diff --git "a/parse/dev/dLL4KXzKUpS/dLL4KXzKUpS_middle.json" "b/parse/dev/dLL4KXzKUpS/dLL4KXzKUpS_middle.json" new file mode 100644--- /dev/null +++ "b/parse/dev/dLL4KXzKUpS/dLL4KXzKUpS_middle.json" @@ -0,0 +1,34064 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 97, + 504, + 137 + ], + "lines": [ + { + "bbox": [ + 153, + 97, + 457, + 117 + ], + "spans": [ + { + "bbox": [ + 153, + 97, + 457, + 117 + ], + "score": 1.0, + "content": "Where2comm: Communication-Efficient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 116, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 506, + 140 + ], + "score": 1.0, + "content": "Collaborative Perception via Spatial Confidence Maps", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 159, + 179, + 451, + 213 + ], + "lines": [ + { + "bbox": [ + 178, + 178, + 433, + 193 + ], + "spans": [ + { + "bbox": [ + 178, + 178, + 214, + 191 + ], + "score": 1.0, + "content": "Yue Hu", + "type": "text" + }, + { + "bbox": [ + 265, + 178, + 334, + 192 + ], + "score": 1.0, + "content": "Shaoheng Fang", + "type": "text" + }, + { + "bbox": [ + 385, + 178, + 433, + 193 + ], + "score": 1.0, + "content": "Zixing Lei", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 158, + 189, + 452, + 204 + ], + "spans": [ + { + "bbox": [ + 158, + 189, + 452, + 204 + ], + "score": 1.0, + "content": "Cooperative Medianet Innovation Center, Shanghai Jiao Tong University", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 201, + 424, + 214 + ], + "spans": [ + { + "bbox": [ + 186, + 201, + 424, + 214 + ], + "score": 1.0, + "content": "{18671129361, shfang, chezacarss}@sjtu.edu.cn", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 118, + 229, + 253, + 263 + ], + "lines": [ + { + "bbox": [ + 159, + 227, + 212, + 244 + ], + "spans": [ + { + "bbox": [ + 159, + 227, + 212, + 244 + ], + "score": 1.0, + "content": "Yiqi Zhong", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 240, + 253, + 252 + ], + "spans": [ + { + "bbox": [ + 117, + 240, + 253, + 252 + ], + "score": 1.0, + "content": "University of Southern California", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 252, + 228, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 252, + 228, + 264 + ], + "score": 1.0, + "content": "yiqizhon@usc.edu", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 267, + 229, + 493, + 263 + ], + "lines": [ + { + "bbox": [ + 351, + 229, + 414, + 241 + ], + "spans": [ + { + "bbox": [ + 351, + 229, + 414, + 241 + ], + "score": 1.0, + "content": "Siheng Chen∗", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 267, + 239, + 494, + 254 + ], + "spans": [ + { + "bbox": [ + 267, + 239, + 494, + 254 + ], + "score": 1.0, + "content": "Shanghai Jiao Tong University, Shanghai AI Laboratory", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 330, + 252, + 432, + 264 + ], + "spans": [ + { + "bbox": [ + 330, + 252, + 432, + 264 + ], + "score": 1.0, + "content": "sihengc@sjtu.edu.cn", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 283, + 291, + 328, + 304 + ], + "lines": [ + { + "bbox": [ + 281, + 290, + 331, + 306 + ], + "spans": [ + { + "bbox": [ + 281, + 290, + 331, + 306 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 143, + 315, + 469, + 534 + ], + "lines": [ + { + "bbox": [ + 141, + 315, + 469, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 315, + 469, + 329 + ], + "score": 1.0, + "content": "Multi-agent collaborative perception could significantly upgrade the perception", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 326, + 469, + 339 + ], + "spans": [ + { + "bbox": [ + 141, + 326, + 469, + 339 + ], + "score": 1.0, + "content": "performance by enabling agents to share complementary information with each", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 337, + 470, + 350 + ], + "spans": [ + { + "bbox": [ + 141, + 337, + 470, + 350 + ], + "score": 1.0, + "content": "other through communication. It inevitably results in a fundamental trade-off", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 348, + 470, + 361 + ], + "spans": [ + { + "bbox": [ + 141, + 348, + 470, + 361 + ], + "score": 1.0, + "content": "between perception performance and communication bandwidth. To tackle this", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 358, + 470, + 372 + ], + "spans": [ + { + "bbox": [ + 141, + 358, + 470, + 372 + ], + "score": 1.0, + "content": "bottleneck issue, we propose a spatial confidence map, which reflects the spatial", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 371, + 469, + 382 + ], + "spans": [ + { + "bbox": [ + 141, + 371, + 469, + 382 + ], + "score": 1.0, + "content": "heterogeneity of perceptual information. It empowers agents to only share spatially", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 381, + 470, + 394 + ], + "spans": [ + { + "bbox": [ + 141, + 381, + 470, + 394 + ], + "score": 1.0, + "content": "sparse, yet perceptually critical information, contributing to where to commu-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 392, + 470, + 405 + ], + "spans": [ + { + "bbox": [ + 141, + 392, + 470, + 405 + ], + "score": 1.0, + "content": "nicate. Based on this novel spatial confidence map, we propose Where2comm,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 403, + 469, + 415 + ], + "spans": [ + { + "bbox": [ + 141, + 403, + 469, + 415 + ], + "score": 1.0, + "content": "a communication-efficient collaborative perception framework. Where2comm", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 414, + 469, + 426 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 469, + 426 + ], + "score": 1.0, + "content": "has two distinct advantages: i) it considers pragmatic compression and uses", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 425, + 469, + 437 + ], + "spans": [ + { + "bbox": [ + 141, + 425, + 469, + 437 + ], + "score": 1.0, + "content": "less communication to achieve higher perception performance by focusing on", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 140, + 435, + 471, + 448 + ], + "spans": [ + { + "bbox": [ + 140, + 435, + 471, + 448 + ], + "score": 1.0, + "content": "perceptually critical areas; and ii) it can handle varying communication band-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 447, + 470, + 458 + ], + "spans": [ + { + "bbox": [ + 141, + 447, + 470, + 458 + ], + "score": 1.0, + "content": "width by dynamically adjusting spatial areas involved in communication. To", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 457, + 470, + 469 + ], + "spans": [ + { + "bbox": [ + 141, + 457, + 470, + 469 + ], + "score": 1.0, + "content": "evaluate Where2comm, we consider 3D object detection in both real-world and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 468, + 470, + 481 + ], + "spans": [ + { + "bbox": [ + 141, + 468, + 470, + 481 + ], + "score": 1.0, + "content": "simulation scenarios with two modalities (camera/LiDAR) and two agent types", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 479, + 470, + 492 + ], + "spans": [ + { + "bbox": [ + 141, + 479, + 470, + 492 + ], + "score": 1.0, + "content": "(cars/drones) on four datasets: OPV2V, V2X-Sim, DAIR-V2X, and our origi-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 490, + 470, + 502 + ], + "spans": [ + { + "bbox": [ + 141, + 490, + 470, + 502 + ], + "score": 1.0, + "content": "nal CoPerception-UAVs. Where2comm consistently outperforms previous meth-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 142, + 501, + 470, + 513 + ], + "spans": [ + { + "bbox": [ + 142, + 501, + 321, + 513 + ], + "score": 1.0, + "content": "ods; for example, it achieves more than 100,", + "type": "text" + }, + { + "bbox": [ + 322, + 501, + 346, + 511 + ], + "score": 0.87, + "content": "0 0 0 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 501, + 470, + 513 + ], + "score": 1.0, + "content": "lower communication volume", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 511, + 469, + 523 + ], + "spans": [ + { + "bbox": [ + 141, + 511, + 469, + 523 + ], + "score": 1.0, + "content": "and still outperforms DiscoNet and V2X-ViT on OPV2V. Our code is available", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 523, + 393, + 535 + ], + "spans": [ + { + "bbox": [ + 141, + 523, + 393, + 535 + ], + "score": 1.0, + "content": "at https://github.com/MediaBrain-SJTU/where2comm.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 107, + 543, + 191, + 557 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 192, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 192, + 559 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 563, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "Collaborative perception enables multiple agents to share complementary perceptual information", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "with each other, promoting more holistic perception. It provides a new direction to fundamentally", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 584, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 506, + 598 + ], + "score": 1.0, + "content": "overcome a number of inevitable limitations of single-agent perception, such as occlusion and long-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "range issues. Related methods and systems are desperately needed in a broad range of real-world", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "applications, such as vehicle-to-everything-communication-aided autonomous driving [1–3], multi-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "robot warehouse automation system [4, 5] and multi-UAVs (unmanned aerial vehicles) for search", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "and rescue [6–8]. To realize collaborative perception, recent works have contributed high-quality", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 639, + 390, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 390, + 651 + ], + "score": 1.0, + "content": "datasets [9–11] and effective collaboration methods [12, 13, 2, 14–19].", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 653, + 505, + 697 + ], + "lines": [ + { + "bbox": [ + 105, + 653, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 665 + ], + "score": 1.0, + "content": "In this emerging field, the current biggest challenge is how to optimize the trade-off between", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 664, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 676 + ], + "score": 1.0, + "content": "perception performance and communication bandwidth. Communication systems in real-world", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 675, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 505, + 686 + ], + "score": 1.0, + "content": "scenarios are always constrained that they can hardly afford huge communication consumption in", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 686, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 506, + 698 + ], + "score": 1.0, + "content": "real-time, such as passing complete raw observations or a large volume of features. Therefore,", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 105, + 730, + 386, + 743 + ], + "spans": [ + { + "bbox": [ + 105, + 730, + 386, + 743 + ], + "score": 1.0, + "content": "36th Conference on Neural Information Processing Systems (NeurIPS 2022).", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 120, + 704, + 203, + 714 + ], + "lines": [ + { + "bbox": [ + 118, + 702, + 204, + 716 + ], + "spans": [ + { + "bbox": [ + 118, + 702, + 204, + 716 + ], + "score": 1.0, + "content": "∗Corresponding author", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 97, + 504, + 137 + ], + "lines": [ + { + "bbox": [ + 153, + 97, + 457, + 117 + ], + "spans": [ + { + "bbox": [ + 153, + 97, + 457, + 117 + ], + "score": 1.0, + "content": "Where2comm: Communication-Efficient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 116, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 506, + 140 + ], + "score": 1.0, + "content": "Collaborative Perception via Spatial Confidence Maps", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 159, + 179, + 451, + 213 + ], + "lines": [ + { + "bbox": [ + 178, + 178, + 433, + 193 + ], + "spans": [ + { + "bbox": [ + 178, + 178, + 214, + 191 + ], + "score": 1.0, + "content": "Yue Hu", + "type": "text" + }, + { + "bbox": [ + 265, + 178, + 334, + 192 + ], + "score": 1.0, + "content": "Shaoheng Fang", + "type": "text" + }, + { + "bbox": [ + 385, + 178, + 433, + 193 + ], + "score": 1.0, + "content": "Zixing Lei", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 158, + 189, + 452, + 204 + ], + "spans": [ + { + "bbox": [ + 158, + 189, + 452, + 204 + ], + "score": 1.0, + "content": "Cooperative Medianet Innovation Center, Shanghai Jiao Tong University", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 201, + 424, + 214 + ], + "spans": [ + { + "bbox": [ + 186, + 201, + 424, + 214 + ], + "score": 1.0, + "content": "{18671129361, shfang, chezacarss}@sjtu.edu.cn", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 158, + 178, + 452, + 214 + ] + }, + { + "type": "text", + "bbox": [ + 118, + 229, + 253, + 263 + ], + "lines": [ + { + "bbox": [ + 159, + 227, + 212, + 244 + ], + "spans": [ + { + "bbox": [ + 159, + 227, + 212, + 244 + ], + "score": 1.0, + "content": "Yiqi Zhong", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 240, + 253, + 252 + ], + "spans": [ + { + "bbox": [ + 117, + 240, + 253, + 252 + ], + "score": 1.0, + "content": "University of Southern California", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 252, + 228, + 264 + ], + "spans": [ + { + "bbox": [ + 141, + 252, + 228, + 264 + ], + "score": 1.0, + "content": "yiqizhon@usc.edu", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 117, + 227, + 253, + 264 + ] + }, + { + "type": "text", + "bbox": [ + 267, + 229, + 493, + 263 + ], + "lines": [ + { + "bbox": [ + 351, + 229, + 414, + 241 + ], + "spans": [ + { + "bbox": [ + 351, + 229, + 414, + 241 + ], + "score": 1.0, + "content": "Siheng Chen∗", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 267, + 239, + 494, + 254 + ], + "spans": [ + { + "bbox": [ + 267, + 239, + 494, + 254 + ], + "score": 1.0, + "content": "Shanghai Jiao Tong University, Shanghai AI Laboratory", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 330, + 252, + 432, + 264 + ], + "spans": [ + { + "bbox": [ + 330, + 252, + 432, + 264 + ], + "score": 1.0, + "content": "sihengc@sjtu.edu.cn", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 267, + 229, + 494, + 264 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 291, + 328, + 304 + ], + "lines": [ + { + "bbox": [ + 281, + 290, + 331, + 306 + ], + "spans": [ + { + "bbox": [ + 281, + 290, + 331, + 306 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 143, + 315, + 469, + 534 + ], + "lines": [ + { + "bbox": [ + 141, + 315, + 469, + 329 + ], + "spans": [ + { + "bbox": [ + 141, + 315, + 469, + 329 + ], + "score": 1.0, + "content": "Multi-agent collaborative perception could significantly upgrade the perception", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 326, + 469, + 339 + ], + "spans": [ + { + "bbox": [ + 141, + 326, + 469, + 339 + ], + "score": 1.0, + "content": "performance by enabling agents to share complementary information with each", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 337, + 470, + 350 + ], + "spans": [ + { + "bbox": [ + 141, + 337, + 470, + 350 + ], + "score": 1.0, + "content": "other through communication. It inevitably results in a fundamental trade-off", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 348, + 470, + 361 + ], + "spans": [ + { + "bbox": [ + 141, + 348, + 470, + 361 + ], + "score": 1.0, + "content": "between perception performance and communication bandwidth. To tackle this", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 358, + 470, + 372 + ], + "spans": [ + { + "bbox": [ + 141, + 358, + 470, + 372 + ], + "score": 1.0, + "content": "bottleneck issue, we propose a spatial confidence map, which reflects the spatial", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 371, + 469, + 382 + ], + "spans": [ + { + "bbox": [ + 141, + 371, + 469, + 382 + ], + "score": 1.0, + "content": "heterogeneity of perceptual information. It empowers agents to only share spatially", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 381, + 470, + 394 + ], + "spans": [ + { + "bbox": [ + 141, + 381, + 470, + 394 + ], + "score": 1.0, + "content": "sparse, yet perceptually critical information, contributing to where to commu-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 392, + 470, + 405 + ], + "spans": [ + { + "bbox": [ + 141, + 392, + 470, + 405 + ], + "score": 1.0, + "content": "nicate. Based on this novel spatial confidence map, we propose Where2comm,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 403, + 469, + 415 + ], + "spans": [ + { + "bbox": [ + 141, + 403, + 469, + 415 + ], + "score": 1.0, + "content": "a communication-efficient collaborative perception framework. Where2comm", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 414, + 469, + 426 + ], + "spans": [ + { + "bbox": [ + 141, + 414, + 469, + 426 + ], + "score": 1.0, + "content": "has two distinct advantages: i) it considers pragmatic compression and uses", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 425, + 469, + 437 + ], + "spans": [ + { + "bbox": [ + 141, + 425, + 469, + 437 + ], + "score": 1.0, + "content": "less communication to achieve higher perception performance by focusing on", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 140, + 435, + 471, + 448 + ], + "spans": [ + { + "bbox": [ + 140, + 435, + 471, + 448 + ], + "score": 1.0, + "content": "perceptually critical areas; and ii) it can handle varying communication band-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 447, + 470, + 458 + ], + "spans": [ + { + "bbox": [ + 141, + 447, + 470, + 458 + ], + "score": 1.0, + "content": "width by dynamically adjusting spatial areas involved in communication. To", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 457, + 470, + 469 + ], + "spans": [ + { + "bbox": [ + 141, + 457, + 470, + 469 + ], + "score": 1.0, + "content": "evaluate Where2comm, we consider 3D object detection in both real-world and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 468, + 470, + 481 + ], + "spans": [ + { + "bbox": [ + 141, + 468, + 470, + 481 + ], + "score": 1.0, + "content": "simulation scenarios with two modalities (camera/LiDAR) and two agent types", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 479, + 470, + 492 + ], + "spans": [ + { + "bbox": [ + 141, + 479, + 470, + 492 + ], + "score": 1.0, + "content": "(cars/drones) on four datasets: OPV2V, V2X-Sim, DAIR-V2X, and our origi-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 490, + 470, + 502 + ], + "spans": [ + { + "bbox": [ + 141, + 490, + 470, + 502 + ], + "score": 1.0, + "content": "nal CoPerception-UAVs. Where2comm consistently outperforms previous meth-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 142, + 501, + 470, + 513 + ], + "spans": [ + { + "bbox": [ + 142, + 501, + 321, + 513 + ], + "score": 1.0, + "content": "ods; for example, it achieves more than 100,", + "type": "text" + }, + { + "bbox": [ + 322, + 501, + 346, + 511 + ], + "score": 0.87, + "content": "0 0 0 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 501, + 470, + 513 + ], + "score": 1.0, + "content": "lower communication volume", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 511, + 469, + 523 + ], + "spans": [ + { + "bbox": [ + 141, + 511, + 469, + 523 + ], + "score": 1.0, + "content": "and still outperforms DiscoNet and V2X-ViT on OPV2V. Our code is available", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 523, + 393, + 535 + ], + "spans": [ + { + "bbox": [ + 141, + 523, + 393, + 535 + ], + "score": 1.0, + "content": "at https://github.com/MediaBrain-SJTU/where2comm.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 21.5, + "bbox_fs": [ + 140, + 315, + 471, + 535 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 543, + 191, + 557 + ], + "lines": [ + { + "bbox": [ + 105, + 542, + 192, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 192, + 559 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 563, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 576 + ], + "score": 1.0, + "content": "Collaborative perception enables multiple agents to share complementary perceptual information", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "with each other, promoting more holistic perception. It provides a new direction to fundamentally", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 584, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 506, + 598 + ], + "score": 1.0, + "content": "overcome a number of inevitable limitations of single-agent perception, such as occlusion and long-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 608 + ], + "score": 1.0, + "content": "range issues. Related methods and systems are desperately needed in a broad range of real-world", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 619 + ], + "score": 1.0, + "content": "applications, such as vehicle-to-everything-communication-aided autonomous driving [1–3], multi-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "robot warehouse automation system [4, 5] and multi-UAVs (unmanned aerial vehicles) for search", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "and rescue [6–8]. To realize collaborative perception, recent works have contributed high-quality", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 639, + 390, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 390, + 651 + ], + "score": 1.0, + "content": "datasets [9–11] and effective collaboration methods [12, 13, 2, 14–19].", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 563, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 653, + 505, + 697 + ], + "lines": [ + { + "bbox": [ + 105, + 653, + 505, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 665 + ], + "score": 1.0, + "content": "In this emerging field, the current biggest challenge is how to optimize the trade-off between", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 664, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 676 + ], + "score": 1.0, + "content": "perception performance and communication bandwidth. Communication systems in real-world", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 675, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 675, + 505, + 686 + ], + "score": 1.0, + "content": "scenarios are always constrained that they can hardly afford huge communication consumption in", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 686, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 506, + 698 + ], + "score": 1.0, + "content": "real-time, such as passing complete raw observations or a large volume of features. Therefore,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 267, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 506, + 280 + ], + "score": 1.0, + "content": "we cannot solely promote the perception performance without evaluating the expense of every bit", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 278, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 506, + 290 + ], + "score": 1.0, + "content": "of precious communication bandwidth. To achieve a better performance and bandwidth trade-off,", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "score": 1.0, + "content": "previous works put forth solutions from several perspectives. For example, When2com [12] considers", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 300, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 506, + 312 + ], + "score": 1.0, + "content": "a handshake mechanism which selects the most relevant collaborators; V2VNet [1] considers end-", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 310, + 507, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 507, + 325 + ], + "score": 1.0, + "content": "to-end-learning-based source coding; and DiscoNet [2] uses 1D convolution to compress message.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "score": 1.0, + "content": "However, all previous works make a plausible assumption: once two agents collaborate, they are", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 332, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 346 + ], + "score": 1.0, + "content": "obligated to share perceptual information of all spatial areas equally. This unnecessary assumption can", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 342, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 357 + ], + "score": 1.0, + "content": "hugely waste the bandwidth as a large proportion of spatial areas may contain irrelevant information", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 355, + 483, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 483, + 368 + ], + "score": 1.0, + "content": "for perception task. Figure 1 illustrates such a spatial heterogeneity of perceptual information.", + "type": "text", + "cross_page": true + } + ], + "index": 15 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 653, + 506, + 698 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 70, + 499, + 214 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 70, + 499, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 499, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 499, + 214 + ], + "score": 0.971, + "type": "image", + "image_path": "ecb2942d1760de73e2b387e0da0d9a190d5dadff0b78abe6fd6bdbc50e308f5c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 70, + 499, + 118.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 118.0, + 499, + 166.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 166.0, + 499, + 214.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 216, + 505, + 260 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 216, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 506, + 228 + ], + "score": 1.0, + "content": "Figure 1: Collaborative perception could contribute to safety-critical scenarios, where the white car", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "and the red car may collide due to occlusion. This collision could be avoided when the blue car can", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "share a message about the red car’s position. Such a message is spatially sparse, yet perceptually", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 249, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 261 + ], + "score": 1.0, + "content": "critical. Considering the precious communication bandwidth, each agent needs to speak to the point!", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 267, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 506, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 506, + 280 + ], + "score": 1.0, + "content": "we cannot solely promote the perception performance without evaluating the expense of every bit", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 278, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 506, + 290 + ], + "score": 1.0, + "content": "of precious communication bandwidth. To achieve a better performance and bandwidth trade-off,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 302 + ], + "score": 1.0, + "content": "previous works put forth solutions from several perspectives. For example, When2com [12] considers", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 300, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 506, + 312 + ], + "score": 1.0, + "content": "a handshake mechanism which selects the most relevant collaborators; V2VNet [1] considers end-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 310, + 507, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 507, + 325 + ], + "score": 1.0, + "content": "to-end-learning-based source coding; and DiscoNet [2] uses 1D convolution to compress message.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 335 + ], + "score": 1.0, + "content": "However, all previous works make a plausible assumption: once two agents collaborate, they are", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 332, + 505, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 505, + 346 + ], + "score": 1.0, + "content": "obligated to share perceptual information of all spatial areas equally. This unnecessary assumption can", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 342, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 357 + ], + "score": 1.0, + "content": "hugely waste the bandwidth as a large proportion of spatial areas may contain irrelevant information", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 355, + 483, + 368 + ], + "spans": [ + { + "bbox": [ + 106, + 355, + 483, + 368 + ], + "score": 1.0, + "content": "for perception task. Figure 1 illustrates such a spatial heterogeneity of perceptual information.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "To fill this gap, we consider a novel spatial-confidence-aware communication strategy. The core idea", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "score": 1.0, + "content": "is to enable a spatial confidence map for each agent, where each element reflects the perceptually", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "score": 1.0, + "content": "critical level of a corresponding spatial area. Based on this map, agents decide which spatial area", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 404, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 416 + ], + "score": 1.0, + "content": "(where) to communicate about. That is, each agent offers spatially sparse, yet critical features", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "to support other agents, and meanwhile requests complementary information from others through", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 426, + 454, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 454, + 438 + ], + "score": 1.0, + "content": "multi-round communication to perform efficient and mutually beneficial collaboration.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 505, + 530 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 507, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 507, + 455 + ], + "score": 1.0, + "content": "Following this strategy, we propose Where2comm, a novel communication-efficient multi-agent collab-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "orative perception framework with the guidance of spatial confidence maps; see Fig. 2. Where2comm", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "includes three key modules: i) a spatial confidence generator, which produces a spatial confidence", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 475, + 507, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 507, + 487 + ], + "score": 1.0, + "content": "map to indicate perceptually critical areas; ii) a spatial confidence-aware communication module,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "score": 1.0, + "content": "which leverages the spatial confidence map to decide where to communicate via novel message", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 497, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 508 + ], + "score": 1.0, + "content": "packing, and who to communicate via novel communication graph construction; and iii) a spatial", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 508, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 519 + ], + "score": 1.0, + "content": "confidence-aware message fusion module, which uses novel confidence-aware multi-head attention", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 519, + 468, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 468, + 531 + ], + "score": 1.0, + "content": "to fuse all messages received from other agents, upgrading the feature map for each agent.