": 0.78, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 80, + 459, + 96 + ], + "score": 1.0, + "content": "denotes the inheriting strategy.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 104, + 505, + 209 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 104, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 107, + 104, + 505, + 209 + ], + "score": 0.98, + "html": "
MethodFaster R-CNN [36] APsRetinaNet [30]
APAPMAPLAPAPsAPMAPL
Teacher w. ResNet-101 (3×)42.025.2 45.654.640.424.044.352.2
Student w. ResNet-50 (1×)37.922.4 41.149.137.423.141.648.3
+ FitNet [37]39.3 (+1.4)22.7 42.351.738.2(+0.8)21.842.648.8
+ Li et al. [28]39.5 (+1.5)23.3 43.051.41--
+ Wang et al. [45]39.2 (+1.3)23.2 42.850.438.4 (+1.0)23.342.649.1
+ Zhang et al. [51]40.0 (+2.1)23.2 43.352.539.3 (+1.9)23.443.650.6
+ Ours40.4 (+2.5)23.4 44.052.039.9 (+2.5)25.043.951.0
+Ours+40.9 (+3.0)24.544.2 53.540.7 (+3.3)24.245.052.7
", + "type": "table", + "image_path": "3bf912a07975adcbbbb3c3e894dee5b7538df695449effc856ecbd6e1d8a86a1.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 107, + 104, + 505, + 139.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 107, + 139.0, + 505, + 174.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 107, + 174.0, + 505, + 209.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 115, + 254, + 497, + 509 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 221, + 504, + 244 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "Table 2: Experiments on more detectors with ICD. Type denotes the AP score is evaluated on", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 230, + 453, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 264, + 246 + ], + "score": 1.0, + "content": "bounding box (BBox) or instance mask", + "type": "text" + }, + { + "bbox": [ + 264, + 232, + 291, + 243 + ], + "score": 0.28, + "content": "( \\tt M a s k )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 230, + 296, + 246 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 296, + 232, + 303, + 244 + ], + "score": 0.74, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 230, + 453, + 246 + ], + "score": 1.0, + "content": "denotes using the inheriting strategy.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "table_body", + "bbox": [ + 115, + 254, + 497, + 509 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 115, + 254, + 497, + 509 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 497, + 509 + ], + "score": 0.986, + "html": "
DetectorSettingTypeAPAP50AP75APsAPMAPL
FCOS [42]Teacher (3×) Student (1×)BBox42.661.645.826.246.353.8
39.458.242.424.243.449.4
Teacher: 18.8 FPS /51.2M+ Ours + Ours †41.7(+2.3)60.345.426.945.952.6
Student: 25.0 FPS /32.2M42.9(+3.5)61.646.627.8 46.854.6
Mask R-CNN [16] Teacher: 17.5 FPS /63.3M Student: 22.9 FPS /44.3MTeacher(3×) Student (1×)BBox42.963.346.826.446.656.1
38.659.542.122.542.049.9
+ Ours40.4 (+1.8)60.944.224.443.752.0
+ Ours †41.2 (+2.6)62.045.025.144.553.6
Teacher(3×) Student (1×)38.6 35.260.4 56.341.319.541.355.3
SOLOv2 [46]+ OursMask36.7 (+1.5)37.517.237.750.3
37.4 (+2.2)58.0 58.739.218.438.952.5
+ Ours † Teacher (3×)39.040.119.139.853.7
Student34.659.4 54.741.9 36.916.2 13.243.158.2 53.3
Teacher: 16.6 FPS/65.5M Student: 21.4 FPS /46.5M+ Ours + Ours †Mask37.2 (+2.6)57.639.814.837.9 40.757.0
CondInst [41]Teacher(3×)BBox38.5 (+3.9)59.041.215.942.358.9
44.663.7
Student (1×)39.748.427.547.858.4
+ Ours58.843.123.943.350.1
+ Ours †42.4 (+2.7) 43.7 (+4.0)61.5 62.946.1 47.225.3 27.146.054.3
Teacher(3×)39.847.356.6
Teacher: 16.8FPS/53.5M Student: 21.3 FPS /34.1MMask61.442.619.443.558.3
Student (1×)35.756.737.716.839.150.3
+ Ours37.8 (+2.1)59.140.417.541.454.7
+ Ours †39.1 (+3.4)60.542.019.142.657.0
", + "type": "table", + "image_path": "fb06ca2109088fc5e667745685db7a75060d162dba0e5695edc3df18782e6b02.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 115, + 254, + 497, + 339.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 115, + 339.0, + 497, + 424.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 115, + 424.0, + 497, + 509.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 6.75 + }, + { + "type": "text", + "bbox": [ + 109, + 523, + 503, + 545 + ], + "lines": [ + { + "bbox": [ + 107, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 107, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "We find initialize these modules with teacher’s parameters will lead to faster convergence, we call it", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 534, + 260, + 547 + ], + "spans": [ + { + "bbox": [ + 107, + 534, + 260, + 547 + ], + "score": 1.0, + "content": "the inheriting strategy in experiments.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 551, + 505, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Most experiments are conducted on a large scale object detection benchmark MS-COCO 4[31] with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "80 classes. We train models on MS-COCO 2017 trainval115k subset and validate on minival", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 572, + 507, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 507, + 586 + ], + "score": 1.0, + "content": "subset. Following common protocols, we report mean Average Precision (AP) as an evaluation metric,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 351, + 596 + ], + "score": 1.0, + "content": "together with AP under different thresholds and scales, i.e.,", + "type": "text" + }, + { + "bbox": [ + 352, + 583, + 374, + 595 + ], + "score": 0.89, + "content": "\\mathrm { { A P } _ { 5 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 583, + 378, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 378, + 584, + 401, + 595 + ], + "score": 0.89, + "content": "\\mathsf { A P } _ { 7 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 583, + 405, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 405, + 584, + 424, + 595 + ], + "score": 0.85, + "content": "\\mathsf { A P } _ { \\mathrm { S } }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 583, + 428, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 429, + 583, + 450, + 595 + ], + "score": 0.87, + "content": "\\mathsf { A P _ { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 583, + 454, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 455, + 583, + 474, + 595 + ], + "score": 0.89, + "content": "\\mathsf { A P } _ { \\mathrm { L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 583, + 506, + 596 + ], + "score": 1.0, + "content": ". Other", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 595, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 482, + 607 + ], + "score": 1.0, + "content": "experiments are listed in Appendix B, e.g., on VOC [12] and Cityscapes [10], more ablations.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 107, + 622, + 187, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 188, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 188, + 635 + ], + "score": 1.0, + "content": "4.2 Main Results", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 642, + 507, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 507, + 657 + ], + "score": 1.0, + "content": "Compare with state-of-the-art methods. We compare our method (ICD) with previous state-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "score": 1.0, + "content": "of-the-arts (SOTAs), including a classic distillation method FitNet [37], two typical detection KD", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "methods [28, 45], and a very recent work with strong performance from Zhang et al. [51]. The", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "score": 1.0, + "content": "comparison is conducted on two classic detectors: Faster R-CNN [36] and RetinaNet [30]. We adopt", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 701, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 118, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "4MS-COCO is publicly available, the annotations are licensed under a Creative Commons Attribution", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 712, + 486, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 486, + 722 + ], + "score": 1.0, + "content": "4.0 License and the use of the images follows Flickr Terms of Use. Refer to [31] for more details.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 104, + 505, + 209 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 504, + 93 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 505, + 82 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 505, + 82 + ], + "score": 1.0, + "content": "Table 1: Comparison with previous methods on challenging benchmark MS-COCO. The method", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 80, + 459, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 327, + 96 + ], + "score": 1.0, + "content": "proposed by Li et al. [28] does not apply to RetinaNet.", + "type": "text" + }, + { + "bbox": [ + 327, + 82, + 334, + 93 + ], + "score": 0.78, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 80, + 459, + 96 + ], + "score": 1.0, + "content": "denotes the inheriting strategy.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 104, + 505, + 209 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 104, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 107, + 104, + 505, + 209 + ], + "score": 0.98, + "html": "
MethodFaster R-CNN [36] APsRetinaNet [30]
APAPMAPLAPAPsAPMAPL
Teacher w. ResNet-101 (3×)42.025.2 45.654.640.424.044.352.2
Student w. ResNet-50 (1×)37.922.4 41.149.137.423.141.648.3
+ FitNet [37]39.3 (+1.4)22.7 42.351.738.2(+0.8)21.842.648.8
+ Li et al. [28]39.5 (+1.5)23.3 43.051.41--
+ Wang et al. [45]39.2 (+1.3)23.2 42.850.438.4 (+1.0)23.342.649.1
+ Zhang et al. [51]40.0 (+2.1)23.2 43.352.539.3 (+1.9)23.443.650.6
+ Ours40.4 (+2.5)23.4 44.052.039.9 (+2.5)25.043.951.0
+Ours+40.9 (+3.0)24.544.2 53.540.7 (+3.3)24.245.052.7
", + "type": "table", + "image_path": "3bf912a07975adcbbbb3c3e894dee5b7538df695449effc856ecbd6e1d8a86a1.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 107, + 104, + 505, + 139.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 107, + 139.0, + 505, + 174.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 107, + 174.0, + 505, + 209.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 115, + 254, + 497, + 509 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 221, + 504, + 244 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "Table 2: Experiments on more detectors with ICD. Type denotes the AP score is evaluated on", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 230, + 453, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 264, + 246 + ], + "score": 1.0, + "content": "bounding box (BBox) or instance mask", + "type": "text" + }, + { + "bbox": [ + 264, + 232, + 291, + 243 + ], + "score": 0.28, + "content": "( \\tt M a s k )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 230, + 296, + 246 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 296, + 232, + 303, + 244 + ], + "score": 0.74, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 230, + 453, + 246 + ], + "score": 1.0, + "content": "denotes using the inheriting strategy.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "table_body", + "bbox": [ + 115, + 254, + 497, + 509 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 115, + 254, + 497, + 509 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 497, + 509 + ], + "score": 0.986, + "html": "
DetectorSettingTypeAPAP50AP75APsAPMAPL
FCOS [42]Teacher (3×) Student (1×)BBox42.661.645.826.246.353.8
39.458.242.424.243.449.4
Teacher: 18.8 FPS /51.2M+ Ours + Ours †41.7(+2.3)60.345.426.945.952.6
Student: 25.0 FPS /32.2M42.9(+3.5)61.646.627.8 46.854.6
Mask R-CNN [16] Teacher: 17.5 FPS /63.3M Student: 22.9 FPS /44.3MTeacher(3×) Student (1×)BBox42.963.346.826.446.656.1
38.659.542.122.542.049.9
+ Ours40.4 (+1.8)60.944.224.443.752.0
+ Ours †41.2 (+2.6)62.045.025.144.553.6
Teacher(3×) Student (1×)38.6 35.260.4 56.341.319.541.355.3
SOLOv2 [46]+ OursMask36.7 (+1.5)37.517.237.750.3
37.4 (+2.2)58.0 58.739.218.438.952.5
+ Ours † Teacher (3×)39.040.119.139.853.7
Student34.659.4 54.741.9 36.916.2 13.243.158.2 53.3
Teacher: 16.6 FPS/65.5M Student: 21.4 FPS /46.5M+ Ours + Ours †Mask37.2 (+2.6)57.639.814.837.9 40.757.0
CondInst [41]Teacher(3×)BBox38.5 (+3.9)59.041.215.942.358.9
44.663.7
Student (1×)39.748.427.547.858.4
+ Ours58.843.123.943.350.1
+ Ours †42.4 (+2.7) 43.7 (+4.0)61.5 62.946.1 47.225.3 27.146.054.3
Teacher(3×)39.847.356.6
Teacher: 16.8FPS/53.5M Student: 21.3 FPS /34.1MMask61.442.619.443.558.3
Student (1×)35.756.737.716.839.150.3
+ Ours37.8 (+2.1)59.140.417.541.454.7
+ Ours †39.1 (+3.4)60.542.019.142.657.0
", + "type": "table", + "image_path": "fb06ca2109088fc5e667745685db7a75060d162dba0e5695edc3df18782e6b02.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 115, + 254, + 497, + 339.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 115, + 339.0, + 497, + 424.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 115, + 424.0, + 497, + 509.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 6.75 + }, + { + "type": "text", + "bbox": [ + 109, + 523, + 503, + 545 + ], + "lines": [ + { + "bbox": [ + 107, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 107, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "We find initialize these modules with teacher’s parameters will lead to faster convergence, we call it", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 107, + 534, + 260, + 547 + ], + "spans": [ + { + "bbox": [ + 107, + 534, + 260, + 547 + ], + "score": 1.0, + "content": "the inheriting strategy in experiments.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 107, + 523, + 505, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 551, + 505, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Most experiments are conducted on a large scale object detection benchmark MS-COCO 4[31] with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "80 classes. We train models on MS-COCO 2017 trainval115k subset and validate on minival", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 572, + 507, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 507, + 586 + ], + "score": 1.0, + "content": "subset. Following common protocols, we report mean Average Precision (AP) as an evaluation metric,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 351, + 596 + ], + "score": 1.0, + "content": "together with AP under different thresholds and scales, i.e.,", + "type": "text" + }, + { + "bbox": [ + 352, + 583, + 374, + 595 + ], + "score": 0.89, + "content": "\\mathrm { { A P } _ { 5 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 583, + 378, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 378, + 584, + 401, + 595 + ], + "score": 0.89, + "content": "\\mathsf { A P } _ { 7 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 583, + 405, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 405, + 584, + 424, + 595 + ], + "score": 0.85, + "content": "\\mathsf { A P } _ { \\mathrm { S } }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 583, + 428, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 429, + 583, + 450, + 595 + ], + "score": 0.87, + "content": "\\mathsf { A P _ { M } }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 583, + 454, + 596 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 455, + 583, + 474, + 595 + ], + "score": 0.89, + "content": "\\mathsf { A P } _ { \\mathrm { L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 583, + 506, + 596 + ], + "score": 1.0, + "content": ". Other", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 595, + 482, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 482, + 607 + ], + "score": 1.0, + "content": "experiments are listed in Appendix B, e.g., on VOC [12] and Cityscapes [10], more ablations.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 550, + 507, + 607 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 622, + 187, + 633 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 188, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 188, + 635 + ], + "score": 1.0, + "content": "4.2 Main Results", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 642, + 507, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 507, + 657 + ], + "score": 1.0, + "content": "Compare with state-of-the-art methods. We compare our method (ICD) with previous state-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "score": 1.0, + "content": "of-the-arts (SOTAs), including a classic distillation method FitNet [37], two typical detection KD", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "methods [28, 45], and a very recent work with strong performance from Zhang et al. [51]. The", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 506, + 690 + ], + "score": 1.0, + "content": "comparison is conducted on two classic detectors: Faster R-CNN [36] and RetinaNet [30]. We adopt", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 281, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 406, + 292 + ], + "score": 1.0, + "content": "detectron2 official released models with ResNet-101 backbone trained on", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 406, + 281, + 420, + 291 + ], + "score": 0.87, + "content": "3 \\times", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 421, + 281, + 505, + 292 + ], + "score": 1.0, + "content": "scheduler as teacher", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 290, + 483, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 245, + 305 + ], + "score": 1.0, + "content": "networks, the student is trained on", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 246, + 292, + 260, + 302 + ], + "score": 0.87, + "content": "1 \\times", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 260, + 290, + 483, + 305 + ], + "score": 1.0, + "content": "with ResNet-50 backbone following the above settings.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 642, + 507, + 690 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 93, + 508, + 269 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 137, + 70, + 471, + 82 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 68, + 473, + 86 + ], + "spans": [ + { + "bbox": [ + 136, + 68, + 316, + 86 + ], + "score": 1.0, + "content": "Table 3: Experiments on mobile backbones.", + "type": "text" + }, + { + "bbox": [ + 317, + 71, + 323, + 82 + ], + "score": 0.72, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 68, + 473, + 86 + ], + "score": 1.0, + "content": "denotes using the inheriting strategy.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 106, + 93, + 508, + 269 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 93, + 508, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 508, + 269 + ], + "score": 0.985, + "html": "
DetectorSettingBackboneAPAP50AP75APsAPMAPL
RetinaNet [30]Teacher (3×)ResNet-101[17]40.460.343.224.044.352.2
Student (1×)26.442.027.813.828.834.1
+ OursMBV2 [38]29.5 (+3.1)45.531.216.232.238.3
+ Ours †31.6(+5.2)48.533.417.634.741.3
RetinaNet [30]Teacher (3×)ResNet-101[17]40.460.343.224.044.352.2
Student (1×)34.954.837.020.938.944.8
+ OursEff-B0 [39]36.7 (+1.8)56.038.721.140.648.1
+ Ours †38.0 (+3.1)57.540.222.441.650.3
FRCNN [36]Teacher(3×)ResNet-101[17]42.062.545.925.245.654.6
Student (1×)27.244.728.814.629.635.6
+ OursMBV2 [38]30.2 (+3.0)48.032.517.032.239.1
+ Ours †31.4 (+4.2)49.433.617.633.541.3
FRCNN [36]Teacher (3×)ResNet-101[17]42.062.545.925.245.654.6
Student (1×)35.356.837.820.838.245.1
+ OursEff-B0 [39]37.0 (+1.7)58.039.621.140.048.3
+ Ours †37.9 (+2.6)58.740.821.440.949.5
", + "type": "table", + "image_path": "7b3f861f14bad7a14306b2f7365137293ea6ef759ac0cb2f8f2b061f6345e5c5.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 106, + 93, + 508, + 151.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 151.66666666666666, + 508, + 210.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 210.33333333333331, + 508, + 269.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 108, + 280, + 504, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 281, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 406, + 292 + ], + "score": 1.0, + "content": "detectron2 official released models with ResNet-101 backbone trained on", + "type": "text" + }, + { + "bbox": [ + 406, + 281, + 420, + 291 + ], + "score": 0.87, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 281, + 505, + 292 + ], + "score": 1.0, + "content": "scheduler as teacher", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 290, + 483, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 245, + 305 + ], + "score": 1.0, + "content": "networks, the student is trained on", + "type": "text" + }, + { + "bbox": [ + 246, + 292, + 260, + 302 + ], + "score": 0.87, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 290, + 483, + 305 + ], + "score": 1.0, + "content": "with ResNet-50 backbone following the above settings.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 258, + 321 + ], + "score": 1.0, + "content": "As shown in Table 1, ICD brings about", + "type": "text" + }, + { + "bbox": [ + 259, + 308, + 288, + 318 + ], + "score": 0.32, + "content": "2 . 5 \\mathrm { A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 307, + 505, + 321 + ], + "score": 1.0, + "content": "and 3.0 AP improvement for plain training and training", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "score": 1.0, + "content": "with the inheriting strategy respectively. Especially for RetinaNet, the student with distillation even", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 264, + 342 + ], + "score": 1.0, + "content": "outperforms a strong teacher trained on", + "type": "text" + }, + { + "bbox": [ + 265, + 330, + 279, + 340 + ], + "score": 0.86, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 330, + 505, + 342 + ], + "score": 1.0, + "content": "scheduler. Compare with previous SOTAs, the proposed", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 340, + 456, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 456, + 354 + ], + "score": 1.0, + "content": "method leads to a considerable margin for about 0.5 AP without the inheriting strategy.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 377 + ], + "score": 1.0, + "content": "Results on other settings. We further evaluate ICD under various detectors, e.g., a commonly used", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "anchor-free detector FCOS [42], and three detectors that have been extended to instance segmentation:", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 385, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 489, + 398 + ], + "score": 1.0, + "content": "Mask R-CNN [16], SOLOv2 [46] and CondInst [41]. We adopt networks with ResNet-101 on", + "type": "text" + }, + { + "bbox": [ + 489, + 386, + 504, + 397 + ], + "score": 0.84, + "content": "3 \\times", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 323, + 410 + ], + "score": 1.0, + "content": "scheduler as teachers and networks with ResNet-50 on", + "type": "text" + }, + { + "bbox": [ + 324, + 397, + 338, + 407 + ], + "score": 0.87, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "scheduler as students following the above", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "settings. As shown in Table 2, we observe consistent improvements for both detection and instance", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "segmentation. There are at most around 4 AP improvement on CondInst [41] on object detection and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 428, + 507, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 507, + 443 + ], + "score": 1.0, + "content": "SOLOv2 [46] on instance segmentation. Moreover, students with weaker backbone (ResNet-50 v.s.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 441, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 261, + 453 + ], + "score": 1.0, + "content": "ResNet-101) and less training images", + "type": "text" + }, + { + "bbox": [ + 261, + 441, + 282, + 452 + ], + "score": 0.5, + "content": "( 1 / 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 441, + 506, + 453 + ], + "score": 1.0, + "content": "even outperform (FCOS) or perform on par (SOLOv2,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "CondInst) with teachers. Note that ICD does not introduce extra burden during inference, our method", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 462, + 452, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 170, + 475 + ], + "score": 1.0, + "content": "improves about", + "type": "text" + }, + { + "bbox": [ + 170, + 462, + 190, + 473 + ], + "score": 0.87, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 462, + 202, + 475 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 202, + 462, + 224, + 473 + ], + "score": 0.81, + "content": "\\mathrm { F P S ^ { 5 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 462, + 274, + 475 + ], + "score": 1.0, + "content": "and reduces", + "type": "text" + }, + { + "bbox": [ + 275, + 462, + 294, + 473 + ], + "score": 0.87, + "content": "40 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 462, + 452, + 475 + ], + "score": 1.0, + "content": "of parameters compared with teachers.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "score": 1.0, + "content": "Mobile backbones. Aside from main experiments on commonly used ResNet [17], we also conduct", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 497, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 509 + ], + "score": 1.0, + "content": "experiments on mobile backbones, which are frequently used in low-power devices. We evaluate our", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "method on two prevalent architectures: MobileNet V2 (MBV2) [38] and EfficientNet-B0 (Eff-B0)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "score": 1.0, + "content": "[39]. The latter one adopts the MobileNet V2 as basis, and further extends it with advanced designs", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 530, + 325, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 325, + 542 + ], + "score": 1.0, + "content": "like stronger data augmentation and better activations.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 545, + 505, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 544, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 560 + ], + "score": 1.0, + "content": "Experiments are conducted on Faster R-CNN (abbr., FRCNN) [36] and RetinaNet [30] following", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "the above settings. As shown in Table 3, our method also significantly improves the performance on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "score": 1.0, + "content": "smaller backbones. For instance, we improve the RetinaNet with MobileNet V2 backbone with 5.2", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "score": 1.0, + "content": "AP gain and 3.1 AP gain with and without inheriting strategy respectively, and up to 3.1 AP gain for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "score": 1.0, + "content": "EfficientNet-B0. We also observe consistent improvements over Faster R-CNN, with up to 4.2 AP", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 601, + 350, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 350, + 613 + ], + "score": 1.0, + "content": "gain for MobileNet-V2 and 2.6 AP gain for EfficientNet-B0.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 107, + 625, + 201, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 624, + 203, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 203, + 638 + ], + "score": 1.0, + "content": "4.3 Ablation Studies", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 645, + 503, + 668 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "To verify the design options and the effectiveness of each component, we conduct ablation studies", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 655, + 425, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 425, + 669 + ], + "score": 1.0, + "content": "with the classic RetinaNet detector on MS-COCO following the above settings.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "Design of the auxiliary task. To better understand the role of our auxiliary task, we conduct", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "experiments to evaluate the contribution of each sub-task. Specifically, our auxiliary task is composed", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 711, + 267, + 722 + ], + "lines": [ + { + "bbox": [ + 119, + 709, + 267, + 723 + ], + "spans": [ + { + "bbox": [ + 119, + 709, + 267, + 723 + ], + "score": 1.0, + "content": "5FPS is evaluated on Nvidia Tesla V100.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 93, + 508, + 269 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 137, + 70, + 471, + 82 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 136, + 68, + 473, + 86 + ], + "spans": [ + { + "bbox": [ + 136, + 68, + 316, + 86 + ], + "score": 1.0, + "content": "Table 3: Experiments on mobile backbones.", + "type": "text" + }, + { + "bbox": [ + 317, + 71, + 323, + 82 + ], + "score": 0.72, + "content": "\\dagger", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 68, + 473, + 86 + ], + "score": 1.0, + "content": "denotes using the inheriting strategy.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 106, + 93, + 508, + 269 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 93, + 508, + 269 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 508, + 269 + ], + "score": 0.985, + "html": "
DetectorSettingBackboneAPAP50AP75APsAPMAPL
RetinaNet [30]Teacher (3×)ResNet-101[17]40.460.343.224.044.352.2
Student (1×)26.442.027.813.828.834.1
+ OursMBV2 [38]29.5 (+3.1)45.531.216.232.238.3
+ Ours †31.6(+5.2)48.533.417.634.741.3
RetinaNet [30]Teacher (3×)ResNet-101[17]40.460.343.224.044.352.2
Student (1×)34.954.837.020.938.944.8
+ OursEff-B0 [39]36.7 (+1.8)56.038.721.140.648.1
+ Ours †38.0 (+3.1)57.540.222.441.650.3
FRCNN [36]Teacher(3×)ResNet-101[17]42.062.545.925.245.654.6
Student (1×)27.244.728.814.629.635.6
+ OursMBV2 [38]30.2 (+3.0)48.032.517.032.239.1
+ Ours †31.4 (+4.2)49.433.617.633.541.3
FRCNN [36]Teacher (3×)ResNet-101[17]42.062.545.925.245.654.6
Student (1×)35.356.837.820.838.245.1
+ OursEff-B0 [39]37.0 (+1.7)58.039.621.140.048.3
+ Ours †37.9 (+2.6)58.740.821.440.949.5