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 534, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "score": 1.0, + "content": "Where2comm has two distinct advantages. First, it promotes pragmatic compression at the feature level", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "and uses less communication to achieve higher perception performance by focusing on perceptually", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 557, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 557, + 506, + 568 + ], + "score": 1.0, + "content": "critical areas. Second, it adapts to various communication bandwidths and communication rounds,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 568, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 506, + 579 + ], + "score": 1.0, + "content": "while previous models only handle one predefined communication bandwidth and a fixed number", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "of communication rounds. To evaluate Where2comm, we consider the collaborative 3D object", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 589, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 505, + 600 + ], + "score": 1.0, + "content": "detection task on four datasets: DAIR-V2X [11], V2X-Sim [9], OPV2V [10] and our original", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "dataset CoPerception-UAVs. Our experiments cover both real-world and simulation scenarios, two", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "types of agents (cars and drones) and sensors (LiDAR and cameras). Results show that i) the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "proposed Where2comm consistently and significantly outperforms previous works in the performance-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 632, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 645 + ], + "score": 1.0, + "content": "bandwidth trade-off across multiple datasets and modalities; and ii) Where2comm achieves better", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 644, + 314, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 314, + 655 + ], + "score": 1.0, + "content": "trade-off when the communication round increases.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 107, + 662, + 202, + 675 + ], + "lines": [ + { + "bbox": [ + 104, + 661, + 203, + 677 + ], + "spans": [ + { + "bbox": [ + 104, + 661, + 203, + 677 + ], + "score": 1.0, + "content": "2 Related Works", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "Multi-agent communication. The communication strategy in multi-agent systems has been widely", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "studied [20]. Early works [21–23] often use predefined protocols or heuristics to decide how agents", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "communicate with each other. However, it is difficult to generalize those methods to complex", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "tasks. Recent works, thus, explore learning-based methods for complex scenarios. For example,", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 753 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 70, + 499, + 214 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 70, + 499, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 70, + 499, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 499, + 214 + ], + "score": 0.971, + "type": "image", + "image_path": "ecb2942d1760de73e2b387e0da0d9a190d5dadff0b78abe6fd6bdbc50e308f5c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 70, + 499, + 118.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 118.0, + 499, + 166.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 166.0, + 499, + 214.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 216, + 505, + 260 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 216, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 506, + 228 + ], + "score": 1.0, + "content": "Figure 1: Collaborative perception could contribute to safety-critical scenarios, where the white car", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 505, + 239 + ], + "score": 1.0, + "content": "and the red car may collide due to occlusion. This collision could be avoided when the blue car can", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "share a message about the red car’s position. Such a message is spatially sparse, yet perceptually", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 249, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 261 + ], + "score": 1.0, + "content": "critical. Considering the precious communication bandwidth, each agent needs to speak to the point!", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 107, + 267, + 505, + 366 + ], + "lines": [], + "index": 11, + "bbox_fs": [ + 105, + 267, + 507, + 368 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 505, + 383 + ], + "score": 1.0, + "content": "To fill this gap, we consider a novel spatial-confidence-aware communication strategy. The core idea", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "score": 1.0, + "content": "is to enable a spatial confidence map for each agent, where each element reflects the perceptually", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "score": 1.0, + "content": "critical level of a corresponding spatial area. Based on this map, agents decide which spatial area", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 404, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 416 + ], + "score": 1.0, + "content": "(where) to communicate about. That is, each agent offers spatially sparse, yet critical features", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "to support other agents, and meanwhile requests complementary information from others through", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 426, + 454, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 454, + 438 + ], + "score": 1.0, + "content": "multi-round communication to perform efficient and mutually beneficial collaboration.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 370, + 506, + 438 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 505, + 530 + ], + "lines": [ + { + "bbox": [ + 105, + 442, + 507, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 507, + 455 + ], + "score": 1.0, + "content": "Following this strategy, we propose Where2comm, a novel communication-efficient multi-agent collab-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 505, + 465 + ], + "score": 1.0, + "content": "orative perception framework with the guidance of spatial confidence maps; see Fig. 2. Where2comm", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "includes three key modules: i) a spatial confidence generator, which produces a spatial confidence", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 475, + 507, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 507, + 487 + ], + "score": 1.0, + "content": "map to indicate perceptually critical areas; ii) a spatial confidence-aware communication module,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "score": 1.0, + "content": "which leverages the spatial confidence map to decide where to communicate via novel message", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 497, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 508 + ], + "score": 1.0, + "content": "packing, and who to communicate via novel communication graph construction; and iii) a spatial", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 508, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 519 + ], + "score": 1.0, + "content": "confidence-aware message fusion module, which uses novel confidence-aware multi-head attention", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 519, + 468, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 468, + 531 + ], + "score": 1.0, + "content": "to fuse all messages received from other agents, upgrading the feature map for each agent.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 442, + 507, + 531 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 534, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 505, + 547 + ], + "score": 1.0, + "content": "Where2comm has two distinct advantages. First, it promotes pragmatic compression at the feature level", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "and uses less communication to achieve higher perception performance by focusing on perceptually", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 557, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 557, + 506, + 568 + ], + "score": 1.0, + "content": "critical areas. Second, it adapts to various communication bandwidths and communication rounds,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 568, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 506, + 579 + ], + "score": 1.0, + "content": "while previous models only handle one predefined communication bandwidth and a fixed number", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "of communication rounds. To evaluate Where2comm, we consider the collaborative 3D object", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 589, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 505, + 600 + ], + "score": 1.0, + "content": "detection task on four datasets: DAIR-V2X [11], V2X-Sim [9], OPV2V [10] and our original", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "dataset CoPerception-UAVs. Our experiments cover both real-world and simulation scenarios, two", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 624 + ], + "score": 1.0, + "content": "types of agents (cars and drones) and sensors (LiDAR and cameras). Results show that i) the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "proposed Where2comm consistently and significantly outperforms previous works in the performance-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 632, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 645 + ], + "score": 1.0, + "content": "bandwidth trade-off across multiple datasets and modalities; and ii) Where2comm achieves better", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 644, + 314, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 314, + 655 + ], + "score": 1.0, + "content": "trade-off when the communication round increases.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 535, + 506, + 655 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 662, + 202, + 675 + ], + "lines": [ + { + "bbox": [ + 104, + 661, + 203, + 677 + ], + "spans": [ + { + "bbox": [ + 104, + 661, + 203, + 677 + ], + "score": 1.0, + "content": "2 Related Works", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 678, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "Multi-agent communication. The communication strategy in multi-agent systems has been widely", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "studied [20]. Early works [21–23] often use predefined protocols or heuristics to decide how agents", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "communicate with each other. However, it is difficult to generalize those methods to complex", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "tasks. Recent works, thus, explore learning-based methods for complex scenarios. For example,", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 678, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 89, + 497, + 147 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 149, + 78, + 460, + 89 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 148, + 76, + 462, + 92 + ], + "spans": [ + { + "bbox": [ + 148, + 76, + 462, + 92 + ], + "score": 1.0, + "content": "Table 1: Major components comparisons of collaborative perception systems.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 110, + 89, + 497, + 147 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 89, + 497, + 147 + ], + "spans": [ + { + "bbox": [ + 110, + 89, + 497, + 147 + ], + "score": 0.978, + "html": "
MethodVenueMessage packingCommunication graph constructionMessage fusion
When2com[12]CVPR2020Full feature mapHandshake-based sparse graphAttention per-agent
V2VNet[1]ECCV 2020Full feature mapFully connected graphAverage per-agent
DiscoNet [2]NeurIPS 2021Full feature mapFully connected graphMLP-based attention per-location
V2X-ViT[26]ECCV2022Full feature mapFully connected graphSelf-attention per-location
Where2commNeurIPS 2022Confidence-aware sparse feature map + request mapConfidence-aware sparse graphConfidence-awaremulti-head attention per-location
", + "type": "table", + "image_path": "c3c501a4a0c57a095e7724ab343c779a6d6cc5cbfd5d7d3759f74ca54ad62760.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 110, + 89, + 497, + 108.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 108.33333333333333, + 497, + 127.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 110, + 127.66666666666666, + 497, + 147.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 106, + 151, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "score": 1.0, + "content": "CommNet [24] learns continuous communication in the multi-agent system. Vain [25] adopts the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 162, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 505, + 173 + ], + "score": 1.0, + "content": "attention mechanism to help agents selectively fuse the information from others. Most of these", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 173, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 506, + 185 + ], + "score": 1.0, + "content": "previous works consider decision-making tasks and adopt reinforcement learning due to the lack", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "of explicit supervision. In this work, we focus on the perception task. Based on direct perception", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 195, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 206 + ], + "score": 1.0, + "content": "supervision, we apply supervised learning to optimize the communication strategy in both trade-off", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 206, + 282, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 282, + 218 + ], + "score": 1.0, + "content": "perception ability and communication cost.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 506, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 506, + 235 + ], + "score": 1.0, + "content": "Collaborative perception. As a recent application of multi-agent communication systems to per-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 233, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 506, + 245 + ], + "score": 1.0, + "content": "ception tasks, collaborative perception is still immature. To support this area of research, there is a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 243, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 258 + ], + "score": 1.0, + "content": "surge of high-quality datasets (e.g., V2X-Sim [9], OpenV2V [10], Comap[27] and DAIR-V2X[11]),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 254, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 506, + 267 + ], + "score": 1.0, + "content": "as well as collaboration methods aimed for better performance-bandwidth trade-off (see comparisons", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 264, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 104, + 264, + 506, + 279 + ], + "score": 1.0, + "content": "in Table 1). When2com [12] proposes a handshake communication mechanism to decide when to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 506, + 291 + ], + "score": 1.0, + "content": "communicate and create sparse communication graph. V2VNet [1] proposes multi-round message", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 288, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 300 + ], + "score": 1.0, + "content": "passing based on graph neural networks to achieve better perception and prediction performance.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "DiscoNet [2] adopts knowledge distillation to take the advantage of both early and intermediate", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 308, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 323 + ], + "score": 1.0, + "content": "collaboration. OPV2V [10] proposes a graph-based attentive intermediate fusion to improve percep-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "tion performances. V2X-ViT [26] introduces a novel heterogeneous multi-agent attention module", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "to fuse information across heterogeneous agents. In this work, we leverage the proposed spatial", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "score": 1.0, + "content": "confidence map to promote more compact messages, more sparse communication graphs, and more", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 353, + 392, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 392, + 365 + ], + "score": 1.0, + "content": "comprehensive fusion, resulting in efficient and effective collaboration.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 107, + 370, + 237, + 384 + ], + "lines": [ + { + "bbox": [ + 104, + 370, + 237, + 385 + ], + "spans": [ + { + "bbox": [ + 104, + 370, + 237, + 385 + ], + "score": 1.0, + "content": "3 Problem Formulation", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 388, + 506, + 432 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 146, + 401 + ], + "score": 1.0, + "content": "Consider", + "type": "text" + }, + { + "bbox": [ + 146, + 389, + 156, + 399 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 389, + 257, + 401 + ], + "score": 1.0, + "content": "agents in the scene. Let", + "type": "text" + }, + { + "bbox": [ + 257, + 389, + 269, + 400 + ], + "score": 0.88, + "content": "\\mathcal { X } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 389, + 287, + 401 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 288, + 389, + 299, + 400 + ], + "score": 0.87, + "content": "\\mathcal { \\mathrm { V } } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "be the observation and the perception supervision", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 399, + 507, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 134, + 412 + ], + "score": 1.0, + "content": "of the", + "type": "text" + }, + { + "bbox": [ + 134, + 401, + 139, + 410 + ], + "score": 0.33, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 399, + 507, + 412 + ], + "score": 1.0, + "content": "th agent, respectively. The objective of collaborative perception is to achieve the maxi-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 477, + 423 + ], + "score": 1.0, + "content": "mized perception performance of all agents as a function of the total communication budge", + "type": "text" + }, + { + "bbox": [ + 478, + 411, + 487, + 421 + ], + "score": 0.83, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 422, + 241, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 196, + 434 + ], + "score": 1.0, + "content": "communication round", + "type": "text" + }, + { + "bbox": [ + 197, + 422, + 207, + 432 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 422, + 241, + 434 + ], + "score": 1.0, + "content": "; that is,", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 431, + 477, + 466 + ], + "lines": [ + { + "bbox": [ + 133, + 431, + 477, + 466 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 477, + 466 + ], + "score": 0.92, + "content": "\\xi _ { \\Phi } ( B , K ) = \\underset { \\theta , \\mathcal { P } } { \\arg \\operatorname* { m a x } } \\sum _ { i = 1 } ^ { N } g ( \\Phi _ { \\theta } ( \\mathcal { X } _ { i } , \\{ \\mathcal { P } _ { i j } ^ { ( K ) } \\} _ { j = 1 } ^ { N } ) , \\mathcal { Y } _ { i } ) , \\mathrm { s . t . } \\sum _ { k = 1 } ^ { K } \\sum _ { i = 1 } ^ { N } | \\mathcal { P } _ { i j } ^ { ( k ) } | \\leq B ,", + "type": "interline_equation", + "image_path": "2c55ef85a09846703b2206d41fd6872b63b1419a1099b6ba3e9e0d0a169281b5.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 133, + 431, + 477, + 442.6666666666667 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 133, + 442.6666666666667, + 477, + 454.33333333333337 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 133, + 454.33333333333337, + 477, + 466.00000000000006 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 465, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 132, + 479 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 465, + 156, + 477 + ], + "score": 0.92, + "content": "g ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 464, + 297, + 479 + ], + "score": 1.0, + "content": "is the perception evaluation metric,", + "type": "text" + }, + { + "bbox": [ + 297, + 466, + 305, + 475 + ], + "score": 0.76, + "content": "\\Phi", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 464, + 506, + 479 + ], + "score": 1.0, + "content": "is the perception network with trainable parameter", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 103, + 475, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 103, + 475, + 158, + 493 + ], + "score": 1.0, + "content": "θ, and P (k)i→j", + "type": "text" + }, + { + "bbox": [ + 156, + 476, + 362, + 493 + ], + "score": 1.0, + "content": "is the message transmitted from the ith agent to the", + "type": "text" + }, + { + "bbox": [ + 362, + 480, + 368, + 491 + ], + "score": 0.73, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 476, + 425, + 493 + ], + "score": 1.0, + "content": "th agent at the", + "type": "text" + }, + { + "bbox": [ + 425, + 480, + 432, + 489 + ], + "score": 0.6, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 476, + 506, + 493 + ], + "score": 1.0, + "content": "th communication", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 213, + 504 + ], + "score": 1.0, + "content": "round. Note that i) when", + "type": "text" + }, + { + "bbox": [ + 213, + 492, + 267, + 502 + ], + "score": 0.91, + "content": "B = K = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 491, + 393, + 504 + ], + "score": 1.0, + "content": ", there is no collaboration and", + "type": "text" + }, + { + "bbox": [ + 393, + 492, + 426, + 504 + ], + "score": 0.94, + "content": "\\xi _ { \\Phi } ( 0 , 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "reflects the single-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "agent perception performance; ii) through optimizing the communication strategy and the network", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "parameter, collaborative perception should perform well consistently at any communication bandwidth", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "score": 1.0, + "content": "or round; and iii) we consider multi-round communication, where each agent serves as both a supporter", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 535, + 445, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 445, + 548 + ], + "score": 1.0, + "content": "(offering message to help others) and a requester (requesting messages from others).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 107, + 551, + 505, + 607 + ], + "lines": [ + { + "bbox": [ + 106, + 552, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 506, + 564 + ], + "score": 1.0, + "content": "In this work, we consider the perception task of 3D object detection and present three contributions:", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "i) we make communication more efficient by designing compact messages and sparse communication", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "graphs; ii) we boost the perception performance by implementing more comprehensive message", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "fusion; iii) we enable the overall system to adapt to varying communication conditions by dynamically", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 596, + 277, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 277, + 608 + ], + "score": 1.0, + "content": "adjusting where and who to communicate.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "title", + "bbox": [ + 104, + 615, + 497, + 630 + ], + "lines": [ + { + "bbox": [ + 103, + 614, + 500, + 633 + ], + "spans": [ + { + "bbox": [ + 103, + 614, + 500, + 633 + ], + "score": 1.0, + "content": "4 Where2comm: Spatial Confidence-Aware Collaborative Perception System", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 104, + 633, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 104, + 633, + 506, + 648 + ], + "score": 1.0, + "content": "This section presents Where2comm, a multi-round, multi-modality, multi-agent collaborative per-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "ception framework based on a spatial-confidence-aware communication strategy; see the overview", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "in Fig. 2. Where2comm includes an observation encoder, a spatial confidence generator, the spatial", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "confidence-aware communication module, the spatial confidence-aware message fusion module and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "a detection decoder. Among five modules, the proposed spatial confidence generator generates the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "spatial confidence map. Based on this spatial confidence map, the proposed spatial confidence-aware", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 701, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 506, + 712 + ], + "score": 1.0, + "content": "communication generates compact messages and sparse communication graphs to save commu-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "nication bandwidth; and the proposed spatial confidence-aware message fusion module leverages", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 89, + 497, + 147 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 149, + 78, + 460, + 89 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 148, + 76, + 462, + 92 + ], + "spans": [ + { + "bbox": [ + 148, + 76, + 462, + 92 + ], + "score": 1.0, + "content": "Table 1: Major components comparisons of collaborative perception systems.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 110, + 89, + 497, + 147 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 89, + 497, + 147 + ], + "spans": [ + { + "bbox": [ + 110, + 89, + 497, + 147 + ], + "score": 0.978, + "html": "
MethodVenueMessage packingCommunication graph constructionMessage fusion
When2com[12]CVPR2020Full feature mapHandshake-based sparse graphAttention per-agent
V2VNet[1]ECCV 2020Full feature mapFully connected graphAverage per-agent
DiscoNet [2]NeurIPS 2021Full feature mapFully connected graphMLP-based attention per-location
V2X-ViT[26]ECCV2022Full feature mapFully connected graphSelf-attention per-location
Where2commNeurIPS 2022Confidence-aware sparse feature map + request mapConfidence-aware sparse graphConfidence-awaremulti-head attention per-location
", + "type": "table", + "image_path": "c3c501a4a0c57a095e7724ab343c779a6d6cc5cbfd5d7d3759f74ca54ad62760.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 110, + 89, + 497, + 108.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 108.33333333333333, + 497, + 127.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 110, + 127.66666666666666, + 497, + 147.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "text", + "bbox": [ + 106, + 151, + 505, + 217 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 505, + 163 + ], + "score": 1.0, + "content": "CommNet [24] learns continuous communication in the multi-agent system. Vain [25] adopts the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 162, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 505, + 173 + ], + "score": 1.0, + "content": "attention mechanism to help agents selectively fuse the information from others. Most of these", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 173, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 506, + 185 + ], + "score": 1.0, + "content": "previous works consider decision-making tasks and adopt reinforcement learning due to the lack", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 196 + ], + "score": 1.0, + "content": "of explicit supervision. In this work, we focus on the perception task. Based on direct perception", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 195, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 206 + ], + "score": 1.0, + "content": "supervision, we apply supervised learning to optimize the communication strategy in both trade-off", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 206, + 282, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 282, + 218 + ], + "score": 1.0, + "content": "perception ability and communication cost.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 151, + 506, + 218 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 222, + 506, + 364 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 506, + 235 + ], + "score": 1.0, + "content": "Collaborative perception. As a recent application of multi-agent communication systems to per-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 233, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 506, + 245 + ], + "score": 1.0, + "content": "ception tasks, collaborative perception is still immature. To support this area of research, there is a", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 243, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 258 + ], + "score": 1.0, + "content": "surge of high-quality datasets (e.g., V2X-Sim [9], OpenV2V [10], Comap[27] and DAIR-V2X[11]),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 254, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 506, + 267 + ], + "score": 1.0, + "content": "as well as collaboration methods aimed for better performance-bandwidth trade-off (see comparisons", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 264, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 104, + 264, + 506, + 279 + ], + "score": 1.0, + "content": "in Table 1). When2com [12] proposes a handshake communication mechanism to decide when to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 506, + 291 + ], + "score": 1.0, + "content": "communicate and create sparse communication graph. V2VNet [1] proposes multi-round message", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 288, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 300 + ], + "score": 1.0, + "content": "passing based on graph neural networks to achieve better perception and prediction performance.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "DiscoNet [2] adopts knowledge distillation to take the advantage of both early and intermediate", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 308, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 323 + ], + "score": 1.0, + "content": "collaboration. OPV2V [10] proposes a graph-based attentive intermediate fusion to improve percep-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 320, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 506, + 333 + ], + "score": 1.0, + "content": "tion performances. V2X-ViT [26] introduces a novel heterogeneous multi-agent attention module", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "to fuse information across heterogeneous agents. In this work, we leverage the proposed spatial", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "score": 1.0, + "content": "confidence map to promote more compact messages, more sparse communication graphs, and more", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 353, + 392, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 392, + 365 + ], + "score": 1.0, + "content": "comprehensive fusion, resulting in efficient and effective collaboration.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 16, + "bbox_fs": [ + 104, + 221, + 506, + 365 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 370, + 237, + 384 + ], + "lines": [ + { + "bbox": [ + 104, + 370, + 237, + 385 + ], + "spans": [ + { + "bbox": [ + 104, + 370, + 237, + 385 + ], + "score": 1.0, + "content": "3 Problem Formulation", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 388, + 506, + 432 + ], + "lines": [ + { + "bbox": [ + 106, + 389, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 146, + 401 + ], + "score": 1.0, + "content": "Consider", + "type": "text" + }, + { + "bbox": [ + 146, + 389, + 156, + 399 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 389, + 257, + 401 + ], + "score": 1.0, + "content": "agents in the scene. Let", + "type": "text" + }, + { + "bbox": [ + 257, + 389, + 269, + 400 + ], + "score": 0.88, + "content": "\\mathcal { X } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 389, + 287, + 401 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 288, + 389, + 299, + 400 + ], + "score": 0.87, + "content": "\\mathcal { \\mathrm { V } } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 389, + 505, + 401 + ], + "score": 1.0, + "content": "be the observation and the perception supervision", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 399, + 507, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 134, + 412 + ], + "score": 1.0, + "content": "of the", + "type": "text" + }, + { + "bbox": [ + 134, + 401, + 139, + 410 + ], + "score": 0.33, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 399, + 507, + 412 + ], + "score": 1.0, + "content": "th agent, respectively. The objective of collaborative perception is to achieve the maxi-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 477, + 423 + ], + "score": 1.0, + "content": "mized perception performance of all agents as a function of the total communication budge", + "type": "text" + }, + { + "bbox": [ + 478, + 411, + 487, + 421 + ], + "score": 0.83, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 422, + 241, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 196, + 434 + ], + "score": 1.0, + "content": "communication round", + "type": "text" + }, + { + "bbox": [ + 197, + 422, + 207, + 432 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 422, + 241, + 434 + ], + "score": 1.0, + "content": "; that is,", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 389, + 507, + 434 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 133, + 431, + 477, + 466 + ], + "lines": [ + { + "bbox": [ + 133, + 431, + 477, + 466 + ], + "spans": [ + { + "bbox": [ + 133, + 431, + 477, + 466 + ], + "score": 0.92, + "content": "\\xi _ { \\Phi } ( B , K ) = \\underset { \\theta , \\mathcal { P } } { \\arg \\operatorname* { m a x } } \\sum _ { i = 1 } ^ { N } g ( \\Phi _ { \\theta } ( \\mathcal { X } _ { i } , \\{ \\mathcal { P } _ { i j } ^ { ( K ) } \\} _ { j = 1 } ^ { N } ) , \\mathcal { Y } _ { i } ) , \\mathrm { s . t . } \\sum _ { k = 1 } ^ { K } \\sum _ { i = 1 } ^ { N } | \\mathcal { P } _ { i j } ^ { ( k ) } | \\leq B ,", + "type": "interline_equation", + "image_path": "2c55ef85a09846703b2206d41fd6872b63b1419a1099b6ba3e9e0d0a169281b5.