", + "type": "table", + "image_path": "7b3f861f14bad7a14306b2f7365137293ea6ef759ac0cb2f8f2b061f6345e5c5.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 106, + 93, + 508, + 151.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 151.66666666666666, + 508, + 210.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 210.33333333333331, + 508, + 269.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 108, + 280, + 504, + 303 + ], + "lines": [], + "index": 4.5, + "bbox_fs": [ + 105, + 281, + 505, + 305 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 307, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 258, + 321 + ], + "score": 1.0, + "content": "As shown in Table 1, ICD brings about", + "type": "text" + }, + { + "bbox": [ + 259, + 308, + 288, + 318 + ], + "score": 0.32, + "content": "2 . 5 \\mathrm { A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 307, + 505, + 321 + ], + "score": 1.0, + "content": "and 3.0 AP improvement for plain training and training", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 506, + 332 + ], + "score": 1.0, + "content": "with the inheriting strategy respectively. Especially for RetinaNet, the student with distillation even", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 330, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 264, + 342 + ], + "score": 1.0, + "content": "outperforms a strong teacher trained on", + "type": "text" + }, + { + "bbox": [ + 265, + 330, + 279, + 340 + ], + "score": 0.86, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 330, + 505, + 342 + ], + "score": 1.0, + "content": "scheduler. Compare with previous SOTAs, the proposed", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 340, + 456, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 456, + 354 + ], + "score": 1.0, + "content": "method leads to a considerable margin for about 0.5 AP without the inheriting strategy.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 307, + 506, + 354 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 505, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 506, + 377 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 377 + ], + "score": 1.0, + "content": "Results on other settings. We further evaluate ICD under various detectors, e.g., a commonly used", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "anchor-free detector FCOS [42], and three detectors that have been extended to instance segmentation:", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 385, + 504, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 489, + 398 + ], + "score": 1.0, + "content": "Mask R-CNN [16], SOLOv2 [46] and CondInst [41]. We adopt networks with ResNet-101 on", + "type": "text" + }, + { + "bbox": [ + 489, + 386, + 504, + 397 + ], + "score": 0.84, + "content": "3 \\times", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 396, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 323, + 410 + ], + "score": 1.0, + "content": "scheduler as teachers and networks with ResNet-50 on", + "type": "text" + }, + { + "bbox": [ + 324, + 397, + 338, + 407 + ], + "score": 0.87, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 396, + 505, + 410 + ], + "score": 1.0, + "content": "scheduler as students following the above", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "settings. As shown in Table 2, we observe consistent improvements for both detection and instance", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 506, + 432 + ], + "score": 1.0, + "content": "segmentation. There are at most around 4 AP improvement on CondInst [41] on object detection and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 428, + 507, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 507, + 443 + ], + "score": 1.0, + "content": "SOLOv2 [46] on instance segmentation. Moreover, students with weaker backbone (ResNet-50 v.s.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 441, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 261, + 453 + ], + "score": 1.0, + "content": "ResNet-101) and less training images", + "type": "text" + }, + { + "bbox": [ + 261, + 441, + 282, + 452 + ], + "score": 0.5, + "content": "( 1 / 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 441, + 506, + 453 + ], + "score": 1.0, + "content": "even outperform (FCOS) or perform on par (SOLOv2,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 451, + 506, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 506, + 464 + ], + "score": 1.0, + "content": "CondInst) with teachers. Note that ICD does not introduce extra burden during inference, our method", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 462, + 452, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 170, + 475 + ], + "score": 1.0, + "content": "improves about", + "type": "text" + }, + { + "bbox": [ + 170, + 462, + 190, + 473 + ], + "score": 0.87, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 462, + 202, + 475 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 202, + 462, + 224, + 473 + ], + "score": 0.81, + "content": "\\mathrm { F P S ^ { 5 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 462, + 274, + 475 + ], + "score": 1.0, + "content": "and reduces", + "type": "text" + }, + { + "bbox": [ + 275, + 462, + 294, + 473 + ], + "score": 0.87, + "content": "40 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 462, + 452, + 475 + ], + "score": 1.0, + "content": "of parameters compared with teachers.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 363, + 507, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "score": 1.0, + "content": "Mobile backbones. Aside from main experiments on commonly used ResNet [17], we also conduct", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 497, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 509 + ], + "score": 1.0, + "content": "experiments on mobile backbones, which are frequently used in low-power devices. We evaluate our", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "method on two prevalent architectures: MobileNet V2 (MBV2) [38] and EfficientNet-B0 (Eff-B0)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "score": 1.0, + "content": "[39]. The latter one adopts the MobileNet V2 as basis, and further extends it with advanced designs", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 530, + 325, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 530, + 325, + 542 + ], + "score": 1.0, + "content": "like stronger data augmentation and better activations.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 485, + 506, + 542 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 545, + 505, + 612 + ], + "lines": [ + { + "bbox": [ + 105, + 544, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 560 + ], + "score": 1.0, + "content": "Experiments are conducted on Faster R-CNN (abbr., FRCNN) [36] and RetinaNet [30] following", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "the above settings. As shown in Table 3, our method also significantly improves the performance on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 581 + ], + "score": 1.0, + "content": "smaller backbones. For instance, we improve the RetinaNet with MobileNet V2 backbone with 5.2", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "score": 1.0, + "content": "AP gain and 3.1 AP gain with and without inheriting strategy respectively, and up to 3.1 AP gain for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "score": 1.0, + "content": "EfficientNet-B0. We also observe consistent improvements over Faster R-CNN, with up to 4.2 AP", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 601, + 350, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 350, + 613 + ], + "score": 1.0, + "content": "gain for MobileNet-V2 and 2.6 AP gain for EfficientNet-B0.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 544, + 506, + 613 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 625, + 201, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 624, + 203, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 203, + 638 + ], + "score": 1.0, + "content": "4.3 Ablation Studies", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 645, + 503, + 668 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "To verify the design options and the effectiveness of each component, we conduct ablation studies", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 655, + 425, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 425, + 669 + ], + "score": 1.0, + "content": "with the classic RetinaNet detector on MS-COCO following the above settings.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 645, + 505, + 669 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 503, + 702 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "Design of the auxiliary task. To better understand the role of our auxiliary task, we conduct", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "experiments to evaluate the contribution of each sub-task. Specifically, our auxiliary task is composed", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "score": 1.0, + "content": "of an identification task with binary cross-entropy loss and a localization task with regression loss,", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "the localization task is further augmented with a hint on bounding box scales. As shown in Table", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 265, + 106 + ], + "score": 1.0, + "content": "4, the identification task itself leads to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 265, + 95, + 296, + 105 + ], + "score": 0.4, + "content": "2 . 2 \\ : \\mathrm { A P }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 296, + 95, + 505, + 106 + ], + "score": 1.0, + "content": "gain compare with the baseline, this high-light the", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 435, + 118 + ], + "score": 1.0, + "content": "importance of knowledge on object perception. The regression task itself leads to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 435, + 106, + 465, + 116 + ], + "score": 0.32, + "content": "1 . 8 \\mathrm { A P }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 465, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "gain, and", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 473, + 128 + ], + "score": 1.0, + "content": "the scale information boosts it for extra 0.2 AP gain. Combine two of them, we achieve", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 474, + 117, + 505, + 127 + ], + "score": 0.3, + "content": "2 . 5 \\mathrm { \\ A P }", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "overall improvement, which indicates the fusion of two knowledge brings extra benefits. Note the", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "auxiliary task only update the decoder and does not introduce extra data, which is very different from", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 279, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 279, + 162 + ], + "score": 1.0, + "content": "multitask learning, e.g., Mask R-CNN[16].", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 680, + 505, + 704 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 161 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 85 + ], + "score": 1.0, + "content": "of an identification task with binary cross-entropy loss and a localization task with regression loss,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "the localization task is further augmented with a hint on bounding box scales. As shown in Table", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 265, + 106 + ], + "score": 1.0, + "content": "4, the identification task itself leads to", + "type": "text" + }, + { + "bbox": [ + 265, + 95, + 296, + 105 + ], + "score": 0.4, + "content": "2 . 2 \\ : \\mathrm { A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 95, + 505, + 106 + ], + "score": 1.0, + "content": "gain compare with the baseline, this high-light the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 435, + 118 + ], + "score": 1.0, + "content": "importance of knowledge on object perception. The regression task itself leads to", + "type": "text" + }, + { + "bbox": [ + 435, + 106, + 465, + 116 + ], + "score": 0.32, + "content": "1 . 8 \\mathrm { A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "gain, and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 473, + 128 + ], + "score": 1.0, + "content": "the scale information boosts it for extra 0.2 AP gain. Combine two of them, we achieve", + "type": "text" + }, + { + "bbox": [ + 474, + 117, + 505, + 127 + ], + "score": 0.3, + "content": "2 . 5 \\mathrm { \\ A P }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "overall improvement, which indicates the fusion of two knowledge brings extra benefits. Note the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 505, + 150 + ], + "score": 1.0, + "content": "auxiliary task only update the decoder and does not introduce extra data, which is very different from", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 149, + 279, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 279, + 162 + ], + "score": 1.0, + "content": "multitask learning, e.g., Mask R-CNN[16].", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + }, + { + "type": "table", + "bbox": [ + 126, + 190, + 486, + 260 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 187, + 167, + 422, + 180 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 165, + 423, + 182 + ], + "spans": [ + { + "bbox": [ + 187, + 165, + 423, + 182 + ], + "score": 1.0, + "content": "Table 4: Comparison with different auxiliary task designs.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 126, + 190, + 486, + 260 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 190, + 486, + 260 + ], + "spans": [ + { + "bbox": [ + 126, + 190, + 486, + 260 + ], + "score": 0.978, + "html": "
IdentificationLocalization+ ScaleAPAP50AP75APsAPMAPL
√37.456.740.323.141.648.3
39.659.242.823.444.050.4
√39.258.642.423.143.550.3
√√39.458.842.423.243.850.5
√√√39.959.443.125.043.951.0
", + "type": "table", + "image_path": "81966c3720a2a79edec3a4a7cc46ee443928549cc0034f310e601c0a9022f06d.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 126, + 190, + 486, + 213.33333333333334 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 126, + 213.33333333333334, + 486, + 236.66666666666669 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 126, + 236.66666666666669, + 486, + 260.0 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "Impact of the instance-aware attention. To verify the effectiveness of instance-aware attention", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 279, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 506, + 292 + ], + "score": 1.0, + "content": "learned by conditional computation, we directly replace it with different variations: the fine-grained", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "score": 1.0, + "content": "mask in [45], pixel-wise attention activation [51, 43], foreground mask analog to [14] and no attention", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "score": 1.0, + "content": "mask. The result in Fig. 5 shows our instance-aware mask leads to about 0.9 AP gain over the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 312, + 352, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 158, + 324 + ], + "score": 1.0, + "content": "baseline and", + "type": "text" + }, + { + "bbox": [ + 158, + 312, + 188, + 322 + ], + "score": 0.34, + "content": "0 . 4 \\ : \\mathrm { A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 312, + 352, + 324 + ], + "score": 1.0, + "content": "gain compare with the best replacement.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 143, + 353, + 468, + 423 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 197, + 330, + 414, + 343 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 196, + 329, + 414, + 344 + ], + "spans": [ + { + "bbox": [ + 196, + 329, + 414, + 344 + ], + "score": 1.0, + "content": "Table 5: Comparison with different types of attention.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "table_body", + "bbox": [ + 143, + 353, + 468, + 423 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 143, + 353, + 468, + 423 + ], + "spans": [ + { + "bbox": [ + 143, + 353, + 468, + 423 + ], + "score": 0.951, + "html": "
Attention TypeAPAP50AP75APsAPmAPL
No Attention39.058.442.123.543.249.9
Foreground Mask39.458.942.423.543.650.0
Fine-grained Mask [45]39.559.042.423.443.850.2
Attention Activation [43]39.358.642.322.643.650.1
Instance-conditional Attention39.959.443.125.043.951.0