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 133, + 431, + 477, + 442.6666666666667 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 133, + 442.6666666666667, + 477, + 454.33333333333337 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 133, + 454.33333333333337, + 477, + 466.00000000000006 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 465, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 132, + 479 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 465, + 156, + 477 + ], + "score": 0.92, + "content": "g ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 464, + 297, + 479 + ], + "score": 1.0, + "content": "is the perception evaluation metric,", + "type": "text" + }, + { + "bbox": [ + 297, + 466, + 305, + 475 + ], + "score": 0.76, + "content": "\\Phi", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 464, + 506, + 479 + ], + "score": 1.0, + "content": "is the perception network with trainable parameter", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 103, + 475, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 103, + 475, + 158, + 493 + ], + "score": 1.0, + "content": "θ, and P (k)i→j", + "type": "text" + }, + { + "bbox": [ + 156, + 476, + 362, + 493 + ], + "score": 1.0, + "content": "is the message transmitted from the ith agent to the", + "type": "text" + }, + { + "bbox": [ + 362, + 480, + 368, + 491 + ], + "score": 0.73, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 476, + 425, + 493 + ], + "score": 1.0, + "content": "th agent at the", + "type": "text" + }, + { + "bbox": [ + 425, + 480, + 432, + 489 + ], + "score": 0.6, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 476, + 506, + 493 + ], + "score": 1.0, + "content": "th communication", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 213, + 504 + ], + "score": 1.0, + "content": "round. Note that i) when", + "type": "text" + }, + { + "bbox": [ + 213, + 492, + 267, + 502 + ], + "score": 0.91, + "content": "B = K = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 491, + 393, + 504 + ], + "score": 1.0, + "content": ", there is no collaboration and", + "type": "text" + }, + { + "bbox": [ + 393, + 492, + 426, + 504 + ], + "score": 0.94, + "content": "\\xi _ { \\Phi } ( 0 , 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "reflects the single-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 515 + ], + "score": 1.0, + "content": "agent perception performance; ii) through optimizing the communication strategy and the network", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "parameter, collaborative perception should perform well consistently at any communication bandwidth", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "score": 1.0, + "content": "or round; and iii) we consider multi-round communication, where each agent serves as both a supporter", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 535, + 445, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 535, + 445, + 548 + ], + "score": 1.0, + "content": "(offering message to help others) and a requester (requesting messages from others).", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 103, + 464, + 506, + 548 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 551, + 505, + 607 + ], + "lines": [ + { + "bbox": [ + 106, + 552, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 506, + 564 + ], + "score": 1.0, + "content": "In this work, we consider the perception task of 3D object detection and present three contributions:", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "i) we make communication more efficient by designing compact messages and sparse communication", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "graphs; ii) we boost the perception performance by implementing more comprehensive message", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "fusion; iii) we enable the overall system to adapt to varying communication conditions by dynamically", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 596, + 277, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 277, + 608 + ], + "score": 1.0, + "content": "adjusting where and who to communicate.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 552, + 506, + 608 + ] + }, + { + "type": "title", + "bbox": [ + 104, + 615, + 497, + 630 + ], + "lines": [ + { + "bbox": [ + 103, + 614, + 500, + 633 + ], + "spans": [ + { + "bbox": [ + 103, + 614, + 500, + 633 + ], + "score": 1.0, + "content": "4 Where2comm: Spatial Confidence-Aware Collaborative Perception System", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 106, + 634, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 104, + 633, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 104, + 633, + 506, + 648 + ], + "score": 1.0, + "content": "This section presents Where2comm, a multi-round, multi-modality, multi-agent collaborative per-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "ception framework based on a spatial-confidence-aware communication strategy; see the overview", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "in Fig. 2. Where2comm includes an observation encoder, a spatial confidence generator, the spatial", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "confidence-aware communication module, the spatial confidence-aware message fusion module and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "a detection decoder. Among five modules, the proposed spatial confidence generator generates the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "spatial confidence map. Based on this spatial confidence map, the proposed spatial confidence-aware", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 701, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 506, + 712 + ], + "score": 1.0, + "content": "communication generates compact messages and sparse communication graphs to save commu-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "nication bandwidth; and the proposed spatial confidence-aware message fusion module leverages", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 253, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 267 + ], + "score": 1.0, + "content": "informative spatial confidence priors to achieve better aggregation; also see an algorithmic summary", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 264, + 466, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 466, + 277 + ], + "score": 1.0, + "content": "in Algorithm 1 and the optimization-oriented design rationale in Section 7.3 in Appendix.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 47.5, + "bbox_fs": [ + 104, + 633, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 71, + 504, + 205 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 71, + 504, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 71, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 504, + 205 + ], + "score": 0.975, + "type": "image", + "image_path": "63724c901bc4ff631e74d37f129418530be55f4216a76533593a817347b39d8a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 71, + 504, + 115.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 115.66666666666666, + 504, + 160.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 160.33333333333331, + 504, + 204.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 212, + 506, + 246 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 212, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 506, + 226 + ], + "score": 1.0, + "content": "Figure 2: System overview. In Where2comm, spatial confidence generator enables the awareness of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "spatial heterogeneous of perceptual information, spatial confidence-aware communication enables", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 234, + 488, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 488, + 247 + ], + "score": 1.0, + "content": "efficient communication, and spatial confidence-aware message fusion boosts the performance.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 254, + 504, + 276 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 267 + ], + "score": 1.0, + "content": "informative spatial confidence priors to achieve better aggregation; also see an algorithmic summary", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 264, + 466, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 466, + 277 + ], + "score": 1.0, + "content": "in Algorithm 1 and the optimization-oriented design rationale in Section 7.3 in Appendix.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 107, + 284, + 219, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 220, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 220, + 297 + ], + "score": 1.0, + "content": "4.1 Observation encoder", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "The observation encoder extracts feature maps from the sensor data. Where2comm accepts single/multi-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "modality inputs, such as RGB images and 3D point clouds. This work adopts the feature represen-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 319, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 333 + ], + "score": 1.0, + "content": "tations in bird’s eye view (BEV), where all agents project their individual perceptual information", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 330, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 104, + 330, + 506, + 345 + ], + "score": 1.0, + "content": "to the same global coordinate system, avoiding complex coordinate transformations and supporting", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 342, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 411, + 355 + ], + "score": 1.0, + "content": "better shared cross-agent collaboration. For the ith agent, given its input", + "type": "text" + }, + { + "bbox": [ + 411, + 343, + 423, + 353 + ], + "score": 0.87, + "content": "\\mathcal { X } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 342, + 506, + 355 + ], + "score": 1.0, + "content": ", the feature map is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 351, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 238, + 368 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { i } ^ { ( 0 ) } = \\Phi _ { \\mathrm { e n c } } ( \\mathcal { X } _ { i } ) \\in \\mathrm { \\overline { { \\mathbb { R } } } } ^ { H \\times W \\times D }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 351, + 270, + 372 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 271, + 355, + 302, + 367 + ], + "score": 0.91, + "content": "\\Phi _ { \\mathrm { e n c } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 351, + 506, + 372 + ], + "score": 1.0, + "content": "is the encoder, the superscript 0 reflects that the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 288, + 380 + ], + "score": 1.0, + "content": "feature is obtained before communication and", + "type": "text" + }, + { + "bbox": [ + 289, + 367, + 325, + 378 + ], + "score": 0.87, + "content": "H , W , D", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "are its height, weight and channel. All agents", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 375, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 104, + 375, + 365, + 392 + ], + "score": 1.0, + "content": "share the same BEV coordinate system. For the image input,", + "type": "text" + }, + { + "bbox": [ + 366, + 378, + 396, + 389 + ], + "score": 0.92, + "content": "\\Phi _ { \\mathrm { e n c } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 375, + 506, + 392 + ], + "score": 1.0, + "content": "is followed by a warping", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 505, + 401 + ], + "score": 1.0, + "content": "function that transforms the extracted feature from front-view to BEV. For 3D point cloud input, we", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 398, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 265, + 413 + ], + "score": 1.0, + "content": "discretize 3D points as a BEV map and", + "type": "text" + }, + { + "bbox": [ + 266, + 399, + 296, + 411 + ], + "score": 0.92, + "content": "\\Phi _ { \\mathrm { e n c } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 398, + 506, + 413 + ], + "score": 1.0, + "content": "extracts features in BEV. The extracted feature map", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 410, + 413, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 413, + 423 + ], + "score": 1.0, + "content": "is output to the spatial confidence generator and the message fusion module.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 107, + 429, + 252, + 441 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 253, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 253, + 444 + ], + "score": 1.0, + "content": "4.2 Spatial confidence generator", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 444, + 505, + 521 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "The spatial confidence generator generates a spatial confidence map from the feature map of each", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 455, + 507, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 507, + 468 + ], + "score": 1.0, + "content": "agent. The spatial confidence map reflects the perceptually critical level of various spatial areas.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "Intuitively, for object detection task, the areas that contain objects are more critical than background", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "areas. During collaboration, areas with objects could help recover the miss-detected objects due to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 486, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 501 + ], + "score": 1.0, + "content": "the limited view; and background areas could be omitted to save the precious bandwidth. So we", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "represent the spatial confidence map with the detection confidence map, where the area with high", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 510, + 464, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 464, + 522 + ], + "score": 1.0, + "content": "perceptually critical level is the area that contains an object with a high confidence score.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 503, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "To implement, we use a detection decoder structure to produce the detection confidence map. Given", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 534, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 104, + 534, + 194, + 554 + ], + "score": 1.0, + "content": "the feature map at the", + "type": "text" + }, + { + "bbox": [ + 195, + 540, + 201, + 549 + ], + "score": 0.75, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 534, + 302, + 554 + ], + "score": 1.0, + "content": "th communication round,", + "type": "text" + }, + { + "bbox": [ + 303, + 536, + 322, + 551 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 534, + 505, + 554 + ], + "score": 1.0, + "content": ", the corresponding spatial confidence map is", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 554, + 385, + 571 + ], + "lines": [ + { + "bbox": [ + 226, + 554, + 385, + 571 + ], + "spans": [ + { + "bbox": [ + 226, + 554, + 385, + 571 + ], + "score": 0.93, + "content": "\\mathbf { C } _ { i } ^ { ( k ) } = \\Phi _ { \\mathrm { g e n e r a t o r } } ( \\mathcal { F } _ { i } ^ { ( k ) } ) \\in [ 0 , 1 ] ^ { H \\times W } ,", + "type": "interline_equation", + "image_path": "d258ec6d4eafb313c1161b996f914e60814f979b2b5b00b4667f9f5db7684d14.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 226, + 554, + 385, + 571 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 505, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 194, + 590 + ], + "score": 1.0, + "content": "where the generator", + "type": "text" + }, + { + "bbox": [ + 194, + 576, + 245, + 588 + ], + "score": 0.92, + "content": "\\Phi _ { \\mathrm { g e n e r a t o r } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 575, + 506, + 590 + ], + "score": 1.0, + "content": "follows a detection decoder. Since we consider multi-round", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "score": 1.0, + "content": "collaboration, Where2comm iteratively updates the feature map by aggregating information from", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 103, + 595, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 103, + 595, + 185, + 615 + ], + "score": 1.0, + "content": "other agents. Once", + "type": "text" + }, + { + "bbox": [ + 185, + 598, + 205, + 613 + ], + "score": 0.93, + "content": "\\mathcal { F } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 595, + 507, + 615 + ], + "score": 1.0, + "content": "is obtained, (1) is triggered to reflect the perceptually critical level at each", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "spatial location. The proposed spatial confidence map answers a crucial question that was ignored by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 622, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 636 + ], + "score": 1.0, + "content": "previous works: for each agent, information at which spatial area is worth sharing with others. By", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 633, + 496, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 496, + 646 + ], + "score": 1.0, + "content": "answering this, it provides a solid base for efficient communication and effective message fusion.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "title", + "bbox": [ + 106, + 653, + 306, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 651, + 307, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 307, + 666 + ], + "score": 1.0, + "content": "4.3 Spatial confidence-aware communication", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 680 + ], + "score": 1.0, + "content": "With the guidance of spatial confidence maps, the proposed communication module packs compact", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "messages with spatially sparse feature maps and transmits messages through a sparsely-connected", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "communication graph. Most existing collaboration perception systems [1, 2, 26] considers full feature", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "maps in the messages and fully-connected communication graphs. To reduce the communication", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 711, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 724 + ], + "score": 1.0, + "content": "bandwidth without affecting perception, we leverage the spatial confidence map to select the most", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + } + ], + "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": "image", + "bbox": [ + 108, + 71, + 504, + 205 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 71, + 504, + 205 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 71, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 504, + 205 + ], + "score": 0.975, + "type": "image", + "image_path": "63724c901bc4ff631e74d37f129418530be55f4216a76533593a817347b39d8a.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 71, + 504, + 115.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 115.66666666666666, + 504, + 160.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 160.33333333333331, + 504, + 204.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 212, + 506, + 246 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 212, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 506, + 226 + ], + "score": 1.0, + "content": "Figure 2: System overview. In Where2comm, spatial confidence generator enables the awareness of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 506, + 236 + ], + "score": 1.0, + "content": "spatial heterogeneous of perceptual information, spatial confidence-aware communication enables", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 234, + 488, + 247 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 488, + 247 + ], + "score": 1.0, + "content": "efficient communication, and spatial confidence-aware message fusion boosts the performance.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 254, + 504, + 276 + ], + "lines": [], + "index": 6.5, + "bbox_fs": [ + 105, + 253, + 505, + 277 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 284, + 219, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 283, + 220, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 283, + 220, + 297 + ], + "score": 1.0, + "content": "4.1 Observation encoder", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 298, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 311 + ], + "score": 1.0, + "content": "The observation encoder extracts feature maps from the sensor data. Where2comm accepts single/multi-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "modality inputs, such as RGB images and 3D point clouds. This work adopts the feature represen-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 319, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 333 + ], + "score": 1.0, + "content": "tations in bird’s eye view (BEV), where all agents project their individual perceptual information", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 330, + 506, + 345 + ], + "spans": [ + { + "bbox": [ + 104, + 330, + 506, + 345 + ], + "score": 1.0, + "content": "to the same global coordinate system, avoiding complex coordinate transformations and supporting", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 342, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 411, + 355 + ], + "score": 1.0, + "content": "better shared cross-agent collaboration. For the ith agent, given its input", + "type": "text" + }, + { + "bbox": [ + 411, + 343, + 423, + 353 + ], + "score": 0.87, + "content": "\\mathcal { X } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 342, + 506, + 355 + ], + "score": 1.0, + "content": ", the feature map is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 351, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 238, + 368 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { i } ^ { ( 0 ) } = \\Phi _ { \\mathrm { e n c } } ( \\mathcal { X } _ { i } ) \\in \\mathrm { \\overline { { \\mathbb { R } } } } ^ { H \\times W \\times D }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 351, + 270, + 372 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 271, + 355, + 302, + 367 + ], + "score": 0.91, + "content": "\\Phi _ { \\mathrm { e n c } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 351, + 506, + 372 + ], + "score": 1.0, + "content": "is the encoder, the superscript 0 reflects that the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 366, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 366, + 288, + 380 + ], + "score": 1.0, + "content": "feature is obtained before communication and", + "type": "text" + }, + { + "bbox": [ + 289, + 367, + 325, + 378 + ], + "score": 0.87, + "content": "H , W , D", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 366, + 505, + 380 + ], + "score": 1.0, + "content": "are its height, weight and channel. All agents", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 375, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 104, + 375, + 365, + 392 + ], + "score": 1.0, + "content": "share the same BEV coordinate system. For the image input,", + "type": "text" + }, + { + "bbox": [ + 366, + 378, + 396, + 389 + ], + "score": 0.92, + "content": "\\Phi _ { \\mathrm { e n c } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 375, + 506, + 392 + ], + "score": 1.0, + "content": "is followed by a warping", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 388, + 505, + 401 + ], + "score": 1.0, + "content": "function that transforms the extracted feature from front-view to BEV. For 3D point cloud input, we", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 398, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 265, + 413 + ], + "score": 1.0, + "content": "discretize 3D points as a BEV map and", + "type": "text" + }, + { + "bbox": [ + 266, + 399, + 296, + 411 + ], + "score": 0.92, + "content": "\\Phi _ { \\mathrm { e n c } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 398, + 506, + 413 + ], + "score": 1.0, + "content": "extracts features in BEV. The extracted feature map", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 410, + 413, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 413, + 423 + ], + "score": 1.0, + "content": "is output to the spatial confidence generator and the message fusion module.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 298, + 506, + 423 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 429, + 252, + 441 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 253, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 253, + 444 + ], + "score": 1.0, + "content": "4.2 Spatial confidence generator", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 444, + 505, + 521 + ], + "lines": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "The spatial confidence generator generates a spatial confidence map from the feature map of each", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 455, + 507, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 507, + 468 + ], + "score": 1.0, + "content": "agent. The spatial confidence map reflects the perceptually critical level of various spatial areas.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "Intuitively, for object detection task, the areas that contain objects are more critical than background", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "areas. During collaboration, areas with objects could help recover the miss-detected objects due to", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 486, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 501 + ], + "score": 1.0, + "content": "the limited view; and background areas could be omitted to save the precious bandwidth. So we", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 499, + 505, + 511 + ], + "score": 1.0, + "content": "represent the spatial confidence map with the detection confidence map, where the area with high", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 510, + 464, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 464, + 522 + ], + "score": 1.0, + "content": "perceptually critical level is the area that contains an object with a high confidence score.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 444, + 507, + 522 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 525, + 503, + 551 + ], + "lines": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "To implement, we use a detection decoder structure to produce the detection confidence map. Given", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 534, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 104, + 534, + 194, + 554 + ], + "score": 1.0, + "content": "the feature map at the", + "type": "text" + }, + { + "bbox": [ + 195, + 540, + 201, + 549 + ], + "score": 0.75, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 534, + 302, + 554 + ], + "score": 1.0, + "content": "th communication round,", + "type": "text" + }, + { + "bbox": [ + 303, + 536, + 322, + 551 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 534, + 505, + 554 + ], + "score": 1.0, + "content": ", the corresponding spatial confidence map is", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 104, + 525, + 505, + 554 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 226, + 554, + 385, + 571 + ], + "lines": [ + { + "bbox": [ + 226, + 554, + 385, + 571 + ], + "spans": [ + { + "bbox": [ + 226, + 554, + 385, + 571 + ], + "score": 0.93, + "content": "\\mathbf { C } _ { i } ^ { ( k ) } = \\Phi _ { \\mathrm { g e n e r a t o r } } ( \\mathcal { F } _ { i } ^ { ( k ) } ) \\in [ 0 , 1 ] ^ { H \\times W } ,", + "type": "interline_equation", + "image_path": "d258ec6d4eafb313c1161b996f914e60814f979b2b5b00b4667f9f5db7684d14.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 226, + 554, + 385, + 571 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 576, + 505, + 645 + ], + "lines": [ + { + "bbox": [ + 105, + 575, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 575, + 194, + 590 + ], + "score": 1.0, + "content": "where the generator", + "type": "text" + }, + { + "bbox": [ + 194, + 576, + 245, + 588 + ], + "score": 0.92, + "content": "\\Phi _ { \\mathrm { g e n e r a t o r } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 575, + 506, + 590 + ], + "score": 1.0, + "content": "follows a detection decoder. Since we consider multi-round", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 505, + 600 + ], + "score": 1.0, + "content": "collaboration, Where2comm iteratively updates the feature map by aggregating information from", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 103, + 595, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 103, + 595, + 185, + 615 + ], + "score": 1.0, + "content": "other agents. Once", + "type": "text" + }, + { + "bbox": [ + 185, + 598, + 205, + 613 + ], + "score": 0.93, + "content": "\\mathcal { F } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 595, + 507, + 615 + ], + "score": 1.0, + "content": "is obtained, (1) is triggered to reflect the perceptually critical level at each", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 505, + 624 + ], + "score": 1.0, + "content": "spatial location. The proposed spatial confidence map answers a crucial question that was ignored by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 622, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 636 + ], + "score": 1.0, + "content": "previous works: for each agent, information at which spatial area is worth sharing with others. By", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 633, + 496, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 496, + 646 + ], + "score": 1.0, + "content": "answering this, it provides a solid base for efficient communication and effective message fusion.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 103, + 575, + 507, + 646 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 653, + 306, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 651, + 307, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 307, + 666 + ], + "score": 1.0, + "content": "4.3 Spatial confidence-aware communication", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 680 + ], + "score": 1.0, + "content": "With the guidance of spatial confidence maps, the proposed communication module packs compact", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 691 + ], + "score": 1.0, + "content": "messages with spatially sparse feature maps and transmits messages through a sparsely-connected", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "communication graph. Most existing collaboration perception systems [1, 2, 26] considers full feature", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "maps in the messages and fully-connected communication graphs. To reduce the communication", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 711, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 506, + 724 + ], + "score": 1.0, + "content": "bandwidth without affecting perception, we leverage the spatial confidence map to select the most", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "informative spatial areas in the feature map (where to communicate) and decide the most beneficial", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 84, + 291, + 96 + ], + "spans": [ + { + "bbox": [ + 107, + 84, + 291, + 96 + ], + "score": 1.0, + "content": "collaboration partners (who to communicate).", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 666, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "informative spatial areas in the feature map (where to communicate) and decide the most beneficial", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 84, + 291, + 96 + ], + "spans": [ + { + "bbox": [ + 107, + 84, + 291, + 96 + ], + "score": 1.0, + "content": "collaboration partners (who to communicate).