", + "type": "table", + "image_path": "ab0fa2b066378a37d2bc46132df1422e259d8a688766ddb42de589811ac1025e.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 143, + 353, + 468, + 376.3333333333333 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 143, + 376.3333333333333, + 468, + 399.66666666666663 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 143, + 399.66666666666663, + 468, + 422.99999999999994 + ], + "spans": [], + "index": 20 + } + ] + } + ], + "index": 18.0 + }, + { + "type": "text", + "bbox": [ + 106, + 431, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 430, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 444 + ], + "score": 1.0, + "content": "Design of the decoder. We mainly verify two properties of the decoder design, as shown in Table", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "score": 1.0, + "content": "6. First, we find a proper number of heads is important for the best performance. The head number", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 454, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 505, + 465 + ], + "score": 1.0, + "content": "balances the dimension for each sub-space and the number of spaces. The best number lies around", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 464, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 475 + ], + "score": 1.0, + "content": "8, which is consistent with former studies [5]. Second, we evaluate the effectiveness of cascaded", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "decoders, as it is widely adopted in former studies [43, 5, 8]. However, we do not find significant", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "differences between different cascade levels, it might because we use a fixed teacher to provide", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 496, + 282, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 282, + 510 + ], + "score": 1.0, + "content": "knowledge, that limited the learning ability.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 115, + 552, + 275, + 611 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 157, + 537, + 235, + 547 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 156, + 536, + 236, + 549 + ], + "spans": [ + { + "bbox": [ + 156, + 536, + 236, + 549 + ], + "score": 1.0, + "content": "(a) Number of heads.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "table_body", + "bbox": [ + 115, + 552, + 275, + 611 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 115, + 552, + 275, + 611 + ], + "spans": [ + { + "bbox": [ + 115, + 552, + 275, + 611 + ], + "score": 0.974, + "html": "
HeadsAPAPsAPMAPL
139.423.743.850.5
439.724.043.951.2
839.925.043.951.0
1639.723.644.250.5
", + "type": "table", + "image_path": "8feb400515edcbaaee94c8f52f59fe111aeeaaf69f6cf3d4c6a738d2de639d84.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 115, + 552, + 275, + 566.75 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 115, + 566.75, + 275, + 581.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 115, + 581.5, + 275, + 596.25 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 115, + 596.25, + 275, + 611.0 + ], + "spans": [], + "index": 34 + } + ] + } + ], + "index": 30.75 + }, + { + "type": "table", + "bbox": [ + 329, + 558, + 492, + 605 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 181, + 519, + 428, + 531 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 181, + 518, + 429, + 531 + ], + "spans": [ + { + "bbox": [ + 181, + 518, + 429, + 531 + ], + "score": 1.0, + "content": "Table 6: Ablation on the number of heads and cascade levels.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "table_caption", + "bbox": [ + 357, + 543, + 465, + 553 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 356, + 542, + 466, + 554 + ], + "spans": [ + { + "bbox": [ + 356, + 542, + 466, + 554 + ], + "score": 1.0, + "content": "(b) Number of cascade levels.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "table_body", + "bbox": [ + 329, + 558, + 492, + 605 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 329, + 558, + 492, + 605 + ], + "spans": [ + { + "bbox": [ + 329, + 558, + 492, + 605 + ], + "score": 0.972, + "html": "
LevelsAPAPsAPMAPL
139.925.043.951.0
239.924.144.051.4
439.824.344.050.9
", + "type": "table", + "image_path": "ae7499a9a1da8cba22a496ebb3cfd8da975caa167214f496ce486abd04e1de77.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 329, + 558, + 492, + 573.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 329, + 573.6666666666666, + 492, + 589.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 329, + 589.3333333333333, + 492, + 604.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 106, + 623, + 180, + 636 + ], + "lines": [ + { + "bbox": [ + 104, + 621, + 181, + 639 + ], + "spans": [ + { + "bbox": [ + 104, + 621, + 181, + 639 + ], + "score": 1.0, + "content": "5 Discussion", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 647, + 506, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 507, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 507, + 660 + ], + "score": 1.0, + "content": "Qualitative analysis. We present a visualization of our learned instance-aware attention in Fig. 4.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 657, + 507, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 507, + 672 + ], + "score": 1.0, + "content": "We find different heads attend to different components related to each instance, e.g., salient parts,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "boundaries. This might suggest that salient parts are important and knowledge for regression mostly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 680, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 693 + ], + "score": 1.0, + "content": "lies around boundaries, instead of averaged on the foreground as assumed in former studies. The head", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "(vi) and (vii) are smoother around key regions, which may relate to context. The head (iv) mainly", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 703, + 504, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 504, + 714 + ], + "score": 1.0, + "content": "attends on image corners, which might be a degenerated case or relate to some implicit descriptors.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 161 + ], + "lines": [], + "index": 3.5, + "bbox_fs": [ + 105, + 72, + 506, + 162 + ], + "lines_deleted": true + }, + { + "type": "table", + "bbox": [ + 126, + 190, + 486, + 260 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 187, + 167, + 422, + 180 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 187, + 165, + 423, + 182 + ], + "spans": [ + { + "bbox": [ + 187, + 165, + 423, + 182 + ], + "score": 1.0, + "content": "Table 4: Comparison with different auxiliary task designs.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 126, + 190, + 486, + 260 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 190, + 486, + 260 + ], + "spans": [ + { + "bbox": [ + 126, + 190, + 486, + 260 + ], + "score": 0.978, + "html": "
IdentificationLocalization+ ScaleAPAP50AP75APsAPMAPL
√37.456.740.323.141.648.3
39.659.242.823.444.050.4
√39.258.642.423.143.550.3
√√39.458.842.423.243.850.5
√√√39.959.443.125.043.951.0
", + "type": "table", + "image_path": "81966c3720a2a79edec3a4a7cc46ee443928549cc0034f310e601c0a9022f06d.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 126, + 190, + 486, + 213.33333333333334 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 126, + 213.33333333333334, + 486, + 236.66666666666669 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 126, + 236.66666666666669, + 486, + 260.0 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 505, + 324 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "Impact of the instance-aware attention. To verify the effectiveness of instance-aware attention", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 279, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 506, + 292 + ], + "score": 1.0, + "content": "learned by conditional computation, we directly replace it with different variations: the fine-grained", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "score": 1.0, + "content": "mask in [45], pixel-wise attention activation [51, 43], foreground mask analog to [14] and no attention", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 301, + 505, + 313 + ], + "score": 1.0, + "content": "mask. The result in Fig. 5 shows our instance-aware mask leads to about 0.9 AP gain over the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 312, + 352, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 158, + 324 + ], + "score": 1.0, + "content": "baseline and", + "type": "text" + }, + { + "bbox": [ + 158, + 312, + 188, + 322 + ], + "score": 0.34, + "content": "0 . 4 \\ : \\mathrm { A P }", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 312, + 352, + 324 + ], + "score": 1.0, + "content": "gain compare with the best replacement.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 268, + 506, + 324 + ] + }, + { + "type": "table", + "bbox": [ + 143, + 353, + 468, + 423 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 197, + 330, + 414, + 343 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 196, + 329, + 414, + 344 + ], + "spans": [ + { + "bbox": [ + 196, + 329, + 414, + 344 + ], + "score": 1.0, + "content": "Table 5: Comparison with different types of attention.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "table_body", + "bbox": [ + 143, + 353, + 468, + 423 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 143, + 353, + 468, + 423 + ], + "spans": [ + { + "bbox": [ + 143, + 353, + 468, + 423 + ], + "score": 0.951, + "html": "
Attention TypeAPAP50AP75APsAPmAPL
No Attention39.058.442.123.543.249.9
Foreground Mask39.458.942.423.543.650.0
Fine-grained Mask [45]39.559.042.423.443.850.2
Attention Activation [43]39.358.642.322.643.650.1
Instance-conditional Attention39.959.443.125.043.951.0
", + "type": "table", + "image_path": "ab0fa2b066378a37d2bc46132df1422e259d8a688766ddb42de589811ac1025e.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 143, + 353, + 468, + 376.3333333333333 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 143, + 376.3333333333333, + 468, + 399.66666666666663 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 143, + 399.66666666666663, + 468, + 422.99999999999994 + ], + "spans": [], + "index": 20 + } + ] + } + ], + "index": 18.0 + }, + { + "type": "text", + "bbox": [ + 106, + 431, + 505, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 430, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 444 + ], + "score": 1.0, + "content": "Design of the decoder. We mainly verify two properties of the decoder design, as shown in Table", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "score": 1.0, + "content": "6. First, we find a proper number of heads is important for the best performance. The head number", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 454, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 505, + 465 + ], + "score": 1.0, + "content": "balances the dimension for each sub-space and the number of spaces. The best number lies around", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 464, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 475 + ], + "score": 1.0, + "content": "8, which is consistent with former studies [5]. Second, we evaluate the effectiveness of cascaded", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "decoders, as it is widely adopted in former studies [43, 5, 8]. However, we do not find significant", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "differences between different cascade levels, it might because we use a fixed teacher to provide", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 496, + 282, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 282, + 510 + ], + "score": 1.0, + "content": "knowledge, that limited the learning ability.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 430, + 506, + 510 + ] + }, + { + "type": "table", + "bbox": [ + 115, + 552, + 275, + 611 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 157, + 537, + 235, + 547 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 156, + 536, + 236, + 549 + ], + "spans": [ + { + "bbox": [ + 156, + 536, + 236, + 549 + ], + "score": 1.0, + "content": "(a) Number of heads.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "table_body", + "bbox": [ + 115, + 552, + 275, + 611 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 115, + 552, + 275, + 611 + ], + "spans": [ + { + "bbox": [ + 115, + 552, + 275, + 611 + ], + "score": 0.974, + "html": "
HeadsAPAPsAPMAPL
139.423.743.850.5
439.724.043.951.2
839.925.043.951.0
1639.723.644.250.5
", + "type": "table", + "image_path": "8feb400515edcbaaee94c8f52f59fe111aeeaaf69f6cf3d4c6a738d2de639d84.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 115, + 552, + 275, + 566.75 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 115, + 566.75, + 275, + 581.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 115, + 581.5, + 275, + 596.25 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 115, + 596.25, + 275, + 611.0 + ], + "spans": [], + "index": 34 + } + ] + } + ], + "index": 30.75 + }, + { + "type": "table", + "bbox": [ + 329, + 558, + 492, + 605 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 181, + 519, + 428, + 531 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 181, + 518, + 429, + 531 + ], + "spans": [ + { + "bbox": [ + 181, + 518, + 429, + 531 + ], + "score": 1.0, + "content": "Table 6: Ablation on the number of heads and cascade levels.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "table_caption", + "bbox": [ + 357, + 543, + 465, + 553 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 356, + 542, + 466, + 554 + ], + "spans": [ + { + "bbox": [ + 356, + 542, + 466, + 554 + ], + "score": 1.0, + "content": "(b) Number of cascade levels.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "table_body", + "bbox": [ + 329, + 558, + 492, + 605 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 329, + 558, + 492, + 605 + ], + "spans": [ + { + "bbox": [ + 329, + 558, + 492, + 605 + ], + "score": 0.972, + "html": "
LevelsAPAPsAPMAPL
139.925.043.951.0
239.924.144.051.4
439.824.344.050.9
", + "type": "table", + "image_path": "ae7499a9a1da8cba22a496ebb3cfd8da975caa167214f496ce486abd04e1de77.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 329, + 558, + 492, + 573.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 329, + 573.6666666666666, + 492, + 589.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 329, + 589.3333333333333, + 492, + 604.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 106, + 623, + 180, + 636 + ], + "lines": [ + { + "bbox": [ + 104, + 621, + 181, + 639 + ], + "spans": [ + { + "bbox": [ + 104, + 621, + 181, + 639 + ], + "score": 1.0, + "content": "5 Discussion", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 647, + 506, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 647, + 507, + 660 + ], + "spans": [ + { + "bbox": [ + 106, + 647, + 507, + 660 + ], + "score": 1.0, + "content": "Qualitative analysis. We present a visualization of our learned instance-aware attention in Fig. 4.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 657, + 507, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 507, + 672 + ], + "score": 1.0, + "content": "We find different heads attend to different components related to each instance, e.g., salient parts,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "boundaries. This might suggest that salient parts are important and knowledge for regression mostly", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 680, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 693 + ], + "score": 1.0, + "content": "lies around boundaries, instead of averaged on the foreground as assumed in former studies. The head", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 691, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 704 + ], + "score": 1.0, + "content": "(vi) and (vii) are smoother around key regions, which may relate to context. The head (iv) mainly", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 703, + 504, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 504, + 714 + ], + "score": 1.0, + "content": "attends on image corners, which might be a degenerated case or relate to some implicit descriptors.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 647, + 507, + 714 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 68, + 493, + 299 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 68, + 493, + 299 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 68, + 493, + 299 + ], + "spans": [ + { + "bbox": [ + 118, + 68, + 493, + 299 + ], + "score": 0.977, + "type": "image", + "image_path": "79a4e47f6f93555faf0423d87ee2ec1cf031a9a4f50e4f5deb905de4c0c3f216.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 68, + 493, + 145.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 145.0, + 493, + 222.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 222.0, + 493, + 299.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 109, + 308, + 504, + 330 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 107, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "Figure 4: Visualization of our learned instance-aware attention over each head. Red denotes weak", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 319, + 252, + 331 + ], + "spans": [ + { + "bbox": [ + 107, + 319, + 252, + 331 + ], + "score": 1.0, + "content": "areas and pink denotes strong areas.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 345, + 317, + 488 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 318, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 318, + 356 + ], + "score": 1.0, + "content": "Resource consumption. We benchmark the train-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 356, + 317, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 317, + 366 + ], + "score": 1.0, + "content": "ing speed for our distillation method. As shown in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 366, + 317, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 317, + 379 + ], + "score": 1.0, + "content": "Fig. 5, training the decoder introduces negligible", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 377, + 317, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 259, + 389 + ], + "score": 1.0, + "content": "cost. Specifically, we benchmark on", + "type": "text" + }, + { + "bbox": [ + 260, + 378, + 274, + 388 + ], + "score": 0.86, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 377, + 317, + 389 + ], + "score": 1.0, + "content": "scheduler", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 388, + 317, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 317, + 400 + ], + "score": 1.0, + "content": "on RetinaNet [30] with eight 2080ti, following the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 399, + 317, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 317, + 411 + ], + "score": 1.0, + "content": "same configuration in Section 4.2. The major time", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 410, + 318, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 294, + 423 + ], + "score": 1.0, + "content": "consumption is spent on training the teacher", + "type": "text" + }, + { + "bbox": [ + 294, + 410, + 315, + 421 + ], + "score": 0.89, + "content": "( 3 \\times )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 410, + 318, + 423 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 421, + 317, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 317, + 433 + ], + "score": 1.0, + "content": "which takes about 33 hours. Training (forward and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 431, + 317, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 317, + 443 + ], + "score": 1.0, + "content": "backward) the student takes about 8 hours, while", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 443, + 318, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 318, + 454 + ], + "score": 1.0, + "content": "training decoder and distillation only take 1.3 hours.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 453, + 317, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 317, + 466 + ], + "score": 1.0, + "content": "Besides the training time, memory consumption of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 464, + 318, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 318, + 476 + ], + "score": 1.0, + "content": "our method is also limited. One can update the de-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 476, + 318, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 318, + 488 + ], + "score": 1.0, + "content": "coder part and the student part in consecutive itera-", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 345, + 352, + 488, + 452 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 345, + 352, + 488, + 452 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 345, + 352, + 488, + 452 + ], + "spans": [ + { + "bbox": [ + 345, + 352, + 488, + 452 + ], + "score": 0.962, + "type": "image", + "image_path": "a47542bc79e0df0de2c5ff11492738d63ba0a63cd78f33e9abc38697a02569c0.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 345, + 352, + 488, + 402.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 345, + 402.0, + 488, + 452.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 355, + 463, + 473, + 475 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 354, + 462, + 474, + 477 + ], + "spans": [ + { + "bbox": [ + 354, + 462, + 474, + 477 + ], + "score": 1.0, + "content": "Figure 5: Time consumption.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + } + ], + "index": 19.25 + }, + { + "type": "text", + "bbox": [ + 108, + 487, + 442, + 497 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 444, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 444, + 500 + ], + "score": 1.0, + "content": "tions, leaving only intermediate features for distillation in memory across iterations.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 506, + 530 + ], + "score": 1.0, + "content": "Real-world impact. The proposed method provides a convenient approach to enhance a detector", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "under a certain setting, resulting in that a small model can perform on par with a larger one. On", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 536, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 104, + 536, + 506, + 554 + ], + "score": 1.0, + "content": "the positive side, it allows users to replace a big model with a smaller one, which reduces energy", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "consumption. On the potentially negative side, the teacher could be costly in training, also the student", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 559, + 360, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 360, + 574 + ], + "score": 1.0, + "content": "might inherit biases from the teacher, which is hard for tracing.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 106, + 594, + 183, + 607 + ], + "lines": [ + { + "bbox": [ + 104, + 592, + 185, + 610 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 185, + 610 + ], + "score": 1.0, + "content": "6 Conclusion", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "We introduce a novel framework for knowledge distillation. The proposed Instance-Conditional", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 635, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 506, + 646 + ], + "score": 1.0, + "content": "knowledge Distillation (ICD) method utilizes instance-feature cross attention to select and locate", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 645, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 506, + 657 + ], + "score": 1.0, + "content": "knowledge that correlates with human observed instances, which provides a new framework for KD.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 656, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 669 + ], + "score": 1.0, + "content": "Our formulation encodes instance as query and teacher’s representation as key. To teach the decoder", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "how to find knowledge, we design an auxiliary task that relies on knowledge for recognition and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "localization. The proposed method consistently improves various detectors, leading to impressive", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "performance gain, some student networks even surpass their teachers. In addition to our design of the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "auxiliary task, we believe there are other alternations that can cover different situations or provide a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 711, + 488, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 488, + 723 + ], + "score": 1.0, + "content": "theoretical formulation of knowledge, which will be a potential direction for further researches.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 301, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 755 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 68, + 493, + 299 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 68, + 493, + 299 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 68, + 493, + 299 + ], + "spans": [ + { + "bbox": [ + 118, + 68, + 493, + 299 + ], + "score": 0.977, + "type": "image", + "image_path": "79a4e47f6f93555faf0423d87ee2ec1cf031a9a4f50e4f5deb905de4c0c3f216.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 68, + 493, + 145.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 145.0, + 493, + 222.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 222.0, + 493, + 299.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 109, + 308, + 504, + 330 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 107, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "Figure 4: Visualization of our learned instance-aware attention over each head. Red denotes weak", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 319, + 252, + 331 + ], + "spans": [ + { + "bbox": [ + 107, + 319, + 252, + 331 + ], + "score": 1.0, + "content": "areas and pink denotes strong areas.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 345, + 317, + 488 + ], + "lines": [ + { + "bbox": [ + 106, + 345, + 318, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 318, + 356 + ], + "score": 1.0, + "content": "Resource consumption. We benchmark the train-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 356, + 317, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 317, + 366 + ], + "score": 1.0, + "content": "ing speed for our distillation method. As shown in", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 366, + 317, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 317, + 379 + ], + "score": 1.0, + "content": "Fig. 5, training the decoder introduces negligible", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 377, + 317, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 259, + 389 + ], + "score": 1.0, + "content": "cost. Specifically, we benchmark on", + "type": "text" + }, + { + "bbox": [ + 260, + 378, + 274, + 388 + ], + "score": 0.86, + "content": "1 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 377, + 317, + 389 + ], + "score": 1.0, + "content": "scheduler", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 388, + 317, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 317, + 400 + ], + "score": 1.0, + "content": "on RetinaNet [30] with eight 2080ti, following the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 399, + 317, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 317, + 411 + ], + "score": 1.0, + "content": "same configuration in Section 4.2. The major time", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 410, + 318, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 294, + 423 + ], + "score": 1.0, + "content": "consumption is spent on training the teacher", + "type": "text" + }, + { + "bbox": [ + 294, + 410, + 315, + 421 + ], + "score": 0.89, + "content": "( 3 \\times )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 410, + 318, + 423 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 421, + 317, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 317, + 433 + ], + "score": 1.0, + "content": "which takes about 33 hours. Training (forward and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 431, + 317, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 317, + 443 + ], + "score": 1.0, + "content": "backward) the student takes about 8 hours, while", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 443, + 318, + 454 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 318, + 454 + ], + "score": 1.0, + "content": "training decoder and distillation only take 1.3 hours.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 453, + 317, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 317, + 466 + ], + "score": 1.0, + "content": "Besides the training time, memory consumption of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 464, + 318, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 318, + 476 + ], + "score": 1.0, + "content": "our method is also limited. One can update the de-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 476, + 318, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 318, + 488 + ], + "score": 1.0, + "content": "coder part and the student part in consecutive itera-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 486, + 444, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 444, + 500 + ], + "score": 1.0, + "content": "tions, leaving only intermediate features for distillation in memory across iterations.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 345, + 318, + 488 + ] + }, + { + "type": "image", + "bbox": [ + 345, + 352, + 488, + 452 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 345, + 352, + 488, + 452 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 345, + 352, + 488, + 452 + ], + "spans": [ + { + "bbox": [ + 345, + 352, + 488, + 452 + ], + "score": 0.962, + "type": "image", + "image_path": "a47542bc79e0df0de2c5ff11492738d63ba0a63cd78f33e9abc38697a02569c0.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 345, + 352, + 488, + 402.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 345, + 402.0, + 488, + 452.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 355, + 463, + 473, + 475 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 354, + 462, + 474, + 477 + ], + "spans": [ + { + "bbox": [ + 354, + 462, + 474, + 477 + ], + "score": 1.0, + "content": "Figure 5: Time consumption.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + } + ], + "index": 19.25 + }, + { + "type": "text", + "bbox": [ + 108, + 487, + 442, + 497 + ], + "lines": [], + "index": 21, + "bbox_fs": [ + 106, + 486, + 444, + 500 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 516, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 506, + 530 + ], + "score": 1.0, + "content": "Real-world impact. The proposed method provides a convenient approach to enhance a detector", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "under a certain setting, resulting in that a small model can perform on par with a larger one. On", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 536, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 104, + 536, + 506, + 554 + ], + "score": 1.0, + "content": "the positive side, it allows users to replace a big model with a smaller one, which reduces energy", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 562 + ], + "score": 1.0, + "content": "consumption. On the potentially negative side, the teacher could be costly in training, also the student", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 559, + 360, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 360, + 574 + ], + "score": 1.0, + "content": "might inherit biases from the teacher, which is hard for tracing.