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 506, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 506, + 112 + ], + "score": 1.0, + "content": "Message packing. Message packing determines what information should be included in the to-be-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "sent message. The proposed message includes: i) a request map that indicates at which spatial areas", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 121, + 481, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 481, + 135 + ], + "score": 1.0, + "content": "the agent needs to know more; and ii) a spatially sparse, yet perceptually critical feature map.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 106, + 137, + 505, + 196 + ], + "lines": [ + { + "bbox": [ + 103, + 136, + 508, + 155 + ], + "spans": [ + { + "bbox": [ + 103, + 136, + 244, + 155 + ], + "score": 1.0, + "content": "The request map of the ith agent is", + "type": "text" + }, + { + "bbox": [ + 244, + 138, + 355, + 153 + ], + "score": 0.93, + "content": "\\mathbf { R } _ { i } ^ { ( k ) } = 1 - \\mathbf { C } _ { i } ^ { ( k ) } \\in \\mathbb { R } ^ { H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 136, + 508, + 155 + ], + "score": 1.0, + "content": ", negatively correlated with the spatial", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 151, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 506, + 163 + ], + "score": 1.0, + "content": "confidence map. The intuition is, for the locations with low confidence score, an agent is hard to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 162, + 507, + 175 + ], + "spans": [ + { + "bbox": [ + 104, + 162, + 507, + 175 + ], + "score": 1.0, + "content": "tell if there is really no objects or it is just caused by the limited information (e.g. occlusion). Thus,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 172, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 187 + ], + "score": 1.0, + "content": "the low confidence score indicates there could be missing information at that location. Requesting", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 183, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 507, + 198 + ], + "score": 1.0, + "content": "information at these locations from other agents could improve the current agent’s detection accuracy.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 200, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 200, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 214 + ], + "score": 1.0, + "content": "The spatially sparse feature map are selected based on each agent’s spatial confidence map and the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "score": 1.0, + "content": "received request maps from others. Specifically, a binary selection matrix is used to represent each", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "location is selected or not, where 1 denotes selected, and 0 elsewhere. For the message sent from the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 234, + 459, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 169, + 246 + ], + "score": 1.0, + "content": "ith agent to the", + "type": "text" + }, + { + "bbox": [ + 169, + 235, + 174, + 245 + ], + "score": 0.76, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 234, + 232, + 246 + ], + "score": 1.0, + "content": "th agent at the", + "type": "text" + }, + { + "bbox": [ + 232, + 234, + 239, + 244 + ], + "score": 0.67, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 234, + 459, + 246 + ], + "score": 1.0, + "content": "th communication round, the binary selection matrix is", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 248, + 433, + 279 + ], + "lines": [ + { + "bbox": [ + 172, + 248, + 433, + 279 + ], + "spans": [ + { + "bbox": [ + 172, + 248, + 433, + 279 + ], + "score": 0.89, + "content": "\\mathbf { M } _ { i j } ^ { ( k ) } = \\{ \\begin{array} { l l } { \\Phi _ { \\mathrm { s e l e c t } } ( \\mathbf { C } _ { i } ^ { ( k ) } ) \\in \\{ 0 , 1 \\} ^ { H \\times W } , } & { \\quad k = 0 ; } \\\\ { \\Phi _ { \\mathrm { s e l e c t } } ( \\mathbf { C } _ { i } ^ { ( k ) } \\odot \\mathbf { R } _ { j } ^ { ( k - 1 ) } ) , \\in \\{ 0 , 1 \\} ^ { H \\times W } , } & { \\quad k > 0 ; } \\end{array} ", + "type": "interline_equation", + "image_path": "e6067067c0ec5ad2999b66753b68061121bf83326853471454ad14dcf9272144.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 172, + 248, + 433, + 258.3333333333333 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 172, + 258.3333333333333, + 433, + 268.66666666666663 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 172, + 268.66666666666663, + 433, + 278.99999999999994 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 278, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 104, + 275, + 508, + 294 + ], + "spans": [ + { + "bbox": [ + 104, + 278, + 132, + 292 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 280, + 142, + 289 + ], + "score": 0.83, + "content": "\\odot", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 278, + 280, + 292 + ], + "score": 1.0, + "content": "is the element-wise multiplication,", + "type": "text" + }, + { + "bbox": [ + 283, + 275, + 508, + 294 + ], + "score": 1.0, + "content": "R(k−1)j is the request map from the jth agent received at", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 291, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 186, + 304 + ], + "score": 1.0, + "content": "the previous round,", + "type": "text" + }, + { + "bbox": [ + 186, + 291, + 225, + 304 + ], + "score": 0.92, + "content": "\\Phi _ { \\mathrm { s e l e c t } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 291, + 506, + 304 + ], + "score": 1.0, + "content": "is the selection function which targets to select the most critical areas", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "conditioned on the input matrix, which represents the critical level at the certain spatial location. We", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 151, + 327 + ], + "score": 1.0, + "content": "implement", + "type": "text" + }, + { + "bbox": [ + 151, + 314, + 189, + 325 + ], + "score": 0.93, + "content": "\\Phi _ { \\mathrm { s e l e c t } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "by selecting the locations where the largest elements at in the given input matrix", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 324, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 506, + 337 + ], + "score": 1.0, + "content": "conditioned on the bandwidth limit; optionally, a Gaussian filter could be applied to filter out the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 506, + 347 + ], + "score": 1.0, + "content": "outliers and introduce some context. In the initial communication round, each agent selects the most", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "score": 1.0, + "content": "critical areas from its own perspective as the request maps from other agents are not available yet; in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 99, + 358, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 99, + 358, + 353, + 393 + ], + "score": 1.0, + "content": "communication. Then, the selected feature map is obtained as", + "type": "text" + }, + { + "bbox": [ + 353, + 368, + 501, + 384 + ], + "score": 0.93, + "content": "\\mathcal { Z } _ { i j } ^ { ( k ) } = \\mathbf { M } _ { i j } ^ { ( k ) } \\odot \\mathcal { F } _ { i } ^ { ( k ) } \\in \\mathbb { R } ^ { H \\times \\bar { W } \\times D }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 358, + 506, + 393 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 257, + 410 + ], + "score": 1.0, + "content": "Overall, the message sent from the", + "type": "text" + }, + { + "bbox": [ + 257, + 399, + 262, + 408 + ], + "score": 0.52, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 397, + 325, + 410 + ], + "score": 1.0, + "content": "th agent to the", + "type": "text" + }, + { + "bbox": [ + 325, + 399, + 331, + 410 + ], + "score": 0.69, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 397, + 394, + 410 + ], + "score": 1.0, + "content": "th agent at the", + "type": "text" + }, + { + "bbox": [ + 395, + 398, + 401, + 408 + ], + "score": 0.64, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "th communication round", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 101, + 404, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 101, + 404, + 117, + 430 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 118, + 408, + 212, + 425 + ], + "score": 0.93, + "content": "\\mathcal { P } _ { i j } ^ { ( k ) } ~ = ~ ( \\mathbf { R } _ { i } ^ { ( k ) } , \\mathcal { \\bar { Z } } _ { i j } ^ { ( k ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 407, + 273, + 426 + ], + "score": 1.0, + "content": "Note that i)", + "type": "text" + }, + { + "bbox": [ + 274, + 409, + 295, + 424 + ], + "score": 0.92, + "content": "\\mathbf { R } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 411, + 505, + 425 + ], + "score": 1.0, + "content": "provides spatial priors to request complementary", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 102, + 424, + 508, + 445 + ], + "spans": [ + { + "bbox": [ + 102, + 424, + 397, + 445 + ], + "score": 1.0, + "content": "information for the ith agent’s need in the next round; the feature map", + "type": "text" + }, + { + "bbox": [ + 398, + 424, + 420, + 441 + ], + "score": 0.92, + "content": "\\mathcal { Z } _ { i j } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 424, + 508, + 445 + ], + "score": 1.0, + "content": "provides supportive", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "information for the ith agent’s need in the this round. They together enable mutually beneficial", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 100, + 450, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 100, + 450, + 197, + 488 + ], + "score": 1.0, + "content": "collaboration; ii) since leading to low commu", + "type": "text" + }, + { + "bbox": [ + 221, + 450, + 335, + 488 + ], + "score": 1.0, + "content": "is sparse, we only transmit nn cost; and iii) the sparsity of", + "type": "text" + }, + { + "bbox": [ + 335, + 464, + 358, + 480 + ], + "score": 0.93, + "content": "\\mathcal { Z } _ { i j } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 450, + 507, + 488 + ], + "score": 1.0, + "content": "o features and corresponding indices,is determined by the binary selection", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 478, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 490 + ], + "score": 1.0, + "content": "matrix, which dynamically allocates the communication budget at various spatial areas based on their", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 489, + 394, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 394, + 501 + ], + "score": 1.0, + "content": "perceptual critical level, adapting to various communication conditions.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 504, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 518 + ], + "score": 1.0, + "content": "Communication graph construction. Communication graph construction targets to identify when", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "score": 1.0, + "content": "and who to communicate to avoid unnecessary communication that wastes the bandwidth. Most", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "previous works [1, 2, 10] consider fully-connected communication graphs. When2com [12] proposes", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "score": 1.0, + "content": "a handshake mechanism, which uses similar global features to match partners. This is hard to interpret", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 549, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 562 + ], + "score": 1.0, + "content": "because two agents, which have similar global features, do not necessarily need information from", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "each other. Different from all previous works, we provide an explicit design rationale: the necessity", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 275, + 583 + ], + "score": 1.0, + "content": "of communication between the ith and the", + "type": "text" + }, + { + "bbox": [ + 275, + 572, + 281, + 582 + ], + "score": 0.73, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "th agents is simply measured by the overlap between the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 355, + 594 + ], + "score": 1.0, + "content": "information that the ith agent has and the information that the", + "type": "text" + }, + { + "bbox": [ + 356, + 583, + 361, + 593 + ], + "score": 0.63, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 582, + 505, + 594 + ], + "score": 1.0, + "content": "th agent needs. With the help of the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 592, + 507, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 507, + 607 + ], + "score": 1.0, + "content": "spatial confidence map and the request map, we construct a more interpretable communication graph.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 608, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "score": 1.0, + "content": "For the initial communication round, every agent in the system is not aware of other agents yet.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "score": 1.0, + "content": "To activate the collaboration, we construct a fully-connected communication graph. Every agent", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 631, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 642 + ], + "score": 1.0, + "content": "will broadcast its message to the rest of the system. For the subsequent communication rounds, we", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 286, + 654 + ], + "score": 1.0, + "content": "examine if the communication between agent", + "type": "text" + }, + { + "bbox": [ + 286, + 642, + 291, + 651 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 641, + 332, + 654 + ], + "score": 1.0, + "content": "and agent", + "type": "text" + }, + { + "bbox": [ + 332, + 642, + 338, + 653 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "is necessary based on the maximum value", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 103, + 651, + 508, + 672 + ], + "spans": [ + { + "bbox": [ + 103, + 651, + 253, + 672 + ], + "score": 1.0, + "content": "of the binary selection matrix M(k)i→j ,", + "type": "text" + }, + { + "bbox": [ + 252, + 652, + 508, + 670 + ], + "score": 1.0, + "content": "i.e. if there is at least one patch is activated, then we regard the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 668, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 259, + 682 + ], + "score": 1.0, + "content": "connection is necessary. Formally, let", + "type": "text" + }, + { + "bbox": [ + 260, + 668, + 280, + 679 + ], + "score": 0.87, + "content": "\\mathbf { A } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 668, + 506, + 682 + ], + "score": 1.0, + "content": "be the adjacency matrix of the communication graph at", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 680, + 331, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 121, + 692 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 681, + 127, + 690 + ], + "score": 0.45, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 680, + 258, + 692 + ], + "score": 1.0, + "content": "th communication round, whose", + "type": "text" + }, + { + "bbox": [ + 258, + 680, + 279, + 692 + ], + "score": 0.91, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 680, + 331, + 692 + ], + "score": 1.0, + "content": "th element is", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44 + }, + { + "type": "interline_equation", + "bbox": [ + 138, + 692, + 466, + 726 + ], + "lines": [ + { + "bbox": [ + 138, + 692, + 466, + 726 + ], + "spans": [ + { + "bbox": [ + 138, + 692, + 466, + 726 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\mathbf { A } _ { i , j } ^ { ( k ) } = \\left\\{ \\begin{array} { l l } { 1 , \\quad } & { k = 0 ; } \\\\ { \\operatorname* { m a x } _ { h \\in \\{ 0 , 1 , \\dots , H - 1 \\} , w \\in \\{ 0 , 1 , \\dots , W - 1 \\} } \\left( \\mathbf { M } _ { i \\to j } ^ { ( k ) } \\right) _ { h , w } \\in \\{ 0 , 1 \\} , \\quad } & { k > 0 ; } \\end{array} \\right. } \\end{array}", + "type": "interline_equation", + "image_path": "9b9be8a18f260845fcfdd7d7078d1ce50a9932aa287637e5e4a0af8fb40291ef.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 138, + 692, + 466, + 703.3333333333334 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 138, + 703.3333333333334, + 466, + 714.6666666666667 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 138, + 714.6666666666667, + 466, + 726.0000000000001 + ], + "spans": [], + "index": 50 + } + ] + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 504, + 95 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 72, + 505, + 96 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 506, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 506, + 112 + ], + "score": 1.0, + "content": "Message packing. Message packing determines what information should be included in the to-be-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "sent message. The proposed message includes: i) a request map that indicates at which spatial areas", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 121, + 481, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 481, + 135 + ], + "score": 1.0, + "content": "the agent needs to know more; and ii) a spatially sparse, yet perceptually critical feature map.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 100, + 506, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 137, + 505, + 196 + ], + "lines": [ + { + "bbox": [ + 103, + 136, + 508, + 155 + ], + "spans": [ + { + "bbox": [ + 103, + 136, + 244, + 155 + ], + "score": 1.0, + "content": "The request map of the ith agent is", + "type": "text" + }, + { + "bbox": [ + 244, + 138, + 355, + 153 + ], + "score": 0.93, + "content": "\\mathbf { R } _ { i } ^ { ( k ) } = 1 - \\mathbf { C } _ { i } ^ { ( k ) } \\in \\mathbb { R } ^ { H \\times W }", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 136, + 508, + 155 + ], + "score": 1.0, + "content": ", negatively correlated with the spatial", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 151, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 506, + 163 + ], + "score": 1.0, + "content": "confidence map. The intuition is, for the locations with low confidence score, an agent is hard to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 162, + 507, + 175 + ], + "spans": [ + { + "bbox": [ + 104, + 162, + 507, + 175 + ], + "score": 1.0, + "content": "tell if there is really no objects or it is just caused by the limited information (e.g. occlusion). Thus,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 172, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 187 + ], + "score": 1.0, + "content": "the low confidence score indicates there could be missing information at that location. Requesting", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 183, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 507, + 198 + ], + "score": 1.0, + "content": "information at these locations from other agents could improve the current agent’s detection accuracy.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 103, + 136, + 508, + 198 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 200, + 505, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 200, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 505, + 214 + ], + "score": 1.0, + "content": "The spatially sparse feature map are selected based on each agent’s spatial confidence map and the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "score": 1.0, + "content": "received request maps from others. Specifically, a binary selection matrix is used to represent each", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 235 + ], + "score": 1.0, + "content": "location is selected or not, where 1 denotes selected, and 0 elsewhere. For the message sent from the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 234, + 459, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 169, + 246 + ], + "score": 1.0, + "content": "ith agent to the", + "type": "text" + }, + { + "bbox": [ + 169, + 235, + 174, + 245 + ], + "score": 0.76, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 234, + 232, + 246 + ], + "score": 1.0, + "content": "th agent at the", + "type": "text" + }, + { + "bbox": [ + 232, + 234, + 239, + 244 + ], + "score": 0.67, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 234, + 459, + 246 + ], + "score": 1.0, + "content": "th communication round, the binary selection matrix is", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 200, + 505, + 246 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 172, + 248, + 433, + 279 + ], + "lines": [ + { + "bbox": [ + 172, + 248, + 433, + 279 + ], + "spans": [ + { + "bbox": [ + 172, + 248, + 433, + 279 + ], + "score": 0.89, + "content": "\\mathbf { M } _ { i j } ^ { ( k ) } = \\{ \\begin{array} { l l } { \\Phi _ { \\mathrm { s e l e c t } } ( \\mathbf { C } _ { i } ^ { ( k ) } ) \\in \\{ 0 , 1 \\} ^ { H \\times W } , } & { \\quad k = 0 ; } \\\\ { \\Phi _ { \\mathrm { s e l e c t } } ( \\mathbf { C } _ { i } ^ { ( k ) } \\odot \\mathbf { R } _ { j } ^ { ( k - 1 ) } ) , \\in \\{ 0 , 1 \\} ^ { H \\times W } , } & { \\quad k > 0 ; } \\end{array} ", + "type": "interline_equation", + "image_path": "e6067067c0ec5ad2999b66753b68061121bf83326853471454ad14dcf9272144.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 172, + 248, + 433, + 258.3333333333333 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 172, + 258.3333333333333, + 433, + 268.66666666666663 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 172, + 268.66666666666663, + 433, + 278.99999999999994 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 278, + 505, + 394 + ], + "lines": [ + { + "bbox": [ + 104, + 275, + 508, + 294 + ], + "spans": [ + { + "bbox": [ + 104, + 278, + 132, + 292 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 132, + 280, + 142, + 289 + ], + "score": 0.83, + "content": "\\odot", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 278, + 280, + 292 + ], + "score": 1.0, + "content": "is the element-wise multiplication,", + "type": "text" + }, + { + "bbox": [ + 283, + 275, + 508, + 294 + ], + "score": 1.0, + "content": "R(k−1)j is the request map from the jth agent received at", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 291, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 186, + 304 + ], + "score": 1.0, + "content": "the previous round,", + "type": "text" + }, + { + "bbox": [ + 186, + 291, + 225, + 304 + ], + "score": 0.92, + "content": "\\Phi _ { \\mathrm { s e l e c t } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 291, + 506, + 304 + ], + "score": 1.0, + "content": "is the selection function which targets to select the most critical areas", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "conditioned on the input matrix, which represents the critical level at the certain spatial location. We", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 313, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 151, + 327 + ], + "score": 1.0, + "content": "implement", + "type": "text" + }, + { + "bbox": [ + 151, + 314, + 189, + 325 + ], + "score": 0.93, + "content": "\\Phi _ { \\mathrm { s e l e c t } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 313, + 506, + 327 + ], + "score": 1.0, + "content": "by selecting the locations where the largest elements at in the given input matrix", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 324, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 506, + 337 + ], + "score": 1.0, + "content": "conditioned on the bandwidth limit; optionally, a Gaussian filter could be applied to filter out the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 335, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 506, + 347 + ], + "score": 1.0, + "content": "outliers and introduce some context. In the initial communication round, each agent selects the most", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 358 + ], + "score": 1.0, + "content": "critical areas from its own perspective as the request maps from other agents are not available yet; in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 99, + 358, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 99, + 358, + 353, + 393 + ], + "score": 1.0, + "content": "communication. Then, the selected feature map is obtained as", + "type": "text" + }, + { + "bbox": [ + 353, + 368, + 501, + 384 + ], + "score": 0.93, + "content": "\\mathcal { Z } _ { i j } ^ { ( k ) } = \\mathbf { M } _ { i j } ^ { ( k ) } \\odot \\mathcal { F } _ { i } ^ { ( k ) } \\in \\mathbb { R } ^ { H \\times \\bar { W } \\times D }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 358, + 506, + 393 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20.5, + "bbox_fs": [ + 99, + 275, + 508, + 393 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 397, + 505, + 501 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 257, + 410 + ], + "score": 1.0, + "content": "Overall, the message sent from the", + "type": "text" + }, + { + "bbox": [ + 257, + 399, + 262, + 408 + ], + "score": 0.52, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 397, + 325, + 410 + ], + "score": 1.0, + "content": "th agent to the", + "type": "text" + }, + { + "bbox": [ + 325, + 399, + 331, + 410 + ], + "score": 0.69, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 397, + 394, + 410 + ], + "score": 1.0, + "content": "th agent at the", + "type": "text" + }, + { + "bbox": [ + 395, + 398, + 401, + 408 + ], + "score": 0.64, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "th communication round", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 101, + 404, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 101, + 404, + 117, + 430 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 118, + 408, + 212, + 425 + ], + "score": 0.93, + "content": "\\mathcal { P } _ { i j } ^ { ( k ) } ~ = ~ ( \\mathbf { R } _ { i } ^ { ( k ) } , \\mathcal { \\bar { Z } } _ { i j } ^ { ( k ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 407, + 273, + 426 + ], + "score": 1.0, + "content": "Note that i)", + "type": "text" + }, + { + "bbox": [ + 274, + 409, + 295, + 424 + ], + "score": 0.92, + "content": "\\mathbf { R } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 411, + 505, + 425 + ], + "score": 1.0, + "content": "provides spatial priors to request complementary", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 102, + 424, + 508, + 445 + ], + "spans": [ + { + "bbox": [ + 102, + 424, + 397, + 445 + ], + "score": 1.0, + "content": "information for the ith agent’s need in the next round; the feature map", + "type": "text" + }, + { + "bbox": [ + 398, + 424, + 420, + 441 + ], + "score": 0.92, + "content": "\\mathcal { Z } _ { i j } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 424, + 508, + 445 + ], + "score": 1.0, + "content": "provides supportive", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 505, + 450 + ], + "score": 1.0, + "content": "information for the ith agent’s need in the this round. They together enable mutually beneficial", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 100, + 450, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 100, + 450, + 197, + 488 + ], + "score": 1.0, + "content": "collaboration; ii) since leading to low commu", + "type": "text" + }, + { + "bbox": [ + 221, + 450, + 335, + 488 + ], + "score": 1.0, + "content": "is sparse, we only transmit nn cost; and iii) the sparsity of", + "type": "text" + }, + { + "bbox": [ + 335, + 464, + 358, + 480 + ], + "score": 0.93, + "content": "\\mathcal { Z } _ { i j } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 450, + 507, + 488 + ], + "score": 1.0, + "content": "o features and corresponding indices,is determined by the binary selection", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 478, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 490 + ], + "score": 1.0, + "content": "matrix, which dynamically allocates the communication budget at various spatial areas based on their", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 489, + 394, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 394, + 501 + ], + "score": 1.0, + "content": "perceptual critical level, adapting to various communication conditions.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 28, + "bbox_fs": [ + 100, + 397, + 508, + 501 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 504, + 505, + 604 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 505, + 518 + ], + "score": 1.0, + "content": "Communication graph construction. Communication graph construction targets to identify when", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "score": 1.0, + "content": "and who to communicate to avoid unnecessary communication that wastes the bandwidth. Most", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "previous works [1, 2, 10] consider fully-connected communication graphs. When2com [12] proposes", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "score": 1.0, + "content": "a handshake mechanism, which uses similar global features to match partners. This is hard to interpret", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 549, + 506, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 562 + ], + "score": 1.0, + "content": "because two agents, which have similar global features, do not necessarily need information from", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "each other. Different from all previous works, we provide an explicit design rationale: the necessity", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 571, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 275, + 583 + ], + "score": 1.0, + "content": "of communication between the ith and the", + "type": "text" + }, + { + "bbox": [ + 275, + 572, + 281, + 582 + ], + "score": 0.73, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 281, + 571, + 505, + 583 + ], + "score": 1.0, + "content": "th agents is simply measured by the overlap between the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 355, + 594 + ], + "score": 1.0, + "content": "information that the ith agent has and the information that the", + "type": "text" + }, + { + "bbox": [ + 356, + 583, + 361, + 593 + ], + "score": 0.63, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 582, + 505, + 594 + ], + "score": 1.0, + "content": "th agent needs. With the help of the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 592, + 507, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 507, + 607 + ], + "score": 1.0, + "content": "spatial confidence map and the request map, we construct a more interpretable communication graph.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36, + "bbox_fs": [ + 104, + 504, + 507, + 607 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 608, + 505, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 622 + ], + "score": 1.0, + "content": "For the initial communication round, every agent in the system is not aware of other agents yet.