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 104, + 515, + 506, + 574 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 594, + 183, + 607 + ], + "lines": [ + { + "bbox": [ + 104, + 592, + 185, + 610 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 185, + 610 + ], + "score": 1.0, + "content": "6 Conclusion", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 623, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 624, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 624, + 505, + 636 + ], + "score": 1.0, + "content": "We introduce a novel framework for knowledge distillation. The proposed Instance-Conditional", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 635, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 635, + 506, + 646 + ], + "score": 1.0, + "content": "knowledge Distillation (ICD) method utilizes instance-feature cross attention to select and locate", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 645, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 506, + 657 + ], + "score": 1.0, + "content": "knowledge that correlates with human observed instances, which provides a new framework for KD.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 656, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 669 + ], + "score": 1.0, + "content": "Our formulation encodes instance as query and teacher’s representation as key. To teach the decoder", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "how to find knowledge, we design an auxiliary task that relies on knowledge for recognition and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "localization. The proposed method consistently improves various detectors, leading to impressive", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "performance gain, some student networks even surpass their teachers. In addition to our design of the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 506, + 712 + ], + "score": 1.0, + "content": "auxiliary task, we believe there are other alternations that can cover different situations or provide a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 711, + 488, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 488, + 723 + ], + "score": 1.0, + "content": "theoretical formulation of knowledge, which will be a potential direction for further researches.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 624, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 70, + 339, + 85 + ], + "lines": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 96, + 505, + 119 + ], + "lines": [ + { + "bbox": [ + 105, + 95, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 110 + ], + "score": 1.0, + "content": "This paper is supported by the National Key R&D Plan of the Ministry of Science and Technology", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 107, + 464, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 464, + 120 + ], + "score": 1.0, + "content": "(Project No. 2020AAA0104400) and Beijing Academy of Artificial Intelligence (BAAI).", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 107, + 134, + 163, + 147 + ], + "lines": [ + { + "bbox": [ + 106, + 133, + 165, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 165, + 149 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 108, + 152, + 506, + 724 + ], + "lines": [ + { + "bbox": [ + 110, + 154, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 110, + 154, + 507, + 168 + ], + "score": 1.0, + "content": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 126, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 177, + 240, + 189 + ], + "spans": [ + { + "bbox": [ + 126, + 177, + 240, + 189 + ], + "score": 1.0, + "content": "answering. In CVPR, 2018.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 196, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 110, + 196, + 506, + 209 + ], + "score": 1.0, + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 206, + 432, + 220 + ], + "spans": [ + { + "bbox": [ + 126, + 206, + 432, + 220 + ], + "score": 1.0, + "content": "Zitnick, and Devi Parikh. VQA: visual question answering. In ICCV, 2015.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 109, + 225, + 507, + 241 + ], + "spans": [ + { + "bbox": [ + 109, + 225, + 507, + 241 + ], + "score": 1.0, + "content": "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. ArXiv preprint,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 236, + 155, + 251 + ], + "spans": [ + { + "bbox": [ + 126, + 236, + 155, + 251 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 110, + 256, + 507, + 272 + ], + "spans": [ + { + "bbox": [ + 110, + 256, + 507, + 272 + ], + "score": 1.0, + "content": "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 268, + 194, + 280 + ], + "spans": [ + { + "bbox": [ + 126, + 268, + 194, + 280 + ], + "score": 1.0, + "content": "In CVPR, 2018.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 109, + 287, + 507, + 302 + ], + "spans": [ + { + "bbox": [ + 109, + 287, + 507, + 302 + ], + "score": 1.0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 299, + 464, + 313 + ], + "spans": [ + { + "bbox": [ + 127, + 299, + 464, + 313 + ], + "score": 1.0, + "content": "Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 108, + 317, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 108, + 317, + 506, + 334 + ], + "score": 1.0, + "content": "[6] Guobin Chen, Wongun Choi, Xiang Yu, Tony X. Han, and Manmohan Chandraker. Learning", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 329, + 451, + 343 + ], + "spans": [ + { + "bbox": [ + 127, + 329, + 451, + 343 + ], + "score": 1.0, + "content": "efficient object detection models with knowledge distillation. In NeurIPS, 2017.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 109, + 348, + 507, + 364 + ], + "spans": [ + { + "bbox": [ + 109, + 348, + 507, + 364 + ], + "score": 1.0, + "content": "[7] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 359, + 507, + 375 + ], + "spans": [ + { + "bbox": [ + 126, + 359, + 507, + 375 + ], + "score": 1.0, + "content": "Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 370, + 507, + 386 + ], + "spans": [ + { + "bbox": [ + 126, + 370, + 507, + 386 + ], + "score": 1.0, + "content": "Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 382, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 127, + 382, + 506, + 396 + ], + "score": 1.0, + "content": "Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 128, + 394, + 390, + 406 + ], + "spans": [ + { + "bbox": [ + 128, + 394, + 390, + 406 + ], + "score": 1.0, + "content": "toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 109, + 412, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 109, + 412, + 506, + 426 + ], + "score": 1.0, + "content": "[8] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 423, + 370, + 438 + ], + "spans": [ + { + "bbox": [ + 126, + 423, + 370, + 438 + ], + "score": 1.0, + "content": "semi-supervised object detection by points. In CVPR, 2021.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 109, + 442, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 109, + 442, + 506, + 458 + ], + "score": 1.0, + "content": "[9] Xu Cheng, Zhefan Rao, Yilan Chen, and Quanshi Zhang. Explaining knowledge distillation by", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 454, + 307, + 468 + ], + "spans": [ + { + "bbox": [ + 126, + 454, + 307, + 468 + ], + "score": 1.0, + "content": "quantifying the knowledge. In CVPR, 2020.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 473, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 488 + ], + "score": 1.0, + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 485, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 127, + 485, + 505, + 499 + ], + "score": 1.0, + "content": "Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 128, + 497, + 306, + 509 + ], + "spans": [ + { + "bbox": [ + 128, + 497, + 306, + 509 + ], + "score": 1.0, + "content": "urban scene understanding. In CVPR, 2016.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "[11] Xing Dai, Zeren Jiang, Zhao Wu, Yiping Bao, Zhicheng Wang, Si Liu, and Erjin Zhou. General", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 128, + 527, + 357, + 539 + ], + "spans": [ + { + "bbox": [ + 128, + 527, + 357, + 539 + ], + "score": 1.0, + "content": "instance distillation for object detection. In CVPR, 2021.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 546, + 507, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 507, + 560 + ], + "score": 1.0, + "content": "[12] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 127, + 557, + 418, + 571 + ], + "spans": [ + { + "bbox": [ + 127, + 557, + 418, + 571 + ], + "score": 1.0, + "content": "The pascal visual object classes challenge: A retrospective. IJCV, 2015.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 576, + 507, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 576, + 507, + 591 + ], + "score": 1.0, + "content": "[13] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 127, + 588, + 210, + 602 + ], + "spans": [ + { + "bbox": [ + 127, + 588, + 210, + 602 + ], + "score": 1.0, + "content": "survey. IJCV, 2021.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 606, + 507, + 623 + ], + "spans": [ + { + "bbox": [ + 104, + 606, + 507, + 623 + ], + "score": 1.0, + "content": "[14] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu, and Chang Xu.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 618, + 396, + 632 + ], + "spans": [ + { + "bbox": [ + 127, + 618, + 396, + 632 + ], + "score": 1.0, + "content": "Distilling object detectors via decoupled features. In CVPR, 2021.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "[15] Song Han, Jeff Pool, John Tran, and William J Dally. Learning both weights and connections", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 648, + 323, + 663 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 323, + 663 + ], + "score": 1.0, + "content": "for efficient neural networks. In NeurIPS, 2015.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 669, + 507, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 507, + 683 + ], + "score": 1.0, + "content": "[16] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross B. Girshick. Mask R-CNN. In ICCV,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 678, + 155, + 694 + ], + "spans": [ + { + "bbox": [ + 126, + 678, + 155, + 694 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 698, + 506, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 715 + ], + "score": 1.0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 711, + 246, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 711, + 246, + 724 + ], + "score": 1.0, + "content": "recognition. In CVPR, 2016.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 23 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 310, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 755 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 755 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 70, + 339, + 85 + ], + "lines": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 96, + 505, + 119 + ], + "lines": [ + { + "bbox": [ + 105, + 95, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 505, + 110 + ], + "score": 1.0, + "content": "This paper is supported by the National Key R&D Plan of the Ministry of Science and Technology", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 107, + 464, + 120 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 464, + 120 + ], + "score": 1.0, + "content": "(Project No. 2020AAA0104400) and Beijing Academy of Artificial Intelligence (BAAI).", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 95, + 505, + 120 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 134, + 163, + 147 + ], + "lines": [ + { + "bbox": [ + 106, + 133, + 165, + 149 + ], + "spans": [ + { + "bbox": [ + 106, + 133, + 165, + 149 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "list", + "bbox": [ + 108, + 152, + 506, + 724 + ], + "lines": [ + { + "bbox": [ + 110, + 154, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 110, + 154, + 507, + 168 + ], + "score": 1.0, + "content": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould,", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 165, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 126, + 165, + 506, + 179 + ], + "score": 1.0, + "content": "and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 177, + 240, + 189 + ], + "spans": [ + { + "bbox": [ + 126, + 177, + 240, + 189 + ], + "score": 1.0, + "content": "answering. In CVPR, 2018.", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 196, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 110, + 196, + 506, + 209 + ], + "score": 1.0, + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 206, + 432, + 220 + ], + "spans": [ + { + "bbox": [ + 126, + 206, + 432, + 220 + ], + "score": 1.0, + "content": "Zitnick, and Devi Parikh. VQA: visual question answering. In ICCV, 2015.", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 225, + 507, + 241 + ], + "spans": [ + { + "bbox": [ + 109, + 225, + 507, + 241 + ], + "score": 1.0, + "content": "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. ArXiv preprint,", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 236, + 155, + 251 + ], + "spans": [ + { + "bbox": [ + 126, + 236, + 155, + 251 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 256, + 507, + 272 + ], + "spans": [ + { + "bbox": [ + 110, + 256, + 507, + 272 + ], + "score": 1.0, + "content": "[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection.", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 268, + 194, + 280 + ], + "spans": [ + { + "bbox": [ + 126, + 268, + 194, + 280 + ], + "score": 1.0, + "content": "In CVPR, 2018.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 287, + 507, + 302 + ], + "spans": [ + { + "bbox": [ + 109, + 287, + 507, + 302 + ], + "score": 1.0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 299, + 464, + 313 + ], + "spans": [ + { + "bbox": [ + 127, + 299, + 464, + 313 + ], + "score": 1.0, + "content": "Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 108, + 317, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 108, + 317, + 506, + 334 + ], + "score": 1.0, + "content": "[6] Guobin Chen, Wongun Choi, Xiang Yu, Tony X. Han, and Manmohan Chandraker. Learning", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 329, + 451, + 343 + ], + "spans": [ + { + "bbox": [ + 127, + 329, + 451, + 343 + ], + "score": 1.0, + "content": "efficient object detection models with knowledge distillation. In NeurIPS, 2017.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 348, + 507, + 364 + ], + "spans": [ + { + "bbox": [ + 109, + 348, + 507, + 364 + ], + "score": 1.0, + "content": "[7] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun,", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 359, + 507, + 375 + ], + "spans": [ + { + "bbox": [ + 126, + 359, + 507, + 375 + ], + "score": 1.0, + "content": "Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 370, + 507, + 386 + ], + "spans": [ + { + "bbox": [ + 126, + 370, + 507, + 386 + ], + "score": 1.0, + "content": "Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 127, + 382, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 127, + 382, + 506, + 396 + ], + "score": 1.0, + "content": "Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 128, + 394, + 390, + 406 + ], + "spans": [ + { + "bbox": [ + 128, + 394, + 390, + 406 + ], + "score": 1.0, + "content": "toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 412, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 109, + 412, + 506, + 426 + ], + "score": 1.0, + "content": "[8] Liangyu Chen, Tong Yang, Xiangyu Zhang, Wei Zhang, and Jian Sun. Points as queries: Weakly", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 423, + 370, + 438 + ], + "spans": [ + { + "bbox": [ + 126, + 423, + 370, + 438 + ], + "score": 1.0, + "content": "semi-supervised object detection by points. In CVPR, 2021.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 442, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 109, + 442, + 506, + 458 + ], + "score": 1.0, + "content": "[9] Xu Cheng, Zhefan Rao, Yilan Chen, and Quanshi Zhang. Explaining knowledge distillation by", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 454, + 307, + 468 + ], + "spans": [ + { + "bbox": [ + 126, + 454, + 307, + 468 + ], + "score": 1.0, + "content": "quantifying the knowledge. In CVPR, 2020.", + "type": "text" + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 473, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 505, + 488 + ], + "score": 1.0, + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 485, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 127, + 485, + 505, + 499 + ], + "score": 1.0, + "content": "Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 128, + 497, + 306, + 509 + ], + "spans": [ + { + "bbox": [ + 128, + 497, + 306, + 509 + ], + "score": 1.0, + "content": "urban scene understanding. In CVPR, 2016.", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "[11] Xing Dai, Zeren Jiang, Zhao Wu, Yiping Bao, Zhicheng Wang, Si Liu, and Erjin Zhou. General", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 527, + 357, + 539 + ], + "spans": [ + { + "bbox": [ + 128, + 527, + 357, + 539 + ], + "score": 1.0, + "content": "instance distillation for object detection. In CVPR, 2021.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 546, + 507, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 507, + 560 + ], + "score": 1.0, + "content": "[12] M. Everingham, S. M. A. Eslami, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman.", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 557, + 418, + 571 + ], + "spans": [ + { + "bbox": [ + 127, + 557, + 418, + 571 + ], + "score": 1.0, + "content": "The pascal visual object classes challenge: A retrospective. IJCV, 2015.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 576, + 507, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 576, + 507, + 591 + ], + "score": 1.0, + "content": "[13] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 588, + 210, + 602 + ], + "spans": [ + { + "bbox": [ + 127, + 588, + 210, + 602 + ], + "score": 1.0, + "content": "survey. IJCV, 2021.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 606, + 507, + 623 + ], + "spans": [ + { + "bbox": [ + 104, + 606, + 507, + 623 + ], + "score": 1.0, + "content": "[14] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xinghao Chen, Chunjing Xu, and Chang Xu.", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 618, + 396, + 632 + ], + "spans": [ + { + "bbox": [ + 127, + 618, + 396, + 632 + ], + "score": 1.0, + "content": "Distilling object detectors via decoupled features. In CVPR, 2021.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "[15] Song Han, Jeff Pool, John Tran, and William J Dally. Learning both weights and connections", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 648, + 323, + 663 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 323, + 663 + ], + "score": 1.0, + "content": "for efficient neural networks. In NeurIPS, 2015.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 669, + 507, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 507, + 683 + ], + "score": 1.0, + "content": "[16] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross B. Girshick. Mask R-CNN. In ICCV,", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 678, + 155, + 694 + ], + "spans": [ + { + "bbox": [ + 126, + 678, + 155, + 694 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 698, + 506, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 715 + ], + "score": 1.0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 711, + 246, + 724 + ], + "spans": [ + { + "bbox": [ + 126, + 711, + 246, + 724 + ], + "score": 1.0, + "content": "recognition. In CVPR, 2016.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[18] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 83, + 235, + 98 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 235, + 98 + ], + "score": 1.0, + "content": "networks. In ICCV, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "score": 1.0, + "content": "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 113, + 193, + 127 + ], + "spans": [ + { + "bbox": [ + 126, + 113, + 193, + 127 + ], + "score": 1.0, + "content": "NeurIPS, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 132, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 506, + 146 + ], + "score": 1.0, + "content": "[20] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 144, + 236, + 156 + ], + "spans": [ + { + "bbox": [ + 127, + 144, + 236, + 156 + ], + "score": 1.0, + "content": "detection. In CVPR, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 161, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 506, + 175 + ], + "score": 1.0, + "content": "[21] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 172, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 126, + 172, + 294, + 187 + ], + "score": 1.0, + "content": "convolutional networks. In CVPR, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "[22] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 203, + 272, + 215 + ], + "spans": [ + { + "bbox": [ + 127, + 203, + 272, + 215 + ], + "score": 1.0, + "content": "neural networks. In NeurIPS, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 221, + 507, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 507, + 235 + ], + "score": 1.0, + "content": "[23] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quan-", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 231, + 507, + 246 + ], + "spans": [ + { + "bbox": [ + 126, + 231, + 507, + 246 + ], + "score": 1.0, + "content": "tized neural networks: Training neural networks with low precision weights and activations.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 242, + 182, + 257 + ], + "spans": [ + { + "bbox": [ + 127, + 242, + 182, + 257 + ], + "score": 1.0, + "content": "JMLR, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 260, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 104, + 260, + 506, + 277 + ], + "score": 1.0, + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 275, + 328, + 285 + ], + "spans": [ + { + "bbox": [ + 128, + 275, + 328, + 285 + ], + "score": 1.0, + "content": "convolutional neural networks. In NeurIPS, 2012.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 291, + 491, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 491, + 305 + ], + "score": 1.0, + "content": "[25] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 310, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 325 + ], + "score": 1.0, + "content": "[26] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 322, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 127, + 322, + 296, + 334 + ], + "score": 1.0, + "content": "for image-text matching. In ECCV, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "[27] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 352, + 266, + 364 + ], + "spans": [ + { + "bbox": [ + 127, + 352, + 266, + 364 + ], + "score": 1.0, + "content": "efficient convnets. In ICLR, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 370, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 384 + ], + "score": 1.0, + "content": "[28] Quanquan Li, Shengying Jin, and Junjie Yan. Mimicking very efficient network for object", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 382, + 237, + 394 + ], + "spans": [ + { + "bbox": [ + 127, + 382, + 237, + 394 + ], + "score": 1.0, + "content": "detection. In CVPR, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 399, + 507, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 507, + 414 + ], + "score": 1.0, + "content": "[29] Tsung-Yi Lin, Piotr Dollár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 410, + 427, + 425 + ], + "spans": [ + { + "bbox": [ + 126, + 410, + 427, + 425 + ], + "score": 1.0, + "content": "Belongie. Feature pyramid networks for object detection. In CVPR, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 506, + 443 + ], + "score": 1.0, + "content": "[30] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollár. Focal loss for", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 441, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 128, + 441, + 287, + 453 + ], + "score": 1.0, + "content": "dense object detection. In ICCV, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 459, + 507, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 507, + 473 + ], + "score": 1.0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan,", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 470, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 127, + 470, + 506, + 482 + ], + "score": 1.0, + "content": "Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV,", + "type": "text", + "cross_page": true + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 480, + 155, + 495 + ], + "spans": [ + { + "bbox": [ + 126, + 480, + 155, + 495 + ], + "score": 1.0, + "content": "2014.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 499, + 507, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 507, + 514 + ], + "score": 1.0, + "content": "[32] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 511, + 416, + 525 + ], + "spans": [ + { + "bbox": [ + 127, + 511, + 416, + 525 + ], + "score": 1.0, + "content": "Knowledge distillation via instance relationship graph. In CVPR, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 529, + 492, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 492, + 542 + ], + "score": 1.0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 548, + 507, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 507, + 563 + ], + "score": 1.0, + "content": "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan,", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 558, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 126, + 558, + 506, + 573 + ], + "score": 1.0, + "content": "Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas", + "type": "text", + "cross_page": true + } + ], + "index": 33 + }, + { + "bbox": [ + 128, + 571, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 128, + 571, + 506, + 583 + ], + "score": 1.0, + "content": "Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy,", + "type": "text", + "cross_page": true + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 579, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 126, + 579, + 506, + 595 + ], + "score": 1.0, + "content": "Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style,", + "type": "text", + "cross_page": true + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 593, + 365, + 605 + ], + "spans": [ + { + "bbox": [ + 127, + 593, + 365, + 605 + ], + "score": 1.0, + "content": "high-performance deep learning library. In NeurIPS, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 506, + 624 + ], + "score": 1.0, + "content": "[35] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 621, + 428, + 636 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 428, + 636 + ], + "score": 1.0, + "content": "classification using binary convolutional neural networks. In ECCV, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "score": 1.0, + "content": "[36] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 651, + 397, + 664 + ], + "spans": [ + { + "bbox": [ + 126, + 651, + 397, + 664 + ], + "score": 1.0, + "content": "object detection with region proposal networks. In NeurIPS, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "[37] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 681, + 390, + 695 + ], + "spans": [ + { + "bbox": [ + 127, + 681, + 390, + 695 + ], + "score": 1.0, + "content": "Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 699, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 507, + 713 + ], + "score": 1.0, + "content": "[38] Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen.", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 711, + 416, + 723 + ], + "spans": [ + { + "bbox": [ + 128, + 711, + 416, + 723 + ], + "score": 1.0, + "content": "Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 73, + 505, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 505, + 87 + ], + "score": 1.0, + "content": "[39] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 83, + 414, + 97 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 414, + 97 + ], + "score": 1.0, + "content": "networks. In ICML, Proceedings of Machine Learning Research, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 116 + ], + "score": 1.0, + "content": "[40] Zhi Tian, Hao Chen, Xinlong Wang, Yuliang Liu, and Chunhua Shen. AdelaiDet: A toolbox for", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 113, + 415, + 127 + ], + "spans": [ + { + "bbox": [ + 127, + 113, + 415, + 127 + ], + "score": 1.0, + "content": "instance-level recognition tasks. https://git.io/adelaidet, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "score": 1.0, + "content": "[41] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 143, + 338, + 157 + ], + "spans": [ + { + "bbox": [ + 126, + 143, + 338, + 157 + ], + "score": 1.0, + "content": "In Proc. Eur. Conf. Computer Vision (ECCV), 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 161, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 505, + 176 + ], + "score": 1.0, + "content": "[42] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: fully convolutional one-stage object", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 174, + 234, + 185 + ], + "spans": [ + { + "bbox": [ + 128, + 174, + 234, + 185 + ], + "score": 1.0, + "content": "detection. In ICCV, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 191, + 507, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 507, + 206 + ], + "score": 1.0, + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 204, + 455, + 215 + ], + "spans": [ + { + "bbox": [ + 127, + 204, + 455, + 215 + ], + "score": 1.0, + "content": "Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "[44] Jianfeng Wang, Lin Song, Zeming Li, Hongbin Sun, Jian Sun, and Nanning Zheng. End-to-end", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 233, + 397, + 246 + ], + "spans": [ + { + "bbox": [ + 127, + 233, + 397, + 246 + ], + "score": 1.0, + "content": "object detection with fully convolutional network. In CVPR, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 251, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 506, + 265 + ], + "score": 1.0, + "content": "[45] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 263, + 298, + 275 + ], + "spans": [ + { + "bbox": [ + 127, + 263, + 298, + 275 + ], + "score": 1.0, + "content": "grained feature imitation. In CVPR, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 281, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 295 + ], + "score": 1.0, + "content": "[46] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. SOLOv2: Dynamic and", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 292, + 314, + 305 + ], + "spans": [ + { + "bbox": [ + 126, + 292, + 314, + 305 + ], + "score": 1.0, + "content": "fast instance segmentation. In NeurIPS, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "score": 1.0, + "content": "[47] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 322, + 399, + 335 + ], + "spans": [ + { + "bbox": [ + 127, + 322, + 399, + 335 + ], + "score": 1.0, + "content": "https://github.com/facebookresearch/detectron2, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "[48] Jianwei Yang, Jiasen Lu, Dhruv Batra, and Devi Parikh. A faster pytorch implementation of", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 352, + 446, + 366 + ], + "spans": [ + { + "bbox": [ + 126, + 352, + 446, + 366 + ], + "score": 1.0, + "content": "faster r-cnn. https://github.com/jwyang/faster-rcnn.pytorch, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 369, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 385 + ], + "score": 1.0, + "content": "[49] Li Yuan, Francis E. H. Tay, Guilin Li, Tao Wang, and Jiashi Feng. Revisiting knowledge", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 381, + 380, + 394 + ], + "spans": [ + { + "bbox": [ + 127, + 381, + 380, + 394 + ], + "score": 1.0, + "content": "distillation via label smoothing regularization. In CVPR, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "[50] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 412, + 465, + 424 + ], + "spans": [ + { + "bbox": [ + 126, + 412, + 465, + 424 + ], + "score": 1.0, + "content": "performance of convolutional neural networks via attention transfer. In ICLR, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 429, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 445 + ], + "score": 1.0, + "content": "[51] Linfeng Zhang and Kaisheng Ma. Improve object detection with feature-based knowledge", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 442, + 404, + 453 + ], + "spans": [ + { + "bbox": [ + 127, + 442, + 404, + 453 + ], + "score": 1.0, + "content": "distillation: Towards accurate and efficient detectors. In ICLR, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + } + ], + "index": 23, + "bbox_fs": [ + 104, + 154, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 44, + 507, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "[18] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 83, + 235, + 98 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 235, + 98 + ], + "score": 1.0, + "content": "networks. In ICCV, 2017.