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 506, + 633 + ], + "score": 1.0, + "content": "To activate the collaboration, we construct a fully-connected communication graph. Every agent", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 631, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 642 + ], + "score": 1.0, + "content": "will broadcast its message to the rest of the system. For the subsequent communication rounds, we", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 286, + 654 + ], + "score": 1.0, + "content": "examine if the communication between agent", + "type": "text" + }, + { + "bbox": [ + 286, + 642, + 291, + 651 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 641, + 332, + 654 + ], + "score": 1.0, + "content": "and agent", + "type": "text" + }, + { + "bbox": [ + 332, + 642, + 338, + 653 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "is necessary based on the maximum value", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 103, + 651, + 508, + 672 + ], + "spans": [ + { + "bbox": [ + 103, + 651, + 253, + 672 + ], + "score": 1.0, + "content": "of the binary selection matrix M(k)i→j ,", + "type": "text" + }, + { + "bbox": [ + 252, + 652, + 508, + 670 + ], + "score": 1.0, + "content": "i.e. if there is at least one patch is activated, then we regard the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 668, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 259, + 682 + ], + "score": 1.0, + "content": "connection is necessary. Formally, let", + "type": "text" + }, + { + "bbox": [ + 260, + 668, + 280, + 679 + ], + "score": 0.87, + "content": "\\mathbf { A } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 668, + 506, + 682 + ], + "score": 1.0, + "content": "be the adjacency matrix of the communication graph at", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 680, + 331, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 121, + 692 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 121, + 681, + 127, + 690 + ], + "score": 0.45, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 680, + 258, + 692 + ], + "score": 1.0, + "content": "th communication round, whose", + "type": "text" + }, + { + "bbox": [ + 258, + 680, + 279, + 692 + ], + "score": 0.91, + "content": "( i , j )", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 680, + 331, + 692 + ], + "score": 1.0, + "content": "th element is", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44, + "bbox_fs": [ + 103, + 608, + 508, + 692 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 138, + 692, + 466, + 726 + ], + "lines": [ + { + "bbox": [ + 138, + 692, + 466, + 726 + ], + "spans": [ + { + "bbox": [ + 138, + 692, + 466, + 726 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\mathbf { A } _ { i , j } ^ { ( k ) } = \\left\\{ \\begin{array} { l l } { 1 , \\quad } & { k = 0 ; } \\\\ { \\operatorname* { m a x } _ { h \\in \\{ 0 , 1 , \\dots , H - 1 \\} , w \\in \\{ 0 , 1 , \\dots , W - 1 \\} } \\left( \\mathbf { M } _ { i \\to j } ^ { ( k ) } \\right) _ { h , w } \\in \\{ 0 , 1 \\} , \\quad } & { k > 0 ; } \\end{array} \\right. } \\end{array}", + "type": "interline_equation", + "image_path": "9b9be8a18f260845fcfdd7d7078d1ce50a9932aa287637e5e4a0af8fb40291ef.jpg" + } + ] + } + ], + "index": 49, + "virtual_lines": [ + { + "bbox": [ + 138, + 692, + 466, + 703.3333333333334 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 138, + 703.3333333333334, + 466, + 714.6666666666667 + ], + "spans": [], + "index": 49 + }, + { + "bbox": [ + 138, + 714.6666666666667, + 466, + 726.0000000000001 + ], + "spans": [], + "index": 50 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 134, + 86 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 73, + 153, + 84 + ], + "score": 0.89, + "content": "h , w", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 72, + 464, + 86 + ], + "score": 1.0, + "content": "index the spatial area, reflecting message passing from the ith agent to the", + "type": "text" + }, + { + "bbox": [ + 464, + 74, + 470, + 84 + ], + "score": 0.77, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "th agent.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 84, + 485, + 96 + ], + "spans": [ + { + "bbox": [ + 107, + 84, + 485, + 96 + ], + "score": 1.0, + "content": "Given this sparse communication graph, agents can exchange messages with selected partners.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 107, + 102, + 302, + 114 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 303, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 303, + 115 + ], + "score": 1.0, + "content": "4.4 Spatial confidence-aware message fusion", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 117, + 506, + 183 + ], + "lines": [ + { + "bbox": [ + 105, + 115, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 130 + ], + "score": 1.0, + "content": "Spatial confidence-aware message fusion targets to augment the feature of each agent by aggregating", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "the received messages from the other agents. To achieve this, we adopt a transformer architecture,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 139, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 505, + 150 + ], + "score": 1.0, + "content": "which leverages multi-head attention to fuse the corresponding features from multiple agents at each", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "individual spatial location. The key technical design is to include the spatial confidence maps of all", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 174 + ], + "score": 1.0, + "content": "the agents to promote cross-agent attention learning. The intuition is that, the spatial confidence map", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 505, + 185 + ], + "score": 1.0, + "content": "could explicitly reflect the perceptually critical level, providing a useful prior for attention learning.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 187, + 506, + 258 + ], + "lines": [ + { + "bbox": [ + 102, + 183, + 509, + 209 + ], + "spans": [ + { + "bbox": [ + 102, + 183, + 401, + 209 + ], + "score": 1.0, + "content": "Specifically, for the ith agent, after receiving the jth agent’s message P(k)j→i,", + "type": "text" + }, + { + "bbox": [ + 372, + 184, + 509, + 208 + ], + "score": 1.0, + "content": "it could unpack to retrieve", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 103, + 200, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 103, + 203, + 173, + 222 + ], + "score": 1.0, + "content": "the feature map", + "type": "text" + }, + { + "bbox": [ + 173, + 202, + 196, + 219 + ], + "score": 0.93, + "content": "\\mathcal { Z } _ { j i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 203, + 327, + 222 + ], + "score": 1.0, + "content": "and the spatial confidence ma", + "type": "text" + }, + { + "bbox": [ + 321, + 200, + 354, + 223 + ], + "score": 1.0, + "content": "p C(k)j", + "type": "text" + }, + { + "bbox": [ + 328, + 203, + 399, + 220 + ], + "score": 0.93, + "content": "{ \\bf C } _ { j } ^ { ( k ) } = 1 - { \\bf R } _ { j } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 202, + 506, + 222 + ], + "score": 1.0, + "content": "R(k)j . We also include the ego", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 101, + 212, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 101, + 212, + 237, + 241 + ], + "score": 1.0, + "content": "feature map in fusion and denote", + "type": "text" + }, + { + "bbox": [ + 238, + 218, + 292, + 234 + ], + "score": 0.95, + "content": "\\mathcal { Z } _ { i i } ^ { ( k ) } = \\mathcal { F } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 216, + 298, + 237 + ], + "score": 1.0, + "content": "= F (k)i t", + "type": "text" + }, + { + "bbox": [ + 292, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "o make the formulation simple and consistent, where", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 229, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 107, + 232, + 129, + 248 + ], + "score": 0.92, + "content": "\\mathcal { Z } _ { i i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 229, + 333, + 252 + ], + "score": 1.0, + "content": "might not be sparse. To fuse the features from the", + "type": "text" + }, + { + "bbox": [ + 333, + 236, + 338, + 247 + ], + "score": 0.77, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 229, + 396, + 252 + ], + "score": 1.0, + "content": "th agent at the", + "type": "text" + }, + { + "bbox": [ + 397, + 235, + 403, + 245 + ], + "score": 0.76, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 229, + 506, + 252 + ], + "score": 1.0, + "content": "th communication round,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 246, + 329, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 329, + 258 + ], + "score": 1.0, + "content": "the cross-agent/ego attention weight for the ith agent is", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 261, + 424, + 283 + ], + "lines": [ + { + "bbox": [ + 186, + 261, + 424, + 283 + ], + "spans": [ + { + "bbox": [ + 186, + 261, + 424, + 283 + ], + "score": 0.93, + "content": "\\mathbf { W } _ { j i } ^ { ( k ) } = \\mathrm { M H A _ { W } } ( \\mathcal { F } _ { i } ^ { ( k ) } , \\mathcal { Z } _ { j i } ^ { ( k ) } , \\mathcal { Z } _ { j i } ^ { ( k ) } ) \\odot \\mathbf { C } _ { j } ^ { ( k ) } \\in \\mathbb { R } ^ { H \\times W } ,", + "type": "interline_equation", + "image_path": "da873b929a98c2c883a9b596b0996dfd598be1834530dc957050bb1c6ea3b9d0.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 186, + 261, + 424, + 283 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 287, + 506, + 353 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 133, + 300 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 287, + 177, + 299 + ], + "score": 0.89, + "content": "\\mathrm { M H A } _ { \\mathrm { W } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 286, + 506, + 300 + ], + "score": 1.0, + "content": "is a multi-head attention applied at each individual spatial location, which outputs", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "the scaled dot-product attention weight. Note that i) the proposed spatial confidence maps contributes", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "score": 1.0, + "content": "to the attention weight, as the features with higher perceptually critical level are more preferred in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "the feature aggregation; ii) the cross-agent attention weight models the collaboration strength with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 331, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 113, + 343 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 114, + 331, + 147, + 341 + ], + "score": 0.9, + "content": "H \\times W", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 331, + 506, + 343 + ], + "score": 1.0, + "content": "spatial resolution, leading to more flexible information fusion at various spatial regions.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 342, + 503, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 387, + 354 + ], + "score": 1.0, + "content": "Then, the feature map of the ith agent after fusing the messages in the", + "type": "text" + }, + { + "bbox": [ + 388, + 343, + 393, + 351 + ], + "score": 0.64, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 342, + 503, + 354 + ], + "score": 1.0, + "content": "th communication round is", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 357, + 426, + 397 + ], + "lines": [ + { + "bbox": [ + 184, + 357, + 426, + 397 + ], + "spans": [ + { + "bbox": [ + 184, + 357, + 426, + 397 + ], + "score": 0.95, + "content": "\\mathcal { F } _ { i } ^ { ( k + 1 ) } = \\mathrm { F F N } \\left( \\sum _ { j \\in \\mathcal { N } _ { i } \\bigcup \\{ i \\} } \\mathbf { W } _ { j \\to i } ^ { ( k ) } \\odot \\mathcal { Z } _ { j \\to i } ^ { ( k ) } \\right) \\in \\mathbb { R } ^ { H \\times W \\times D } ,", + "type": "interline_equation", + "image_path": "5ce0ef1adb90e497ecf0a15640ef57fa4f9dd48549ad05c36eaf46496c230cb6.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 184, + 357, + 426, + 370.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 184, + 370.3333333333333, + 426, + 383.66666666666663 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 184, + 383.66666666666663, + 426, + 396.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 400, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 133, + 413 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 401, + 165, + 413 + ], + "score": 0.63, + "content": "\\mathrm { F F N } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 400, + 300, + 413 + ], + "score": 1.0, + "content": "is the feed-forward network and", + "type": "text" + }, + { + "bbox": [ + 301, + 401, + 313, + 412 + ], + "score": 0.87, + "content": "{ \\mathcal { N } } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 400, + 506, + 413 + ], + "score": 1.0, + "content": "is the neighbors of the ith agent defined in the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 145, + 407, + 508, + 448 + ], + "spans": [ + { + "bbox": [ + 145, + 407, + 197, + 448 + ], + "score": 1.0, + "content": "cation graph round. In the", + "type": "text" + }, + { + "bbox": [ + 198, + 413, + 218, + 425 + ], + "score": 0.87, + "content": "\\mathbf { A } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 407, + 289, + 448 + ], + "score": 1.0, + "content": ". The fused featur round, we output", + "type": "text" + }, + { + "bbox": [ + 296, + 413, + 326, + 427 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { i } ^ { ( k + 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 407, + 508, + 448 + ], + "score": 1.0, + "content": "would serve as the ith agent’s feature in theo the detection decoder to generate detections.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 427, + 318, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 144, + 441 + ], + "score": 0.74, + "content": "( k + 1 ) \\operatorname { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 427, + 318, + 441 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { i } ^ { ( k + 1 ) }", + "type": "inline_equation" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 444, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "score": 1.0, + "content": "Sensor positional encoding. Sensor positional encoding represents the physical distance between", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "each agent’s sensor and its observation. It adopts a standard positional encoding function conditioned", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "on the sensing distance and feature dimension. The features are summed up with the positional", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 479, + 354, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 354, + 490 + ], + "score": 1.0, + "content": "encoding of each location before inputting to the transformer.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 494, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "Compared to existing fusion modules that do not use attention mechanism [1] or only use agent-level", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 506, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 517 + ], + "score": 1.0, + "content": "attentions [12], the per-location attention mechanism adopted by the proposed fusion emphasizes the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "location-specific feature interactions. It makes the feature fusion more targeted. Compared to the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "methods that also use the per-location attention-based fusion module[2, 10, 26], the proposed fusion", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "module leverages multi-head attention with two extra priors, including spatial confidence map and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 549, + 470, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 470, + 562 + ], + "score": 1.0, + "content": "sensing distances. Both assist attention learning to prefer high quality and critical features.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "title", + "bbox": [ + 107, + 567, + 208, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 209, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 209, + 580 + ], + "score": 1.0, + "content": "4.5 Detection decoder", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 506, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "The detection decoder decodes features into objects, including class and regression output. Given", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 102, + 588, + 508, + 611 + ], + "spans": [ + { + "bbox": [ + 102, + 588, + 197, + 611 + ], + "score": 1.0, + "content": "the feature map at the", + "type": "text" + }, + { + "bbox": [ + 198, + 596, + 204, + 605 + ], + "score": 0.56, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 588, + 306, + 611 + ], + "score": 1.0, + "content": "th communication round", + "type": "text" + }, + { + "bbox": [ + 306, + 593, + 326, + 607 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 588, + 420, + 611 + ], + "score": 1.0, + "content": ", the detection decoder", + "type": "text" + }, + { + "bbox": [ + 421, + 595, + 451, + 607 + ], + "score": 0.92, + "content": "\\Phi _ { \\mathrm { d e c } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 588, + 508, + 611 + ], + "score": 1.0, + "content": "generate the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 101, + 604, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 101, + 604, + 159, + 624 + ], + "score": 1.0, + "content": "detections of", + "type": "text" + }, + { + "bbox": [ + 159, + 611, + 164, + 619 + ], + "score": 0.31, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 604, + 208, + 624 + ], + "score": 1.0, + "content": "th agent by", + "type": "text" + }, + { + "bbox": [ + 208, + 606, + 340, + 622 + ], + "score": 0.92, + "content": "\\widehat { \\mathcal { O } } _ { i } ^ { ( k ) } = \\Phi _ { \\mathrm { d e c } } ( \\mathcal { F } _ { i } ^ { ( k ) } ) \\in \\mathbb { R } ^ { H \\times W \\times 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 604, + 435, + 624 + ], + "score": 1.0, + "content": ", where each location of", + "type": "text" + }, + { + "bbox": [ + 435, + 607, + 455, + 622 + ], + "score": 0.93, + "content": "\\widehat { \\mathcal { O } } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 604, + 506, + 624 + ], + "score": 1.0, + "content": "represents a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 619, + 507, + 634 + ], + "spans": [ + { + "bbox": [ + 104, + 619, + 195, + 634 + ], + "score": 1.0, + "content": "rotated box with class", + "type": "text" + }, + { + "bbox": [ + 195, + 622, + 300, + 632 + ], + "score": 0.9, + "content": "( c , x , y , h , w , \\cos { \\alpha } , \\sin { \\alpha } )", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 619, + 507, + 634 + ], + "score": 1.0, + "content": ", denoting class confidence, position, size and angle.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 103, + 631, + 504, + 647 + ], + "spans": [ + { + "bbox": [ + 103, + 632, + 484, + 647 + ], + "score": 1.0, + "content": "The objects are the final output of the proposed collaborative perception system. Note that", + "type": "text" + }, + { + "bbox": [ + 484, + 631, + 504, + 646 + ], + "score": 0.92, + "content": "\\widehat { \\mathcal { O } } _ { i } ^ { ( 0 ) }", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 645, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 287, + 657 + ], + "score": 1.0, + "content": "denotes the detections without collaboration.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5 + }, + { + "type": "title", + "bbox": [ + 108, + 663, + 275, + 675 + ], + "lines": [ + { + "bbox": [ + 105, + 663, + 277, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 277, + 677 + ], + "score": 1.0, + "content": "4.6 Training details and loss functions", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "To train the overall system, we supervise two tasks: spatial confidence generation and object detection", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "at each round. As mentioned before, the functionality of the spatial confidence generator is the same", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "as the classification in the detection decoder. To promote parameter efficiency, our spatial confidence", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "generator reuses the parameters of the detection decoder. For the multi-round settings, each round is", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 106, + 72, + 505, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 134, + 86 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 73, + 153, + 84 + ], + "score": 0.89, + "content": "h , w", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 72, + 464, + 86 + ], + "score": 1.0, + "content": "index the spatial area, reflecting message passing from the ith agent to the", + "type": "text" + }, + { + "bbox": [ + 464, + 74, + 470, + 84 + ], + "score": 0.77, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "th agent.", + "type": "text" + } + ], + "index": 0, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 84, + 485, + 96 + ], + "spans": [ + { + "bbox": [ + 107, + 84, + 485, + 96 + ], + "score": 1.0, + "content": "Given this sparse communication graph, agents can exchange messages with selected partners.", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 72, + 506, + 96 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 102, + 302, + 114 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 303, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 303, + 115 + ], + "score": 1.0, + "content": "4.4 Spatial confidence-aware message fusion", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 117, + 506, + 183 + ], + "lines": [ + { + "bbox": [ + 105, + 115, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 130 + ], + "score": 1.0, + "content": "Spatial confidence-aware message fusion targets to augment the feature of each agent by aggregating", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "the received messages from the other agents. To achieve this, we adopt a transformer architecture,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 139, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 505, + 150 + ], + "score": 1.0, + "content": "which leverages multi-head attention to fuse the corresponding features from multiple agents at each", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 162 + ], + "score": 1.0, + "content": "individual spatial location. The key technical design is to include the spatial confidence maps of all", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 174 + ], + "score": 1.0, + "content": "the agents to promote cross-agent attention learning. The intuition is that, the spatial confidence map", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 170, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 505, + 185 + ], + "score": 1.0, + "content": "could explicitly reflect the perceptually critical level, providing a useful prior for attention learning.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 115, + 506, + 185 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 187, + 506, + 258 + ], + "lines": [ + { + "bbox": [ + 102, + 183, + 509, + 209 + ], + "spans": [ + { + "bbox": [ + 102, + 183, + 401, + 209 + ], + "score": 1.0, + "content": "Specifically, for the ith agent, after receiving the jth agent’s message P(k)j→i,", + "type": "text" + }, + { + "bbox": [ + 372, + 184, + 509, + 208 + ], + "score": 1.0, + "content": "it could unpack to retrieve", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 103, + 200, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 103, + 203, + 173, + 222 + ], + "score": 1.0, + "content": "the feature map", + "type": "text" + }, + { + "bbox": [ + 173, + 202, + 196, + 219 + ], + "score": 0.93, + "content": "\\mathcal { Z } _ { j i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 203, + 327, + 222 + ], + "score": 1.0, + "content": "and the spatial confidence ma", + "type": "text" + }, + { + "bbox": [ + 321, + 200, + 354, + 223 + ], + "score": 1.0, + "content": "p C(k)j", + "type": "text" + }, + { + "bbox": [ + 328, + 203, + 399, + 220 + ], + "score": 0.93, + "content": "{ \\bf C } _ { j } ^ { ( k ) } = 1 - { \\bf R } _ { j } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 202, + 506, + 222 + ], + "score": 1.0, + "content": "R(k)j . We also include the ego", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 101, + 212, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 101, + 212, + 237, + 241 + ], + "score": 1.0, + "content": "feature map in fusion and denote", + "type": "text" + }, + { + "bbox": [ + 238, + 218, + 292, + 234 + ], + "score": 0.95, + "content": "\\mathcal { Z } _ { i i } ^ { ( k ) } = \\mathcal { F } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 216, + 298, + 237 + ], + "score": 1.0, + "content": "= F (k)i t", + "type": "text" + }, + { + "bbox": [ + 292, + 220, + 505, + 233 + ], + "score": 1.0, + "content": "o make the formulation simple and consistent, where", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 229, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 107, + 232, + 129, + 248 + ], + "score": 0.92, + "content": "\\mathcal { Z } _ { i i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 229, + 333, + 252 + ], + "score": 1.0, + "content": "might not be sparse. To fuse the features from the", + "type": "text" + }, + { + "bbox": [ + 333, + 236, + 338, + 247 + ], + "score": 0.77, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 229, + 396, + 252 + ], + "score": 1.0, + "content": "th agent at the", + "type": "text" + }, + { + "bbox": [ + 397, + 235, + 403, + 245 + ], + "score": 0.76, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 229, + 506, + 252 + ], + "score": 1.0, + "content": "th communication round,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 246, + 329, + 258 + ], + "spans": [ + { + "bbox": [ + 106, + 246, + 329, + 258 + ], + "score": 1.0, + "content": "the cross-agent/ego attention weight for the ith agent is", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 101, + 183, + 509, + 258 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 261, + 424, + 283 + ], + "lines": [ + { + "bbox": [ + 186, + 261, + 424, + 283 + ], + "spans": [ + { + "bbox": [ + 186, + 261, + 424, + 283 + ], + "score": 0.93, + "content": "\\mathbf { W } _ { j i } ^ { ( k ) } = \\mathrm { M H A _ { W } } ( \\mathcal { F } _ { i } ^ { ( k ) } , \\mathcal { Z } _ { j i } ^ { ( k ) } , \\mathcal { Z } _ { j i } ^ { ( k ) } ) \\odot \\mathbf { C } _ { j } ^ { ( k ) } \\in \\mathbb { R } ^ { H \\times W } ,", + "type": "interline_equation", + "image_path": "da873b929a98c2c883a9b596b0996dfd598be1834530dc957050bb1c6ea3b9d0.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 186, + 261, + 424, + 283 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 287, + 506, + 353 + ], + "lines": [ + { + "bbox": [ + 105, + 286, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 133, + 300 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 287, + 177, + 299 + ], + "score": 0.89, + "content": "\\mathrm { M H A } _ { \\mathrm { W } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 286, + 506, + 300 + ], + "score": 1.0, + "content": "is a multi-head attention applied at each individual spatial location, which outputs", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "the scaled dot-product attention weight. Note that i) the proposed spatial confidence maps contributes", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 321 + ], + "score": 1.0, + "content": "to the attention weight, as the features with higher perceptually critical level are more preferred in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "the feature aggregation; ii) the cross-agent attention weight models the collaboration strength with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 331, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 113, + 343 + ], + "score": 1.0, + "content": "a", + "type": "text" + }, + { + "bbox": [ + 114, + 331, + 147, + 341 + ], + "score": 0.9, + "content": "H \\times W", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 331, + 506, + 343 + ], + "score": 1.0, + "content": "spatial resolution, leading to more flexible information fusion at various spatial regions.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 342, + 503, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 387, + 354 + ], + "score": 1.0, + "content": "Then, the feature map of the ith agent after fusing the messages in the", + "type": "text" + }, + { + "bbox": [ + 388, + 343, + 393, + 351 + ], + "score": 0.64, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 342, + 503, + 354 + ], + "score": 1.0, + "content": "th communication round is", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 286, + 506, + 354 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 184, + 357, + 426, + 397 + ], + "lines": [ + { + "bbox": [ + 184, + 357, + 426, + 397 + ], + "spans": [ + { + "bbox": [ + 184, + 357, + 426, + 397 + ], + "score": 0.95, + "content": "\\mathcal { F } _ { i } ^ { ( k + 1 ) } = \\mathrm { F F N } \\left( \\sum _ { j \\in \\mathcal { N } _ { i } \\bigcup \\{ i \\} } \\mathbf { W } _ { j \\to i } ^ { ( k ) } \\odot \\mathcal { Z } _ { j \\to i } ^ { ( k ) } \\right) \\in \\mathbb { R } ^ { H \\times W \\times D } ,", + "type": "interline_equation", + "image_path": "5ce0ef1adb90e497ecf0a15640ef57fa4f9dd48549ad05c36eaf46496c230cb6.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 184, + 357, + 426, + 370.3333333333333 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 184, + 370.3333333333333, + 426, + 383.66666666666663 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 184, + 383.66666666666663, + 426, + 396.99999999999994 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 400, + 505, + 441 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 133, + 413 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 401, + 165, + 413 + ], + "score": 0.63, + "content": "\\mathrm { F F N } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 400, + 300, + 413 + ], + "score": 1.0, + "content": "is the feed-forward network and", + "type": "text" + }, + { + "bbox": [ + 301, + 401, + 313, + 412 + ], + "score": 0.87, + "content": "{ \\mathcal { N } } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 400, + 506, + 413 + ], + "score": 1.0, + "content": "is the neighbors of the ith agent defined in the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 145, + 407, + 508, + 448 + ], + "spans": [ + { + "bbox": [ + 145, + 407, + 197, + 448 + ], + "score": 1.0, + "content": "cation graph round. In the", + "type": "text" + }, + { + "bbox": [ + 198, + 413, + 218, + 425 + ], + "score": 0.87, + "content": "\\mathbf { A } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 407, + 289, + 448 + ], + "score": 1.0, + "content": ". The fused featur round, we output", + "type": "text" + }, + { + "bbox": [ + 296, + 413, + 326, + 427 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { i } ^ { ( k + 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 407, + 508, + 448 + ], + "score": 1.0, + "content": "would serve as the ith agent’s feature in theo the detection decoder to generate detections.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 427, + 318, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 144, + 441 + ], + "score": 0.74, + "content": "( k + 1 ) \\operatorname { t h }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 427, + 318, + 441 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { i } ^ { ( k + 1 ) }", + "type": "inline_equation" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 400, + 508, + 448 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 444, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 505, + 457 + ], + "score": 1.0, + "content": "Sensor positional encoding. Sensor positional encoding represents the physical distance between", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "each agent’s sensor and its observation. It adopts a standard positional encoding function conditioned", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 479 + ], + "score": 1.0, + "content": "on the sensing distance and feature dimension. The features are summed up with the positional", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 479, + 354, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 354, + 490 + ], + "score": 1.0, + "content": "encoding of each location before inputting to the transformer.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 445, + 505, + 490 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 494, + 505, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "Compared to existing fusion modules that do not use attention mechanism [1] or only use agent-level", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 506, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 505, + 517 + ], + "score": 1.0, + "content": "attentions [12], the per-location attention mechanism adopted by the proposed fusion emphasizes the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "location-specific feature interactions. It makes the feature fusion more targeted. Compared to the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "methods that also use the per-location attention-based fusion module[2, 10, 26], the proposed fusion", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 505, + 550 + ], + "score": 1.0, + "content": "module leverages multi-head attention with two extra priors, including spatial confidence map and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 549, + 470, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 470, + 562 + ], + "score": 1.0, + "content": "sensing distances. Both assist attention learning to prefer high quality and critical features.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 495, + 505, + 562 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 567, + 208, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 209, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 209, + 580 + ], + "score": 1.0, + "content": "4.5 Detection decoder", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 506, + 657 + ], + "lines": [ + { + "bbox": [ + 106, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "The detection decoder decodes features into objects, including class and regression output. Given", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 102, + 588, + 508, + 611 + ], + "spans": [ + { + "bbox": [ + 102, + 588, + 197, + 611 + ], + "score": 1.0, + "content": "the feature map at the", + "type": "text" + }, + { + "bbox": [ + 198, + 596, + 204, + 605 + ], + "score": 0.56, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 588, + 306, + 611 + ], + "score": 1.0, + "content": "th communication round", + "type": "text" + }, + { + "bbox": [ + 306, + 593, + 326, + 607 + ], + "score": 0.92, + "content": "\\mathcal { F } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 588, + 420, + 611 + ], + "score": 1.0, + "content": ", the detection decoder", + "type": "text" + }, + { + "bbox": [ + 421, + 595, + 451, + 607 + ], + "score": 0.92, + "content": "\\Phi _ { \\mathrm { d e c } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 588, + 508, + 611 + ], + "score": 1.0, + "content": "generate the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 101, + 604, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 101, + 604, + 159, + 624 + ], + "score": 1.0, + "content": "detections of", + "type": "text" + }, + { + "bbox": [ + 159, + 611, + 164, + 619 + ], + "score": 0.31, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 604, + 208, + 624 + ], + "score": 1.0, + "content": "th agent by", + "type": "text" + }, + { + "bbox": [ + 208, + 606, + 340, + 622 + ], + "score": 0.92, + "content": "\\widehat { \\mathcal { O } } _ { i } ^ { ( k ) } = \\Phi _ { \\mathrm { d e c } } ( \\mathcal { F } _ { i } ^ { ( k ) } ) \\in \\mathbb { R } ^ { H \\times W \\times 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 604, + 435, + 624 + ], + "score": 1.0, + "content": ", where each location of", + "type": "text" + }, + { + "bbox": [ + 435, + 607, + 455, + 622 + ], + "score": 0.93, + "content": "\\widehat { \\mathcal { O } } _ { i } ^ { ( k ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 604, + 506, + 624 + ], + "score": 1.0, + "content": "represents a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 619, + 507, + 634 + ], + "spans": [ + { + "bbox": [ + 104, + 619, + 195, + 634 + ], + "score": 1.0, + "content": "rotated box with class", + "type": "text" + }, + { + "bbox": [ + 195, + 622, + 300, + 632 + ], + "score": 0.9, + "content": "( c , x , y , h , w , \\cos { \\alpha } , \\sin { \\alpha } )", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 619, + 507, + 634 + ], + "score": 1.0, + "content": ", denoting class confidence, position, size and angle.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 103, + 631, + 504, + 647 + ], + "spans": [ + { + "bbox": [ + 103, + 632, + 484, + 647 + ], + "score": 1.0, + "content": "The objects are the final output of the proposed collaborative perception system. Note that", + "type": "text" + }, + { + "bbox": [ + 484, + 631, + 504, + 646 + ], + "score": 0.92, + "content": "\\widehat { \\mathcal { O } } _ { i } ^ { ( 0 ) }", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 645, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 287, + 657 + ], + "score": 1.0, + "content": "denotes the detections without collaboration.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5, + "bbox_fs": [ + 101, + 581, + 508, + 657 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 663, + 275, + 675 + ], + "lines": [ + { + "bbox": [ + 105, + 663, + 277, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 277, + 677 + ], + "score": 1.0, + "content": "4.6 Training details and loss functions", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "To train the overall system, we supervise two tasks: spatial confidence generation and object detection", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "at each round. As mentioned before, the functionality of the spatial confidence generator is the same", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "as the classification in the detection decoder. To promote parameter efficiency, our spatial confidence", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "generator reuses the parameters of the detection decoder. For the multi-round settings, each round is", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 101, + 395, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 101, + 395, + 315, + 424 + ], + "score": 1.0, + "content": "supervised with one detection loss, the overall loss is", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 315, + 400, + 452, + 420 + ], + "score": 0.93, + "content": "\\begin{array} { r } { L = \\sum _ { k = 0 } ^ { K } \\sum _ { i } ^ { N } L _ { \\mathrm { d e t } } \\left( \\widehat { \\mathcal { O } } _ { i } ^ { ( k ) } , \\mathcal { O } _ { i } \\right) } \\end{array}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 453, + 395, + 482, + 424 + ], + "score": 1.0, + "content": ", where", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 483, + 404, + 495, + 415 + ], + "score": 0.86, + "content": "\\mathcal { O } _ { i }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 495, + 395, + 506, + 424 + ], + "score": 1.0, + "content": "is", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 418, + 377, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 251, + 430 + ], + "score": 1.0, + "content": "the ith agent’s ground-truth objects,", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 252, + 418, + 271, + 429 + ], + "score": 0.9, + "content": "L _ { \\mathrm { d e t } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 271, + 418, + 377, + 430 + ], + "score": 1.0, + "content": "is the detection loss [28].", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 678, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 68, + 483, + 341 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 68, + 483, + 341 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 68, + 483, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 68, + 483, + 341 + ], + "score": 0.976, + "type": "image", + "image_path": "176f2727c25a10864779659ec6a4f5f489f596a152dcfd46781b9e5966104606.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 68, + 483, + 159.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 159.0, + 483, + 250.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 250.0, + 483, + 341.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 344, + 505, + 389 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Figure 3: Where2comm achieves consistently superior performance-bandwidth trade-off on all the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 373, + 368 + ], + "score": 1.0, + "content": "three collaborative perception datasets, e.g, Where2comm achieves", + "type": "text" + }, + { + "bbox": [ + 373, + 356, + 398, + 366 + ], + "score": 0.4, + "content": "5 , O O O", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 355, + 506, + 368 + ], + "score": 1.0, + "content": "times less communication", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "volume and still outperforms When2com on CoPerception-UAVs dataset. The entire red curve comes", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 378, + 377, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 377, + 389 + ], + "score": 1.0, + "content": "from a single Where2comm model evaluated at varying bandwidths.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 106, + 401, + 502, + 430 + ], + "lines": [ + { + "bbox": [ + 101, + 395, + 506, + 424 + ], + "spans": [ + { + "bbox": [ + 101, + 395, + 315, + 424 + ], + "score": 1.0, + "content": "supervised with one detection loss, the overall loss is", + "type": "text" + }, + { + "bbox": [ + 315, + 400, + 452, + 420 + ], + "score": 0.93, + "content": "\\begin{array} { r } { L = \\sum _ { k = 0 } ^ { K } \\sum _ { i } ^ { N } L _ { \\mathrm { d e t } } \\left( \\widehat { \\mathcal { O } } _ { i } ^ { ( k ) } , \\mathcal { O } _ { i } \\right) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 395, + 482, + 424 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 483, + 404, + 495, + 415 + ], + "score": 0.86, + "content": "\\mathcal { O } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 395, + 506, + 424 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 418, + 377, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 251, + 430 + ], + "score": 1.0, + "content": "the ith agent’s ground-truth objects,", + "type": "text" + }, + { + "bbox": [ + 252, + 418, + 271, + 429 + ], + "score": 0.9, + "content": "L _ { \\mathrm { d e t } }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 418, + 377, + 430 + ], + "score": 1.0, + "content": "is the detection loss [28].", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "Training strategy for multi-round setting. To adapt to multi-round communication and dynamic", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "bandwidth, we train the model under various communication settings with curriculum learning", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 456, + 507, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 507, + 469 + ], + "score": 1.0, + "content": "strategy [29]. We first gradually increase the communication bandwidth and round; and then,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "randomly sample bandwidth and round to promote robustness. Through this training strategy, a single", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 479, + 356, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 356, + 491 + ], + "score": 1.0, + "content": "model can perform well at various communication conditions.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 499, + 236, + 513 + ], + "lines": [ + { + "bbox": [ + 104, + 498, + 237, + 516 + ], + "spans": [ + { + "bbox": [ + 104, + 498, + 237, + 516 + ], + "score": 1.0, + "content": "5 Experimental Results", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "Our experiments covers four datasets, both real-world and simulation scenarios, two types of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "agents (cars and drones) and two types of sensors (LiDAR and cameras). Specifically, we con-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "duct camera-only 3D object detection in the setting of V2X-communication aided autonomous", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "driving on OPV2V dataset [10], camera-only 3D object detection in the setting of drone swarm on", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "the proposed CoPerception-UAVs dataset, and LiDAR-based 3D object detection on DAIR-V2X", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "dataset [11] and V2X-Sim dataset [9]. The detection results are evaluated by Average Precision (AP)", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "score": 1.0, + "content": "at Intersection-over-Union (IoU) threshold of 0.50 and 0.70. The communication results count the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "message size by byte in log scale with base 2. To compare communication results straightforward", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 606, + 391, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 391, + 619 + ], + "score": 1.0, + "content": "and fair, we do not consider any extra data/feature/model compression.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 625, + 279, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 279, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 279, + 640 + ], + "score": 1.0, + "content": "5.1 Datasets and experimental settings", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "OPV2V. OPV2V [10] is a vehicle-to-vehicle collaborative perception dataset, co-simulated by", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 650, + 504, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 504, + 662 + ], + "score": 1.0, + "content": "OpenCDA [10] and Carla [30]. It includes 12K frames of 3D point clouds and RGB images with", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 660, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 309, + 675 + ], + "score": 1.0, + "content": "230K annotated 3D boxes. The perception range is", + "type": "text" + }, + { + "bbox": [ + 309, + 662, + 353, + 672 + ], + "score": 0.89, + "content": "4 0 \\mathrm { m } \\times 4 0 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 660, + 505, + 675 + ], + "score": 1.0, + "content": ". For camera-only 3D object detection", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 672, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 505, + 686 + ], + "score": 1.0, + "content": "task on OPV2V, we implement the detector following CADDN [31]. The input front-view image size", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 682, + 502, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 502, + 697 + ], + "score": 1.0, + "content": "is (416, 160). The front-view input feature map is transformed to BEV with resolution 0.5m/pixel.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "V2X-Sim. V2X-Sim [9] is a vehicle-to-everything collaborative perception dataset, co-simulated", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "score": 1.0, + "content": "by SUMO [32] and Carla, including 10K frames of 3D LiDAR point clouds and 501K 3D boxes.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 68, + 483, + 341 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 68, + 483, + 341 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 68, + 483, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 68, + 483, + 341 + ], + "score": 0.976, + "type": "image", + "image_path": "176f2727c25a10864779659ec6a4f5f489f596a152dcfd46781b9e5966104606.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 68, + 483, + 159.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 159.0, + 483, + 250.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 250.0, + 483, + 341.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 344, + 505, + 389 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Figure 3: Where2comm achieves consistently superior performance-bandwidth trade-off on all the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 355, + 506, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 373, + 368 + ], + "score": 1.0, + "content": "three collaborative perception datasets, e.g, Where2comm achieves", + "type": "text" + }, + { + "bbox": [ + 373, + 356, + 398, + 366 + ], + "score": 0.4, + "content": "5 , O O O", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 355, + 506, + 368 + ], + "score": 1.0, + "content": "times less communication", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 379 + ], + "score": 1.0, + "content": "volume and still outperforms When2com on CoPerception-UAVs dataset. The entire red curve comes", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 378, + 377, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 377, + 389 + ], + "score": 1.0, + "content": "from a single Where2comm model evaluated at varying bandwidths.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "text", + "bbox": [ + 106, + 401, + 502, + 430 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 101, + 395, + 506, + 430 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 434, + 505, + 489 + ], + "lines": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "Training strategy for multi-round setting. To adapt to multi-round communication and dynamic", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "bandwidth, we train the model under various communication settings with curriculum learning", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 456, + 507, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 507, + 469 + ], + "score": 1.0, + "content": "strategy [29]. We first gradually increase the communication bandwidth and round; and then,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 505, + 480 + ], + "score": 1.0, + "content": "randomly sample bandwidth and round to promote robustness. Through this training strategy, a single", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 479, + 356, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 356, + 491 + ], + "score": 1.0, + "content": "model can perform well at various communication conditions.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 433, + 507, + 491 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 499, + 236, + 513 + ], + "lines": [ + { + "bbox": [ + 104, + 498, + 237, + 516 + ], + "spans": [ + { + "bbox": [ + 104, + 498, + 237, + 516 + ], + "score": 1.0, + "content": "5 Experimental Results", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "Our experiments covers four datasets, both real-world and simulation scenarios, two types of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 529, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 506, + 542 + ], + "score": 1.0, + "content": "agents (cars and drones) and two types of sensors (LiDAR and cameras). Specifically, we con-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "duct camera-only 3D object detection in the setting of V2X-communication aided autonomous", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "driving on OPV2V dataset [10], camera-only 3D object detection in the setting of drone swarm on", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 505, + 575 + ], + "score": 1.0, + "content": "the proposed CoPerception-UAVs dataset, and LiDAR-based 3D object detection on DAIR-V2X", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 586 + ], + "score": 1.0, + "content": "dataset [11] and V2X-Sim dataset [9]. The detection results are evaluated by Average Precision (AP)", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "score": 1.0, + "content": "at Intersection-over-Union (IoU) threshold of 0.50 and 0.70. The communication results count the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "message size by byte in log scale with base 2. To compare communication results straightforward", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 606, + 391, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 391, + 619 + ], + "score": 1.0, + "content": "and fair, we do not consider any extra data/feature/model compression.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 519, + 506, + 619 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 625, + 279, + 637 + ], + "lines": [ + { + "bbox": [ + 105, + 623, + 279, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 279, + 640 + ], + "score": 1.0, + "content": "5.1 Datasets and experimental settings", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "score": 1.0, + "content": "OPV2V. OPV2V [10] is a vehicle-to-vehicle collaborative perception dataset, co-simulated by", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 650, + 504, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 504, + 662 + ], + "score": 1.0, + "content": "OpenCDA [10] and Carla [30]. It includes 12K frames of 3D point clouds and RGB images with", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 660, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 309, + 675 + ], + "score": 1.0, + "content": "230K annotated 3D boxes. The perception range is", + "type": "text" + }, + { + "bbox": [ + 309, + 662, + 353, + 672 + ], + "score": 0.89, + "content": "4 0 \\mathrm { m } \\times 4 0 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 660, + 505, + 675 + ], + "score": 1.0, + "content": ". For camera-only 3D object detection", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 672, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 505, + 686 + ], + "score": 1.0, + "content": "task on OPV2V, we implement the detector following CADDN [31]. The input front-view image size", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 682, + 502, + 697 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 502, + 697 + ], + "score": 1.0, + "content": "is (416, 160). The front-view input feature map is transformed to BEV with resolution 0.5m/pixel.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 639, + 505, + 697 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 699, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "V2X-Sim. V2X-Sim [9] is a vehicle-to-everything collaborative perception dataset, co-simulated", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "score": 1.0, + "content": "by SUMO [32] and Carla, including 10K frames of 3D LiDAR point clouds and 501K 3D boxes.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 699, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 70, + 479, + 173 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 70, + 479, + 173 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 70, + 479, + 173 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 479, + 173 + ], + "score": 0.966, + "type": "image", + "image_path": "3b86090779d433b47baa89a2d959ad84c90159d6416807244562537c86cf9192.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 70, + 479, + 104.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 104.33333333333334, + 479, + 138.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 138.66666666666669, + 479, + 173.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 174, + 495, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 172, + 497, + 187 + ], + "spans": [ + { + "bbox": [ + 112, + 172, + 497, + 187 + ], + "score": 1.0, + "content": "Figure 4: More communication rounds continuously improve performance-bandwidth trade-off.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 194, + 505, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 201, + 206 + ], + "score": 1.0, + "content": "The perception range is", + "type": "text" + }, + { + "bbox": [ + 201, + 194, + 245, + 204 + ], + "score": 0.9, + "content": "6 4 \\mathrm { m } \\times 6 4 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 193, + 505, + 206 + ], + "score": 1.0, + "content": ". For LiDAR-based 3D object detection task, our detector follows", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "MotionNet [33]. We discretize 3D points into a BEV map with size (256, 256, 13) and the resolution", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 216, + 311, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 115, + 228 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 116, + 216, + 138, + 227 + ], + "score": 0.62, + "content": "0 . 4 \\mathrm { m } \\iota", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 216, + 243, + 228 + ], + "score": 1.0, + "content": "/pixel in length and width,", + "type": "text" + }, + { + "bbox": [ + 243, + 216, + 271, + 226 + ], + "score": 0.69, + "content": "0 . 2 5 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 216, + 311, + 228 + ], + "score": 1.0, + "content": "in height.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 506, + 244 + ], + "score": 1.0, + "content": "CoPerception-UAVs. To enrich the collaborative perception datasets, we consider the swarm of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 243, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 255 + ], + "score": 1.0, + "content": "unmanned aerial vehicles (UAV) and propose a UAV-swarm-based collaborative perception dataset:", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 254, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 506, + 266 + ], + "score": 1.0, + "content": "CoPerception-UAVs, co-simulated by AirSim [34] and Carla [30], including 131.9K aerial images and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 265, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 270, + 277 + ], + "score": 1.0, + "content": "1.94M 3D boxes. The perception range is", + "type": "text" + }, + { + "bbox": [ + 270, + 265, + 323, + 275 + ], + "score": 0.91, + "content": "2 0 0 \\mathrm { m } \\times 3 5 0 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 265, + 506, + 277 + ], + "score": 1.0, + "content": ". For the camera-only 3D object detection task", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "on CoPerception-UAVs, our detector follows DVDET [8]. The input aerial image size is (800, 450).", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 421, + 299 + ], + "score": 1.0, + "content": "The aerial-view input feature map is transformed to BEV with the resolution of", + "type": "text" + }, + { + "bbox": [ + 421, + 287, + 448, + 297 + ], + "score": 0.42, + "content": "0 . 2 5 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "/pixel, and the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 298, + 302, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 302, + 310 + ], + "score": 1.0, + "content": "size is (192, 352); see more details in Appendix.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "DAIR-V2X. DAIR-V2X [11] is the only public real-world collaborative perception dataset. Each", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "sample contains two agents: a vehicle and an infrastructure, with 3D annotations. The perception", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 335, + 507, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 142, + 349 + ], + "score": 1.0, + "content": "range is", + "type": "text" + }, + { + "bbox": [ + 142, + 336, + 200, + 346 + ], + "score": 0.88, + "content": "2 0 1 . 6 \\mathrm { m } \\times 8 0 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 335, + 507, + 349 + ], + "score": 1.0, + "content": ". Originally DAIR-V2X does not label objects outside the camera’s view,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 347, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 506, + 359 + ], + "score": 1.0, + "content": "we relabel all objects to cover 360-degree detection range. We complement several intermediate", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 358, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 369 + ], + "score": 1.0, + "content": "fusion-based baselines on DAIR-V2X to comprehensively validate our method on real data. For", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "score": 1.0, + "content": "LiDAR-based 3D object detection task, our detector follows PointPillar [35]. We represent the field", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 381, + 392 + ], + "score": 1.0, + "content": "of view into a BEV map with size (200, 504, 64) and the resolution is", + "type": "text" + }, + { + "bbox": [ + 381, + 380, + 403, + 390 + ], + "score": 0.53, + "content": "0 . 4 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "/pixel in length and width.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 107, + 399, + 231, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 231, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 231, + 412 + ], + "score": 1.0, + "content": "5.2 Quantitative evaluation", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 414, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "Benchmark comparison. Fig. 3 compares the proposed Where2comm with the previous methods in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 315, + 437 + ], + "score": 1.0, + "content": "terms of the trade-off between detection performance", + "type": "text" + }, + { + "bbox": [ + 315, + 425, + 379, + 435 + ], + "score": 0.74, + "content": "( \\mathbf { A P } @ \\mathbf { I o U } { = } 0 . 5 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 425, + 506, + 437 + ], + "score": 1.0, + "content": "and communication bandwidth;", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "also see exact values in Table 3 of Appendix. We consider single-agent detection without collaboration", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 444, + 507, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 132, + 461 + ], + "score": 0.88, + "content": "( \\widehat { \\mathcal { O } } _ { i } ^ { ( 0 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 444, + 507, + 465 + ], + "score": 1.0, + "content": ", When2com [12], V2VNet [1], DiscoNet [2], V2X-ViT [26] and late fusion, where agents", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "directly exchange the detected 3D boxes. The red curve comes from a single Where2comm model", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 470, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 483 + ], + "score": 1.0, + "content": "evaluated at varying bandwidths. We see that the proposed Where2comm: i) achieves a far-more", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 482, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 493 + ], + "score": 1.0, + "content": "superior perception-communication trade-off across all the communication bandwidth choices and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "various collaborative perception tasks, including camera-only 3D object detection from aerial view", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "score": 1.0, + "content": "and car front view, and LiDAR-based 3D object detection; ii) achieves significant improvements over", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "previous state-of-the-arts on both real-world (DAIR-V2X) and simulation scenarios, improves the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 201, + 538 + ], + "score": 1.0, + "content": "SOTA performance by", + "type": "text" + }, + { + "bbox": [ + 201, + 525, + 225, + 536 + ], + "score": 0.85, + "content": "7 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 525, + 291, + 538 + ], + "score": 1.0, + "content": "on DAIR-V2X,", + "type": "text" + }, + { + "bbox": [ + 292, + 525, + 320, + 536 + ], + "score": 0.86, + "content": "6 . 6 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 525, + 422, + 538 + ], + "score": 1.0, + "content": "on CoPerception-UAVs,", + "type": "text" + }, + { + "bbox": [ + 423, + 525, + 455, + 536 + ], + "score": 0.88, + "content": "2 5 . 8 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "on OPV2V,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 128, + 547 + ], + "score": 0.85, + "content": "1 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "on V2X-Sim; iii) achieves the same detection performance of previous state-of-the-arts with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "extremely less communication volume: 5128 times less on CoPerception-UAVs, more than 100K", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 559, + 429, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 429, + 569 + ], + "score": 1.0, + "content": "times less on OPV2V, 55 times less on V2X-Sim, 105 times less on DAIR-V2X.