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "score": 1.0, + "content": "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 113, + 193, + 127 + ], + "spans": [ + { + "bbox": [ + 126, + 113, + 193, + 127 + ], + "score": 1.0, + "content": "NeurIPS, 2014.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 506, + 146 + ], + "score": 1.0, + "content": "[20] Han Hu, Jiayuan Gu, Zheng Zhang, Jifeng Dai, and Yichen Wei. Relation networks for object", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 144, + 236, + 156 + ], + "spans": [ + { + "bbox": [ + 127, + 144, + 236, + 156 + ], + "score": 1.0, + "content": "detection. In CVPR, 2018.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 161, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 506, + 175 + ], + "score": 1.0, + "content": "[21] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 172, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 126, + 172, + 294, + 187 + ], + "score": 1.0, + "content": "convolutional networks. In CVPR, 2017.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 191, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 506, + 205 + ], + "score": 1.0, + "content": "[22] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 203, + 272, + 215 + ], + "spans": [ + { + "bbox": [ + 127, + 203, + 272, + 215 + ], + "score": 1.0, + "content": "neural networks. In NeurIPS, 2016.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 221, + 507, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 507, + 235 + ], + "score": 1.0, + "content": "[23] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Quan-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 231, + 507, + 246 + ], + "spans": [ + { + "bbox": [ + 126, + 231, + 507, + 246 + ], + "score": 1.0, + "content": "tized neural networks: Training neural networks with low precision weights and activations.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 127, + 242, + 182, + 257 + ], + "spans": [ + { + "bbox": [ + 127, + 242, + 182, + 257 + ], + "score": 1.0, + "content": "JMLR, 2017.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 260, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 104, + 260, + 506, + 277 + ], + "score": 1.0, + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 128, + 275, + 328, + 285 + ], + "spans": [ + { + "bbox": [ + 128, + 275, + 328, + 285 + ], + "score": 1.0, + "content": "convolutional neural networks. In NeurIPS, 2012.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 291, + 491, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 491, + 305 + ], + "score": 1.0, + "content": "[25] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In ECCV, 2018.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 310, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 506, + 325 + ], + "score": 1.0, + "content": "[26] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 322, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 127, + 322, + 296, + 334 + ], + "score": 1.0, + "content": "for image-text matching. In ECCV, 2018.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "[27] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 352, + 266, + 364 + ], + "spans": [ + { + "bbox": [ + 127, + 352, + 266, + 364 + ], + "score": 1.0, + "content": "efficient convnets. In ICLR, 2017.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 370, + 506, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 384 + ], + "score": 1.0, + "content": "[28] Quanquan Li, Shengying Jin, and Junjie Yan. Mimicking very efficient network for object", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 382, + 237, + 394 + ], + "spans": [ + { + "bbox": [ + 127, + 382, + 237, + 394 + ], + "score": 1.0, + "content": "detection. In CVPR, 2017.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 399, + 507, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 507, + 414 + ], + "score": 1.0, + "content": "[29] Tsung-Yi Lin, Piotr Dollár, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 410, + 427, + 425 + ], + "spans": [ + { + "bbox": [ + 126, + 410, + 427, + 425 + ], + "score": 1.0, + "content": "Belongie. Feature pyramid networks for object detection. In CVPR, 2017.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 429, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 506, + 443 + ], + "score": 1.0, + "content": "[30] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollár. Focal loss for", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 128, + 441, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 128, + 441, + 287, + 453 + ], + "score": 1.0, + "content": "dense object detection. In ICCV, 2017.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 459, + 507, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 507, + 473 + ], + "score": 1.0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 470, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 127, + 470, + 506, + 482 + ], + "score": 1.0, + "content": "Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 480, + 155, + 495 + ], + "spans": [ + { + "bbox": [ + 126, + 480, + 155, + 495 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 499, + 507, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 499, + 507, + 514 + ], + "score": 1.0, + "content": "[32] Yufan Liu, Jiajiong Cao, Bing Li, Chunfeng Yuan, Weiming Hu, Yangxi Li, and Yunqiang Duan.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 127, + 511, + 416, + 525 + ], + "spans": [ + { + "bbox": [ + 127, + 511, + 416, + 525 + ], + "score": 1.0, + "content": "Knowledge distillation via instance relationship graph. In CVPR, 2019.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 529, + 492, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 492, + 542 + ], + "score": 1.0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 548, + 507, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 507, + 563 + ], + "score": 1.0, + "content": "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 126, + 558, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 126, + 558, + 506, + 573 + ], + "score": 1.0, + "content": "Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 128, + 571, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 128, + 571, + 506, + 583 + ], + "score": 1.0, + "content": "Köpf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 579, + 506, + 595 + ], + "spans": [ + { + "bbox": [ + 126, + 579, + 506, + 595 + ], + "score": 1.0, + "content": "Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 593, + 365, + 605 + ], + "spans": [ + { + "bbox": [ + 127, + 593, + 365, + 605 + ], + "score": 1.0, + "content": "high-performance deep learning library. In NeurIPS, 2019.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 506, + 624 + ], + "score": 1.0, + "content": "[35] Mohammad Rastegari, Vicente Ordonez, Joseph Redmon, and Ali Farhadi. Xnor-net: Imagenet", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 621, + 428, + 636 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 428, + 636 + ], + "score": 1.0, + "content": "classification using binary convolutional neural networks. In ECCV, 2016.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "score": 1.0, + "content": "[36] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 126, + 651, + 397, + 664 + ], + "spans": [ + { + "bbox": [ + 126, + 651, + 397, + 664 + ], + "score": 1.0, + "content": "object detection with region proposal networks. In NeurIPS, 2015.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "[37] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 127, + 681, + 390, + 695 + ], + "spans": [ + { + "bbox": [ + 127, + 681, + 390, + 695 + ], + "score": 1.0, + "content": "Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 699, + 507, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 507, + 713 + ], + "score": 1.0, + "content": "[38] Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 128, + 711, + 416, + 723 + ], + "spans": [ + { + "bbox": [ + 128, + 711, + 416, + 723 + ], + "score": 1.0, + "content": "Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 22 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 44, + 507, + 731 + ], + "lines": [], + "index": 22, + "bbox_fs": [ + 104, + 72, + 507, + 723 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 70, + 507, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 505, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 505, + 87 + ], + "score": 1.0, + "content": "[39] Mingxing Tan and Quoc V. Le. Efficientnet: Rethinking model scaling for convolutional neural", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 83, + 414, + 97 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 414, + 97 + ], + "score": 1.0, + "content": "networks. In ICML, Proceedings of Machine Learning Research, 2019.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 116 + ], + "score": 1.0, + "content": "[40] Zhi Tian, Hao Chen, Xinlong Wang, Yuliang Liu, and Chunhua Shen. AdelaiDet: A toolbox for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 113, + 415, + 127 + ], + "spans": [ + { + "bbox": [ + 127, + 113, + 415, + 127 + ], + "score": 1.0, + "content": "instance-level recognition tasks. https://git.io/adelaidet, 2019.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 146 + ], + "score": 1.0, + "content": "[41] Zhi Tian, Chunhua Shen, and Hao Chen. Conditional convolutions for instance segmentation.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 143, + 338, + 157 + ], + "spans": [ + { + "bbox": [ + 126, + 143, + 338, + 157 + ], + "score": 1.0, + "content": "In Proc. Eur. Conf. Computer Vision (ECCV), 2020.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 161, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 505, + 176 + ], + "score": 1.0, + "content": "[42] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: fully convolutional one-stage object", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 128, + 174, + 234, + 185 + ], + "spans": [ + { + "bbox": [ + 128, + 174, + 234, + 185 + ], + "score": 1.0, + "content": "detection. In ICCV, 2019.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 191, + 507, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 507, + 206 + ], + "score": 1.0, + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 127, + 204, + 455, + 215 + ], + "spans": [ + { + "bbox": [ + 127, + 204, + 455, + 215 + ], + "score": 1.0, + "content": "Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "[44] Jianfeng Wang, Lin Song, Zeming Li, Hongbin Sun, Jian Sun, and Nanning Zheng. End-to-end", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 233, + 397, + 246 + ], + "spans": [ + { + "bbox": [ + 127, + 233, + 397, + 246 + ], + "score": 1.0, + "content": "object detection with fully convolutional network. In CVPR, 2021.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 251, + 506, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 506, + 265 + ], + "score": 1.0, + "content": "[45] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling object detectors with fine-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 127, + 263, + 298, + 275 + ], + "spans": [ + { + "bbox": [ + 127, + 263, + 298, + 275 + ], + "score": 1.0, + "content": "grained feature imitation. In CVPR, 2019.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 281, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 295 + ], + "score": 1.0, + "content": "[46] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. SOLOv2: Dynamic and", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 292, + 314, + 305 + ], + "spans": [ + { + "bbox": [ + 126, + 292, + 314, + 305 + ], + "score": 1.0, + "content": "fast instance segmentation. In NeurIPS, 2020.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 506, + 325 + ], + "score": 1.0, + "content": "[47] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 322, + 399, + 335 + ], + "spans": [ + { + "bbox": [ + 127, + 322, + 399, + 335 + ], + "score": 1.0, + "content": "https://github.com/facebookresearch/detectron2, 2019.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 340, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 506, + 354 + ], + "score": 1.0, + "content": "[48] Jianwei Yang, Jiasen Lu, Dhruv Batra, and Devi Parikh. A faster pytorch implementation of", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 352, + 446, + 366 + ], + "spans": [ + { + "bbox": [ + 126, + 352, + 446, + 366 + ], + "score": 1.0, + "content": "faster r-cnn. https://github.com/jwyang/faster-rcnn.pytorch, 2017.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 369, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 385 + ], + "score": 1.0, + "content": "[49] Li Yuan, Francis E. H. Tay, Guilin Li, Tao Wang, and Jiashi Feng. Revisiting knowledge", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 381, + 380, + 394 + ], + "spans": [ + { + "bbox": [ + 127, + 381, + 380, + 394 + ], + "score": 1.0, + "content": "distillation via label smoothing regularization. In CVPR, 2020.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "[50] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 126, + 412, + 465, + 424 + ], + "spans": [ + { + "bbox": [ + 126, + 412, + 465, + 424 + ], + "score": 1.0, + "content": "performance of convolutional neural networks via attention transfer. In ICLR, 2017.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 429, + 505, + 445 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 445 + ], + "score": 1.0, + "content": "[51] Linfeng Zhang and Kaisheng Ma. Improve object detection with feature-based knowledge", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 442, + 404, + 453 + ], + "spans": [ + { + "bbox": [ + 127, + 442, + 404, + 453 + ], + "score": 1.0, + "content": "distillation: Towards accurate and efficient detectors. In ICLR, 2021.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 12.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 70, + 507, + 457 + ], + "lines": [], + "index": 12.5, + "bbox_fs": [ + 105, + 73, + 507, + 453 + ], + "lines_deleted": true + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file