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 574, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "Multi-round evaluation. Fig. 4 presents the performances of Where2comm at communication rounds", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 585, + 507, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 507, + 597 + ], + "score": 1.0, + "content": "ranging from 1 to 3. Each curve comes from a single Where2comm model with a certain communica-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "tion round evaluated at varying bandwidths. Results show that 1 communication round is good, more", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 608, + 504, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 504, + 619 + ], + "score": 1.0, + "content": "rounds are even better. Multi-round communication steadily improves the performance-bandwidth", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "trade-off across all three datasets, reflecting its effectiveness and robustness. This encourages the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 629, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 641 + ], + "score": 1.0, + "content": "agents to actively collaborate without worrying the performance degradation. This also validates", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 640, + 440, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 440, + 651 + ], + "score": 1.0, + "content": "that Where2comm can well work at various communication bandwidths and rounds.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 656, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "Robustness to localization noise. We follow the localization noise setting in V2VNet and V2X-ViT", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 234, + 679 + ], + "score": 1.0, + "content": "(Gaussian noise with a mean of", + "type": "text" + }, + { + "bbox": [ + 235, + 668, + 249, + 677 + ], + "score": 0.35, + "content": "_ { 0 \\mathrm { m } }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 667, + 360, + 679 + ], + "score": 1.0, + "content": "and a standard deviation of", + "type": "text" + }, + { + "bbox": [ + 360, + 667, + 399, + 678 + ], + "score": 0.79, + "content": "0 \\mathrm { m } { - } 0 . 6 \\mathrm { m } \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "and conduct experiments", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "on all the three datasets to validate the robustness against realistic localization noise. Where2comm is", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "more robust to the localization noise than previous SOTAs. Fig. 5 shows the detection performances", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "as a function of localization noise level in CoPerception-UAVs, OPV2V and V2X-Sim datasets,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "score": 1.0, + "content": "respectively We see: i) overall the collaborative perception performance degrades with the increasing", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 70, + 479, + 173 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 70, + 479, + 173 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 70, + 479, + 173 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 479, + 173 + ], + "score": 0.966, + "type": "image", + "image_path": "3b86090779d433b47baa89a2d959ad84c90159d6416807244562537c86cf9192.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 70, + 479, + 104.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 104.33333333333334, + 479, + 138.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 138.66666666666669, + 479, + 173.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 174, + 495, + 186 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 172, + 497, + 187 + ], + "spans": [ + { + "bbox": [ + 112, + 172, + 497, + 187 + ], + "score": 1.0, + "content": "Figure 4: More communication rounds continuously improve performance-bandwidth trade-off.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 194, + 505, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 193, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 201, + 206 + ], + "score": 1.0, + "content": "The perception range is", + "type": "text" + }, + { + "bbox": [ + 201, + 194, + 245, + 204 + ], + "score": 0.9, + "content": "6 4 \\mathrm { m } \\times 6 4 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 193, + 505, + 206 + ], + "score": 1.0, + "content": ". For LiDAR-based 3D object detection task, our detector follows", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "MotionNet [33]. We discretize 3D points into a BEV map with size (256, 256, 13) and the resolution", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 216, + 311, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 115, + 228 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 116, + 216, + 138, + 227 + ], + "score": 0.62, + "content": "0 . 4 \\mathrm { m } \\iota", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 216, + 243, + 228 + ], + "score": 1.0, + "content": "/pixel in length and width,", + "type": "text" + }, + { + "bbox": [ + 243, + 216, + 271, + 226 + ], + "score": 0.69, + "content": "0 . 2 5 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 216, + 311, + 228 + ], + "score": 1.0, + "content": "in height.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 193, + 505, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 505, + 309 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 506, + 244 + ], + "score": 1.0, + "content": "CoPerception-UAVs. To enrich the collaborative perception datasets, we consider the swarm of", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 243, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 255 + ], + "score": 1.0, + "content": "unmanned aerial vehicles (UAV) and propose a UAV-swarm-based collaborative perception dataset:", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 254, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 506, + 266 + ], + "score": 1.0, + "content": "CoPerception-UAVs, co-simulated by AirSim [34] and Carla [30], including 131.9K aerial images and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 265, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 270, + 277 + ], + "score": 1.0, + "content": "1.94M 3D boxes. The perception range is", + "type": "text" + }, + { + "bbox": [ + 270, + 265, + 323, + 275 + ], + "score": 0.91, + "content": "2 0 0 \\mathrm { m } \\times 3 5 0 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 265, + 506, + 277 + ], + "score": 1.0, + "content": ". For the camera-only 3D object detection task", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "on CoPerception-UAVs, our detector follows DVDET [8]. The input aerial image size is (800, 450).", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 287, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 421, + 299 + ], + "score": 1.0, + "content": "The aerial-view input feature map is transformed to BEV with the resolution of", + "type": "text" + }, + { + "bbox": [ + 421, + 287, + 448, + 297 + ], + "score": 0.42, + "content": "0 . 2 5 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 287, + 505, + 299 + ], + "score": 1.0, + "content": "/pixel, and the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 298, + 302, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 302, + 310 + ], + "score": 1.0, + "content": "size is (192, 352); see more details in Appendix.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 232, + 506, + 310 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 313, + 505, + 391 + ], + "lines": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "DAIR-V2X. DAIR-V2X [11] is the only public real-world collaborative perception dataset. Each", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "sample contains two agents: a vehicle and an infrastructure, with 3D annotations. The perception", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 335, + 507, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 142, + 349 + ], + "score": 1.0, + "content": "range is", + "type": "text" + }, + { + "bbox": [ + 142, + 336, + 200, + 346 + ], + "score": 0.88, + "content": "2 0 1 . 6 \\mathrm { m } \\times 8 0 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 335, + 507, + 349 + ], + "score": 1.0, + "content": ". Originally DAIR-V2X does not label objects outside the camera’s view,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 347, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 506, + 359 + ], + "score": 1.0, + "content": "we relabel all objects to cover 360-degree detection range. We complement several intermediate", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 358, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 369 + ], + "score": 1.0, + "content": "fusion-based baselines on DAIR-V2X to comprehensively validate our method on real data. For", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 505, + 380 + ], + "score": 1.0, + "content": "LiDAR-based 3D object detection task, our detector follows PointPillar [35]. We represent the field", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 381, + 392 + ], + "score": 1.0, + "content": "of view into a BEV map with size (200, 504, 64) and the resolution is", + "type": "text" + }, + { + "bbox": [ + 381, + 380, + 403, + 390 + ], + "score": 0.53, + "content": "0 . 4 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "/pixel in length and width.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 313, + 507, + 392 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 399, + 231, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 231, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 231, + 412 + ], + "score": 1.0, + "content": "5.2 Quantitative evaluation", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 414, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 506, + 426 + ], + "score": 1.0, + "content": "Benchmark comparison. Fig. 3 compares the proposed Where2comm with the previous methods in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 425, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 315, + 437 + ], + "score": 1.0, + "content": "terms of the trade-off between detection performance", + "type": "text" + }, + { + "bbox": [ + 315, + 425, + 379, + 435 + ], + "score": 0.74, + "content": "( \\mathbf { A P } @ \\mathbf { I o U } { = } 0 . 5 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 425, + 506, + 437 + ], + "score": 1.0, + "content": "and communication bandwidth;", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "also see exact values in Table 3 of Appendix. We consider single-agent detection without collaboration", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 444, + 507, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 132, + 461 + ], + "score": 0.88, + "content": "( \\widehat { \\mathcal { O } } _ { i } ^ { ( 0 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 444, + 507, + 465 + ], + "score": 1.0, + "content": ", When2com [12], V2VNet [1], DiscoNet [2], V2X-ViT [26] and late fusion, where agents", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "directly exchange the detected 3D boxes. The red curve comes from a single Where2comm model", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 470, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 506, + 483 + ], + "score": 1.0, + "content": "evaluated at varying bandwidths. We see that the proposed Where2comm: i) achieves a far-more", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 482, + 506, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 493 + ], + "score": 1.0, + "content": "superior perception-communication trade-off across all the communication bandwidth choices and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "various collaborative perception tasks, including camera-only 3D object detection from aerial view", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 506, + 516 + ], + "score": 1.0, + "content": "and car front view, and LiDAR-based 3D object detection; ii) achieves significant improvements over", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 506, + 527 + ], + "score": 1.0, + "content": "previous state-of-the-arts on both real-world (DAIR-V2X) and simulation scenarios, improves the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 201, + 538 + ], + "score": 1.0, + "content": "SOTA performance by", + "type": "text" + }, + { + "bbox": [ + 201, + 525, + 225, + 536 + ], + "score": 0.85, + "content": "7 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 525, + 291, + 538 + ], + "score": 1.0, + "content": "on DAIR-V2X,", + "type": "text" + }, + { + "bbox": [ + 292, + 525, + 320, + 536 + ], + "score": 0.86, + "content": "6 . 6 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 525, + 422, + 538 + ], + "score": 1.0, + "content": "on CoPerception-UAVs,", + "type": "text" + }, + { + "bbox": [ + 423, + 525, + 455, + 536 + ], + "score": 0.88, + "content": "2 5 . 8 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "on OPV2V,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 128, + 547 + ], + "score": 0.85, + "content": "1 . 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "on V2X-Sim; iii) achieves the same detection performance of previous state-of-the-arts with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 506, + 559 + ], + "score": 1.0, + "content": "extremely less communication volume: 5128 times less on CoPerception-UAVs, more than 100K", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 559, + 429, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 429, + 569 + ], + "score": 1.0, + "content": "times less on OPV2V, 55 times less on V2X-Sim, 105 times less on DAIR-V2X.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 414, + 507, + 569 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 574, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "Multi-round evaluation. Fig. 4 presents the performances of Where2comm at communication rounds", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 585, + 507, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 507, + 597 + ], + "score": 1.0, + "content": "ranging from 1 to 3. Each curve comes from a single Where2comm model with a certain communica-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 609 + ], + "score": 1.0, + "content": "tion round evaluated at varying bandwidths. Results show that 1 communication round is good, more", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 608, + 504, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 504, + 619 + ], + "score": 1.0, + "content": "rounds are even better. Multi-round communication steadily improves the performance-bandwidth", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 630 + ], + "score": 1.0, + "content": "trade-off across all three datasets, reflecting its effectiveness and robustness. This encourages the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 629, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 506, + 641 + ], + "score": 1.0, + "content": "agents to actively collaborate without worrying the performance degradation. This also validates", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 640, + 440, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 440, + 651 + ], + "score": 1.0, + "content": "that Where2comm can well work at various communication bandwidths and rounds.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 573, + 507, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 656, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "Robustness to localization noise. We follow the localization noise setting in V2VNet and V2X-ViT", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 234, + 679 + ], + "score": 1.0, + "content": "(Gaussian noise with a mean of", + "type": "text" + }, + { + "bbox": [ + 235, + 668, + 249, + 677 + ], + "score": 0.35, + "content": "_ { 0 \\mathrm { m } }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 667, + 360, + 679 + ], + "score": 1.0, + "content": "and a standard deviation of", + "type": "text" + }, + { + "bbox": [ + 360, + 667, + 399, + 678 + ], + "score": 0.79, + "content": "0 \\mathrm { m } { - } 0 . 6 \\mathrm { m } \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "and conduct experiments", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "on all the three datasets to validate the robustness against realistic localization noise. Where2comm is", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "more robust to the localization noise than previous SOTAs. Fig. 5 shows the detection performances", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "as a function of localization noise level in CoPerception-UAVs, OPV2V and V2X-Sim datasets,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "score": 1.0, + "content": "respectively We see: i) overall the collaborative perception performance degrades with the increasing", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "localization noise, while where2comm outperforms previous SOTAs (When2com, V2VNet,DiscoNet)", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "under all the localization noise. ii) where2comm keeps being superior to No Collaboration while", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 458, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 237, + 470 + ], + "score": 1.0, + "content": "V2VNet fails when noise is over", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 237, + 459, + 259, + 469 + ], + "score": 0.65, + "content": "0 . 4 \\mathrm { m }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 260, + 458, + 412, + 470 + ], + "score": 1.0, + "content": "and DiscoNet fails when noise is over", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 412, + 459, + 434, + 469 + ], + "score": 0.63, + "content": "0 . 5 \\mathrm { m }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 434, + 458, + 506, + 470 + ], + "score": 1.0, + "content": "on CoPerception-", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "score": 1.0, + "content": "UAVs. The reasons are: i) the powerful transformer architecture in fusion module attentively select", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 479, + 507, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 507, + 494 + ], + "score": 1.0, + "content": "the most suitable collaborative feature; ii) the spatial confidence map helps filter out noisy features,", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 491, + 427, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 427, + 503 + ], + "score": 1.0, + "content": "these two designs work together to mitigate noise localization distortion effects.", + "type": "text", + "cross_page": true + } + ], + "index": 19 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 656, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 69, + 492, + 176 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 69, + 492, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 69, + 492, + 176 + ], + "spans": [ + { + "bbox": [ + 110, + 69, + 492, + 176 + ], + "score": 0.968, + "type": "image", + "image_path": "e81be3f9beccbf3077247e2c41e4663271e0e8ccc1c3dcea2bf8dc123852fe69.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 69, + 492, + 104.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 104.66666666666666, + 492, + 140.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 140.33333333333331, + 492, + 175.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 180, + 504, + 203 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 180, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 505, + 192 + ], + "score": 1.0, + "content": "Figure 5: Robustness to localization error. Gaussian noise with zero mean and varying std is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 191, + 472, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 472, + 203 + ], + "score": 1.0, + "content": "introduced. Where2comm consistently outperforms previous SOTAs and No Collaboration.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 111, + 207, + 498, + 347 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 207, + 498, + 347 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 111, + 207, + 498, + 347 + ], + "spans": [ + { + "bbox": [ + 111, + 207, + 498, + 347 + ], + "score": 0.861, + "type": "image", + "image_path": "6ddd22e7a49a063f8a100e6e80d7a1b645b5b68a459f6286f53ab8758bfe4a63.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 111, + 207, + 498, + 253.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 111, + 253.66666666666666, + 498, + 300.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 300.3333333333333, + 498, + 347.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 348, + 506, + 422 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "score": 1.0, + "content": "Figure 6: Visualization of collaboration between Drone 1 and Drone 2 on CoPerception-UAVs", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 356, + 503, + 377 + ], + "spans": [ + { + "bbox": [ + 102, + 356, + 270, + 377 + ], + "score": 1.0, + "content": "dataset, including spatial confidence map", + "type": "text" + }, + { + "bbox": [ + 270, + 358, + 295, + 373 + ], + "score": 0.87, + "content": "( \\mathbf { C } _ { 1 } ^ { ( 0 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 356, + 366, + 377 + ], + "score": 1.0, + "content": ", selection matrix", + "type": "text" + }, + { + "bbox": [ + 366, + 358, + 399, + 373 + ], + "score": 0.89, + "content": "( \\mathbf { M } _ { 1 2 } ^ { ( 0 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 356, + 441, + 377 + ], + "score": 1.0, + "content": ", message", + "type": "text" + }, + { + "bbox": [ + 441, + 358, + 503, + 373 + ], + "score": 0.9, + "content": "( \\{ \\bar { \\mathbf { R } _ { 2 } ^ { ( 0 ) } } , \\mathcal { Z } _ { 2 1 } ^ { ( 0 ) } \\} )", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 366, + 511, + 393 + ], + "spans": [ + { + "bbox": [ + 102, + 369, + 410, + 391 + ], + "score": 1.0, + "content": "in the communication module, attention weight in the fusion module (W(0)1→1", + "type": "text" + }, + { + "bbox": [ + 378, + 372, + 443, + 388 + ], + "score": 0.9, + "content": "( \\mathbf { \\bar { W } } _ { 1 \\to 1 } ^ { ( 0 ) } , \\mathbf { W } _ { 2 \\to 1 } ^ { ( 0 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 366, + 511, + 393 + ], + "score": 1.0, + "content": "2 2 1, and Drone 1’s", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 384, + 508, + 403 + ], + "spans": [ + { + "bbox": [ + 104, + 384, + 204, + 403 + ], + "score": 1.0, + "content": "detection results before", + "type": "text" + }, + { + "bbox": [ + 204, + 386, + 230, + 401 + ], + "score": 0.88, + "content": "( \\widehat { \\mathcal { O } } _ { 1 } ^ { ( 0 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 384, + 270, + 403 + ], + "score": 1.0, + "content": "and after", + "type": "text" + }, + { + "bbox": [ + 270, + 386, + 296, + 401 + ], + "score": 0.87, + "content": "( \\widehat { \\mathcal { O } } _ { 1 } ^ { ( 1 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 384, + 508, + 403 + ], + "score": 1.0, + "content": "collaboration. Green and red boxes denote ground-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "truth and detection, respectively. The objects occluded by a tall building can be detected through", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 410, + 359, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 359, + 424 + ], + "score": 1.0, + "content": "transmitting spatially sparse, yet perceptually critical message.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + } + ], + "index": 8.25 + }, + { + "type": "text", + "bbox": [ + 107, + 436, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "localization noise, while where2comm outperforms previous SOTAs (When2com, V2VNet,DiscoNet)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "under all the localization noise. ii) where2comm keeps being superior to No Collaboration while", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 458, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 237, + 470 + ], + "score": 1.0, + "content": "V2VNet fails when noise is over", + "type": "text" + }, + { + "bbox": [ + 237, + 459, + 259, + 469 + ], + "score": 0.65, + "content": "0 . 4 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 458, + 412, + 470 + ], + "score": 1.0, + "content": "and DiscoNet fails when noise is over", + "type": "text" + }, + { + "bbox": [ + 412, + 459, + 434, + 469 + ], + "score": 0.63, + "content": "0 . 5 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 458, + 506, + 470 + ], + "score": 1.0, + "content": "on CoPerception-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "score": 1.0, + "content": "UAVs. The reasons are: i) the powerful transformer architecture in fusion module attentively select", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 479, + 507, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 507, + 494 + ], + "score": 1.0, + "content": "the most suitable collaborative feature; ii) the spatial confidence map helps filter out noisy features,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 491, + 427, + 503 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 427, + 503 + ], + "score": 1.0, + "content": "these two designs work together to mitigate noise localization distortion effects.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 107, + 514, + 224, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 226, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 226, + 528 + ], + "score": 1.0, + "content": "5.3 Qualitative evaluation", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 530, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "Visualization of spatial confidence map. Fig. 6 illustrates how Where2comm is empowered by the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "proposed spatial confidence map. In the scene, Drone 1’s view is occluded by a tall building. With", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 551, + 507, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 507, + 566 + ], + "score": 1.0, + "content": "Drone 2’s help, Drone 1 is able to detect through occlusion. Fig. 6 (a-d) shows Drone 1’s observation,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "spatial confidence map (1), binary selection matrix (2), and ego attention weight (3). Fig. 6 (f-h)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "shows Drone 2’s observation and message sent to Drone 1, including the request map (opposite of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 585, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 506, + 598 + ], + "score": 1.0, + "content": "confidence map) and the sparse feature map, achieving efficient communication. Fig. 6 (i) shows", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "the attention weight for Drone 1 to fuse Drone 2’s messages, which is sparse, yet highlights the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "objects’ positions. Fig. 6 (e) and (j) compares the detection results before and after the collaboration", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "with Drone 2. We see that the proposed spatial confidence map contributes to spatially sparse, yet", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 629, + 456, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 456, + 642 + ], + "score": 1.0, + "content": "perceptually critical message, which effectively helps Drone 1 detect occluded objects.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "Visualization of detection results. Fig. 7 shows that compared to No Collaboration, When2com", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "and DiscoNet, Where2comm is able to achieves more complete and accurate detection results. The", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "reason is that When2com employs a scalar to denote the agent-to-agent attention, which cannot", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "distinguish which spatial area is more informative; DiscoNet employs a MLP-based fusion weight", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "learning, which cannot well capture the complex collaboration attention; while Where2comm can", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "zoom in to critical spatial areas in a cell-level resolution and leverage the spatial confidence map and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 711, + 372, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 372, + 724 + ], + "score": 1.0, + "content": "sensing distances as priors to achieve more comprehensive fusion.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 69, + 492, + 176 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 69, + 492, + 176 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 69, + 492, + 176 + ], + "spans": [ + { + "bbox": [ + 110, + 69, + 492, + 176 + ], + "score": 0.968, + "type": "image", + "image_path": "e81be3f9beccbf3077247e2c41e4663271e0e8ccc1c3dcea2bf8dc123852fe69.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 69, + 492, + 104.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 104.66666666666666, + 492, + 140.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 140.33333333333331, + 492, + 175.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 180, + 504, + 203 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 180, + 505, + 192 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 505, + 192 + ], + "score": 1.0, + "content": "Figure 5: Robustness to localization error. Gaussian noise with zero mean and varying std is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 191, + 472, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 472, + 203 + ], + "score": 1.0, + "content": "introduced. Where2comm consistently outperforms previous SOTAs and No Collaboration.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 111, + 207, + 498, + 347 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 207, + 498, + 347 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 111, + 207, + 498, + 347 + ], + "spans": [ + { + "bbox": [ + 111, + 207, + 498, + 347 + ], + "score": 0.861, + "type": "image", + "image_path": "6ddd22e7a49a063f8a100e6e80d7a1b645b5b68a459f6286f53ab8758bfe4a63.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 111, + 207, + 498, + 253.66666666666666 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 111, + 253.66666666666666, + 498, + 300.3333333333333 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 111, + 300.3333333333333, + 498, + 347.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 348, + 506, + 422 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 506, + 360 + ], + "score": 1.0, + "content": "Figure 6: Visualization of collaboration between Drone 1 and Drone 2 on CoPerception-UAVs", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 356, + 503, + 377 + ], + "spans": [ + { + "bbox": [ + 102, + 356, + 270, + 377 + ], + "score": 1.0, + "content": "dataset, including spatial confidence map", + "type": "text" + }, + { + "bbox": [ + 270, + 358, + 295, + 373 + ], + "score": 0.87, + "content": "( \\mathbf { C } _ { 1 } ^ { ( 0 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 356, + 366, + 377 + ], + "score": 1.0, + "content": ", selection matrix", + "type": "text" + }, + { + "bbox": [ + 366, + 358, + 399, + 373 + ], + "score": 0.89, + "content": "( \\mathbf { M } _ { 1 2 } ^ { ( 0 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 356, + 441, + 377 + ], + "score": 1.0, + "content": ", message", + "type": "text" + }, + { + "bbox": [ + 441, + 358, + 503, + 373 + ], + "score": 0.9, + "content": "( \\{ \\bar { \\mathbf { R } _ { 2 } ^ { ( 0 ) } } , \\mathcal { Z } _ { 2 1 } ^ { ( 0 ) } \\} )", + "type": "inline_equation" + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 366, + 511, + 393 + ], + "spans": [ + { + "bbox": [ + 102, + 369, + 410, + 391 + ], + "score": 1.0, + "content": "in the communication module, attention weight in the fusion module (W(0)1→1", + "type": "text" + }, + { + "bbox": [ + 378, + 372, + 443, + 388 + ], + "score": 0.9, + "content": "( \\mathbf { \\bar { W } } _ { 1 \\to 1 } ^ { ( 0 ) } , \\mathbf { W } _ { 2 \\to 1 } ^ { ( 0 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 366, + 511, + 393 + ], + "score": 1.0, + "content": "2 2 1, and Drone 1’s", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 384, + 508, + 403 + ], + "spans": [ + { + "bbox": [ + 104, + 384, + 204, + 403 + ], + "score": 1.0, + "content": "detection results before", + "type": "text" + }, + { + "bbox": [ + 204, + 386, + 230, + 401 + ], + "score": 0.88, + "content": "( \\widehat { \\mathcal { O } } _ { 1 } ^ { ( 0 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 384, + 270, + 403 + ], + "score": 1.0, + "content": "and after", + "type": "text" + }, + { + "bbox": [ + 270, + 386, + 296, + 401 + ], + "score": 0.87, + "content": "( \\widehat { \\mathcal { O } } _ { 1 } ^ { ( 1 ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 384, + 508, + 403 + ], + "score": 1.0, + "content": "collaboration. Green and red boxes denote ground-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 505, + 412 + ], + "score": 1.0, + "content": "truth and detection, respectively. The objects occluded by a tall building can be detected through", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 410, + 359, + 424 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 359, + 424 + ], + "score": 1.0, + "content": "transmitting spatially sparse, yet perceptually critical message.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + } + ], + "index": 8.25 + }, + { + "type": "text", + "bbox": [ + 107, + 436, + 505, + 503 + ], + "lines": [], + "index": 16.5, + "bbox_fs": [ + 105, + 437, + 507, + 503 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 514, + 224, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 226, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 226, + 528 + ], + "score": 1.0, + "content": "5.3 Qualitative evaluation", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 530, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 505, + 543 + ], + "score": 1.0, + "content": "Visualization of spatial confidence map. Fig. 6 illustrates how Where2comm is empowered by the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "proposed spatial confidence map. In the scene, Drone 1’s view is occluded by a tall building. With", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 551, + 507, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 507, + 566 + ], + "score": 1.0, + "content": "Drone 2’s help, Drone 1 is able to detect through occlusion. Fig. 6 (a-d) shows Drone 1’s observation,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "spatial confidence map (1), binary selection matrix (2), and ego attention weight (3). Fig. 6 (f-h)", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 106, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "shows Drone 2’s observation and message sent to Drone 1, including the request map (opposite of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 585, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 506, + 598 + ], + "score": 1.0, + "content": "confidence map) and the sparse feature map, achieving efficient communication. Fig. 6 (i) shows", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "the attention weight for Drone 1 to fuse Drone 2’s messages, which is sparse, yet highlights the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "objects’ positions. Fig. 6 (e) and (j) compares the detection results before and after the collaboration", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "with Drone 2. We see that the proposed spatial confidence map contributes to spatially sparse, yet", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 629, + 456, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 456, + 642 + ], + "score": 1.0, + "content": "perceptually critical message, which effectively helps Drone 1 detect occluded objects.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 531, + 507, + 642 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 645, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "Visualization of detection results. Fig. 7 shows that compared to No Collaboration, When2com", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "and DiscoNet, Where2comm is able to achieves more complete and accurate detection results. The", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "reason is that When2com employs a scalar to denote the agent-to-agent attention, which cannot", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "distinguish which spatial area is more informative; DiscoNet employs a MLP-based fusion weight", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "learning, which cannot well capture the complex collaboration attention; while Where2comm can", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "zoom in to critical spatial areas in a cell-level resolution and leverage the spatial confidence map and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 711, + 372, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 372, + 724 + ], + "score": 1.0, + "content": "sensing distances as priors to achieve more comprehensive fusion.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 645, + 505, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 70, + 498, + 171 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 70, + 498, + 171 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 70, + 498, + 171 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 498, + 171 + ], + "score": 0.968, + "type": "image", + "image_path": "3499a8151367e542cac1f416c572dabfcfb8c5a24caf3d31831a2c1b86868797.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 70, + 498, + 103.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 103.66666666666666, + 498, + 137.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 137.33333333333331, + 498, + 170.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "Figure 7: Where2comm qualitatively outperforms When2com and DiscoNet in DAIR-V2X dataset.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "Green and red boxes denote ground-truth and detection, respectively. Yellow and blue denote the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 193, + 377, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 377, + 206 + ], + "score": 1.0, + "content": "point clouds collected from vehicle and infrastructure, respectively.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 106, + 216, + 257, + 288 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 216, + 257, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 216, + 257, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 257, + 288 + ], + "score": 0.952, + "type": "image", + "image_path": "92a0f45a8ee309a48a0fcf0d9162333ae2041e4109919f0079b386d0f640b45d.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 106, + 216, + 257, + 252.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 252.0, + 257, + 288.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 292, + 260, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 293, + 260, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 260, + 304 + ], + "score": 1.0, + "content": "Figure 8: Selection matrix ablation", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 303, + 261, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 261, + 315 + ], + "score": 1.0, + "content": "study. Applying Gaussian filter im-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 315, + 191, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 191, + 326 + ], + "score": 1.0, + "content": "proves performance.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + } + ], + "index": 7.75 + }, + { + "type": "table", + "bbox": [ + 267, + 263, + 509, + 331 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 266, + 213, + 507, + 256 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 266, + 212, + 508, + 225 + ], + "spans": [ + { + "bbox": [ + 266, + 212, + 508, + 225 + ], + "score": 1.0, + "content": "Table 2: Fusion component ablation study. Multi-head atten-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 266, + 224, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 266, + 224, + 506, + 236 + ], + "score": 1.0, + "content": "tion (MHA), sensor positional encoding (SPE) and spatial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 266, + 234, + 508, + 247 + ], + "spans": [ + { + "bbox": [ + 266, + 234, + 508, + 247 + ], + "score": 1.0, + "content": "confidence map (SCM) all improves the performances. Re-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 265, + 245, + 435, + 257 + ], + "spans": [ + { + "bbox": [ + 265, + 245, + 348, + 257 + ], + "score": 1.0, + "content": "sults are reported in", + "type": "text" + }, + { + "bbox": [ + 348, + 245, + 430, + 256 + ], + "score": 0.85, + "content": "\\mathrm { A P @ 0 . 5 0 / A P @ 0 . 7 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 245, + 435, + 257 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "table_body", + "bbox": [ + 267, + 263, + 509, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 267, + 263, + 509, + 331 + ], + "spans": [ + { + "bbox": [ + 267, + 263, + 509, + 331 + ], + "score": 0.98, + "html": "
MHA SPE SCMOPV2VCoPerception-UAVsV2X-Sim
√34.96/13.9263.48/44.23 51.2/45.7
38.75/13.2863.99/44.4657.3/50.8
39.82/16.4364.34/46.8659.1/52.0
√ √√ √ 卜47.30/19.3064.83/47.6259.1/52.2
", + "type": "table", + "image_path": "3b81b83e062a9e7778202536bb4d505e27eb2c0336a7cc12e5387e4bbfeb48a2.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 267, + 263, + 509, + 276.6 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 267, + 276.6, + 509, + 290.20000000000005 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 267, + 290.20000000000005, + 509, + 303.80000000000007 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 267, + 303.80000000000007, + 509, + 317.4000000000001 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 267, + 317.4000000000001, + 509, + 331.0000000000001 + ], + "spans": [], + "index": 19 + } + ] + } + ], + "index": 14.75 + }, + { + "type": "title", + "bbox": [ + 106, + 341, + 200, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 201, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 201, + 354 + ], + "score": 1.0, + "content": "5.4 Ablation studies", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 355, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "score": 1.0, + "content": "Effect of Gaussian filter in perceptually critical area selection. Fig. 8 compares two versions", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 365, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 380 + ], + "score": 1.0, + "content": "of the selection matrix (2) with and without Gaussian filter. We see that applying Gaussian filter", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "score": 1.0, + "content": "improves the overall performance. The reason is that: i) Gaussian filter could help filter out the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 386, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 402 + ], + "score": 1.0, + "content": "outliers in the input map, selecting more robust critical regions; ii) it considers the context, benefiting", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 464, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 464, + 414 + ], + "score": 1.0, + "content": "the independent feature selection at each certain location by providing more information.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 415, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "score": 1.0, + "content": "Effect of components in spatial confidence-aware message fusion. Tab. 2 assesses the effectiveness", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 426, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 506, + 439 + ], + "score": 1.0, + "content": "of the proposed fusion with two priors. We see that: i) per-location multi-head attention (MHA)", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 436, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 258, + 451 + ], + "score": 1.0, + "content": "outperforms the vanilla attention by", + "type": "text" + }, + { + "bbox": [ + 259, + 438, + 291, + 448 + ], + "score": 0.88, + "content": "1 0 . 8 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 436, + 356, + 451 + ], + "score": 1.0, + "content": "on OPV2V on", + "type": "text" + }, + { + "bbox": [ + 356, + 438, + 397, + 448 + ], + "score": 0.85, + "content": "\\mathbf { A P @ 0 . 5 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 436, + 506, + 451 + ], + "score": 1.0, + "content": ", because MHA leverages", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "information from multiple heads, better capturing cross-agent attention; and ii) As two informative", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "priors, both sensing position encoding (SPE) and spatial confidence map (SCM) can consistently", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "improve the performance. Especially, the version with all three designs improves the detection", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 481, + 309, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 171, + 493 + ], + "score": 1.0, + "content": "performance by", + "type": "text" + }, + { + "bbox": [ + 172, + 481, + 204, + 492 + ], + "score": 0.87, + "content": "2 2 . 0 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 481, + 264, + 493 + ], + "score": 1.0, + "content": "on OPV2V on", + "type": "text" + }, + { + "bbox": [ + 265, + 481, + 305, + 492 + ], + "score": 0.77, + "content": "\\mathrm { A P @ } 0 . 5 0", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 481, + 309, + 493 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 108, + 499, + 258, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 497, + 259, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 259, + 515 + ], + "score": 1.0, + "content": "6 Conclusion and limitation", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 506, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "We propose Where2comm, a novel communication-efficient collaborative perception framework. The", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "core idea is to exploit a spatial confidence map at each agent to promote pragmatic compression,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 537, + 507, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 507, + 551 + ], + "score": 1.0, + "content": "assisting agents to decide what to communicate with whom, and whose information to aggregate.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "Each agent offers spatially sparse, yet perceptually critical features to support other agents; meanwhile,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "score": 1.0, + "content": "requests complementary information from others in multi-round communication. Comprehensive", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 570, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 583 + ], + "score": 1.0, + "content": "experiments covering multi-type agents and multi-modality inputs show that Where2comm achieves", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 581, + 454, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 454, + 594 + ], + "score": 1.0, + "content": "far superior trade-off between perception performance and communication bandwidth.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 597, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Limitation and future work. The current work focuses on perceptually critical spatial areas. In", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 608, + 507, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 507, + 621 + ], + "score": 1.0, + "content": "future, we plan to expand a similar idea to the temporal dimension and determine critical time stamps.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "More cost will be reduced by exploring when to communicate. We also expect that more methods on", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 630, + 503, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 503, + 644 + ], + "score": 1.0, + "content": "pragmatic compression and emergent communication could be applied to collaborative perception.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 646, + 506, + 690 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "score": 1.0, + "content": "Acknowledgment. This research is partially supported by the National Key R&D Program of China", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 658, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 506, + 669 + ], + "score": 1.0, + "content": "under Grant 2021ZD0112801, National Natural Science Foundation of China under Grant 62171276,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 668, + 508, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 508, + 680 + ], + "score": 1.0, + "content": "the Science and Technology Commission of Shanghai Municipal under Grant 21511100900, CCF-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 679, + 413, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 413, + 691 + ], + "score": 1.0, + "content": "DiDi GAIA Research Collaboration Plan 202112 and CALT Grant 2021-01.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 742, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 70, + 498, + 171 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 70, + 498, + 171 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 70, + 498, + 171 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 498, + 171 + ], + "score": 0.968, + "type": "image", + "image_path": "3499a8151367e542cac1f416c572dabfcfb8c5a24caf3d31831a2c1b86868797.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 70, + 498, + 103.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 103.66666666666666, + 498, + 137.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 137.33333333333331, + 498, + 170.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "Figure 7: Where2comm qualitatively outperforms When2com and DiscoNet in DAIR-V2X dataset.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "Green and red boxes denote ground-truth and detection, respectively. Yellow and blue denote the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 193, + 377, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 193, + 377, + 206 + ], + "score": 1.0, + "content": "point clouds collected from vehicle and infrastructure, respectively.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 171, + 506, + 206 + ] + }, + { + "type": "image", + "bbox": [ + 106, + 216, + 257, + 288 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 216, + 257, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 216, + 257, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 257, + 288 + ], + "score": 0.952, + "type": "image", + "image_path": "92a0f45a8ee309a48a0fcf0d9162333ae2041e4109919f0079b386d0f640b45d.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 106, + 216, + 257, + 252.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 252.0, + 257, + 288.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 292, + 260, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 293, + 260, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 260, + 304 + ], + "score": 1.0, + "content": "Figure 8: Selection matrix ablation", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 303, + 261, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 261, + 315 + ], + "score": 1.0, + "content": "study. Applying Gaussian filter im-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 315, + 191, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 191, + 326 + ], + "score": 1.0, + "content": "proves performance.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + } + ], + "index": 7.75 + }, + { + "type": "table", + "bbox": [ + 267, + 263, + 509, + 331 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 266, + 213, + 507, + 256 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 266, + 212, + 508, + 225 + ], + "spans": [ + { + "bbox": [ + 266, + 212, + 508, + 225 + ], + "score": 1.0, + "content": "Table 2: Fusion component ablation study. Multi-head atten-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 266, + 224, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 266, + 224, + 506, + 236 + ], + "score": 1.0, + "content": "tion (MHA), sensor positional encoding (SPE) and spatial", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 266, + 234, + 508, + 247 + ], + "spans": [ + { + "bbox": [ + 266, + 234, + 508, + 247 + ], + "score": 1.0, + "content": "confidence map (SCM) all improves the performances. Re-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 265, + 245, + 435, + 257 + ], + "spans": [ + { + "bbox": [ + 265, + 245, + 348, + 257 + ], + "score": 1.0, + "content": "sults are reported in", + "type": "text" + }, + { + "bbox": [ + 348, + 245, + 430, + 256 + ], + "score": 0.85, + "content": "\\mathrm { A P @ 0 . 5 0 / A P @ 0 . 7 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 245, + 435, + 257 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "table_body", + "bbox": [ + 267, + 263, + 509, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 267, + 263, + 509, + 331 + ], + "spans": [ + { + "bbox": [ + 267, + 263, + 509, + 331 + ], + "score": 0.98, + "html": "
MHA SPE SCMOPV2VCoPerception-UAVsV2X-Sim
√34.96/13.9263.48/44.23 51.2/45.7
38.75/13.2863.99/44.4657.3/50.8
39.82/16.4364.34/46.8659.1/52.0
√ √√ √ 卜47.30/19.3064.83/47.6259.1/52.2
", + "type": "table", + "image_path": "3b81b83e062a9e7778202536bb4d505e27eb2c0336a7cc12e5387e4bbfeb48a2.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 267, + 263, + 509, + 276.6 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 267, + 276.6, + 509, + 290.20000000000005 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 267, + 290.20000000000005, + 509, + 303.80000000000007 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 267, + 303.80000000000007, + 509, + 317.4000000000001 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 267, + 317.4000000000001, + 509, + 331.0000000000001 + ], + "spans": [], + "index": 19 + } + ] + } + ], + "index": 14.75 + }, + { + "type": "title", + "bbox": [ + 106, + 341, + 200, + 352 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 201, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 201, + 354 + ], + "score": 1.0, + "content": "5.4 Ablation studies", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 355, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "score": 1.0, + "content": "Effect of Gaussian filter in perceptually critical area selection. Fig. 8 compares two versions", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 365, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 380 + ], + "score": 1.0, + "content": "of the selection matrix (2) with and without Gaussian filter. We see that applying Gaussian filter", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 505, + 390 + ], + "score": 1.0, + "content": "improves the overall performance. The reason is that: i) Gaussian filter could help filter out the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 386, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 386, + 506, + 402 + ], + "score": 1.0, + "content": "outliers in the input map, selecting more robust critical regions; ii) it considers the context, benefiting", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 464, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 464, + 414 + ], + "score": 1.0, + "content": "the independent feature selection at each certain location by providing more information.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 354, + 506, + 414 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 415, + 505, + 492 + ], + "lines": [ + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "score": 1.0, + "content": "Effect of components in spatial confidence-aware message fusion. Tab. 2 assesses the effectiveness", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 426, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 506, + 439 + ], + "score": 1.0, + "content": "of the proposed fusion with two priors. We see that: i) per-location multi-head attention (MHA)", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 436, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 258, + 451 + ], + "score": 1.0, + "content": "outperforms the vanilla attention by", + "type": "text" + }, + { + "bbox": [ + 259, + 438, + 291, + 448 + ], + "score": 0.88, + "content": "1 0 . 8 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 436, + 356, + 451 + ], + "score": 1.0, + "content": "on OPV2V on", + "type": "text" + }, + { + "bbox": [ + 356, + 438, + 397, + 448 + ], + "score": 0.85, + "content": "\\mathbf { A P @ 0 . 5 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 436, + 506, + 451 + ], + "score": 1.0, + "content": ", because MHA leverages", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "information from multiple heads, better capturing cross-agent attention; and ii) As two informative", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "priors, both sensing position encoding (SPE) and spatial confidence map (SCM) can consistently", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "improve the performance. Especially, the version with all three designs improves the detection", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 481, + 309, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 171, + 493 + ], + "score": 1.0, + "content": "performance by", + "type": "text" + }, + { + "bbox": [ + 172, + 481, + 204, + 492 + ], + "score": 0.87, + "content": "2 2 . 0 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 481, + 264, + 493 + ], + "score": 1.0, + "content": "on OPV2V on", + "type": "text" + }, + { + "bbox": [ + 265, + 481, + 305, + 492 + ], + "score": 0.77, + "content": "\\mathrm { A P @ } 0 . 5 0", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 481, + 309, + 493 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 415, + 506, + 493 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 499, + 258, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 497, + 259, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 259, + 515 + ], + "score": 1.0, + "content": "6 Conclusion and limitation", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 515, + 506, + 592 + ], + "lines": [ + { + "bbox": [ + 106, + 515, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 515, + 505, + 528 + ], + "score": 1.0, + "content": "We propose Where2comm, a novel communication-efficient collaborative perception framework. The", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 540 + ], + "score": 1.0, + "content": "core idea is to exploit a spatial confidence map at each agent to promote pragmatic compression,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 537, + 507, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 507, + 551 + ], + "score": 1.0, + "content": "assisting agents to decide what to communicate with whom, and whose information to aggregate.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "Each agent offers spatially sparse, yet perceptually critical features to support other agents; meanwhile,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 506, + 572 + ], + "score": 1.0, + "content": "requests complementary information from others in multi-round communication. Comprehensive", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 570, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 583 + ], + "score": 1.0, + "content": "experiments covering multi-type agents and multi-modality inputs show that Where2comm achieves", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 581, + 454, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 581, + 454, + 594 + ], + "score": 1.0, + "content": "far superior trade-off between perception performance and communication bandwidth.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 515, + 507, + 594 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 597, + 505, + 642 + ], + "lines": [ + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 505, + 609 + ], + "score": 1.0, + "content": "Limitation and future work. The current work focuses on perceptually critical spatial areas. In", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 608, + 507, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 507, + 621 + ], + "score": 1.0, + "content": "future, we plan to expand a similar idea to the temporal dimension and determine critical time stamps.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "More cost will be reduced by exploring when to communicate. We also expect that more methods on", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 630, + 503, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 503, + 644 + ], + "score": 1.0, + "content": "pragmatic compression and emergent communication could be applied to collaborative perception.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 597, + 507, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 646, + 506, + 690 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 506, + 659 + ], + "score": 1.0, + "content": "Acknowledgment. This research is partially supported by the National Key R&D Program of China", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 658, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 506, + 669 + ], + "score": 1.0, + "content": "under Grant 2021ZD0112801, National Natural Science Foundation of China under Grant 62171276,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 668, + 508, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 508, + 680 + ], + "score": 1.0, + "content": "the Science and Technology Commission of Shanghai Municipal under Grant 21511100900, CCF-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 679, + 413, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 679, + 413, + 691 + ], + "score": 1.0, + "content": "DiDi GAIA Research Collaboration Plan 202112 and CALT Grant 2021-01.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 646, + 508, + 691 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 61, + 507, + 728 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 109, + 89, + 507, + 102 + ], + "spans": [ + { + "bbox": [ + 109, + 89, + 507, + 102 + ], + "score": 1.0, + "content": "[1] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 124, + 97, + 506, + 112 + ], + "spans": [ + { + "bbox": [ + 124, + 97, + 506, + 112 + ], + "score": 1.0, + "content": "V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 109, + 319, + 122 + ], + "spans": [ + { + "bbox": [ + 124, + 109, + 319, + 122 + ], + "score": 1.0, + "content": "on Computer Vision, pages 605–621. Springer, 2020.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 126, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 109, + 126, + 506, + 138 + ], + "score": 1.0, + "content": "[2] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 136, + 507, + 150 + ], + "spans": [ + { + "bbox": [ + 124, + 136, + 507, + 150 + ], + "score": 1.0, + "content": "collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 146, + 150, + 159 + ], + "spans": [ + { + "bbox": [ + 125, + 146, + 150, + 159 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 109, + 163, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 109, + 163, + 506, + 177 + ], + "score": 1.0, + "content": "[3] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl K. Wellington. 3d point cloud", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 173, + 507, + 187 + ], + "spans": [ + { + "bbox": [ + 125, + 173, + 507, + 187 + ], + "score": 1.0, + "content": "processing and learning for autonomous driving: Impacting map creation, localization, and perception.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 184, + 316, + 196 + ], + "spans": [ + { + "bbox": [ + 125, + 184, + 316, + 196 + ], + "score": 1.0, + "content": "IEEE Signal Processing Magazine, 38:68–86, 2021.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 109, + 201, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 109, + 201, + 506, + 215 + ], + "score": 1.0, + "content": "[4] Zhi Li, Ali Vatankhah Barenji, Jiazhi Jiang, Ray Y Zhong, and Gangyan Xu. A mechanism for scheduling", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 124, + 209, + 507, + 225 + ], + "spans": [ + { + "bbox": [ + 124, + 209, + 507, + 225 + ], + "score": 1.0, + "content": "multi robot intelligent warehouse system face with dynamic demand. Journal of Intelligent Manufacturing,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 221, + 208, + 233 + ], + "spans": [ + { + "bbox": [ + 126, + 221, + 208, + 233 + ], + "score": 1.0, + "content": "31(2):469–480, 2020.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 110, + 239, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 110, + 239, + 505, + 250 + ], + "score": 1.0, + "content": "[5] Michela Zaccaria, Mikhail Giorgini, Riccardo Monica, and Jacopo Aleotti. Multi-robot multiple camera", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 124, + 248, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 124, + 248, + 506, + 262 + ], + "score": 1.0, + "content": "people detection and tracking in automated warehouses. In 2021 IEEE 19th International Conference on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 258, + 331, + 272 + ], + "spans": [ + { + "bbox": [ + 125, + 258, + 331, + 272 + ], + "score": 1.0, + "content": "Industrial Informatics (INDIN), pages 1–6. IEEE, 2021.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 110, + 276, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 110, + 276, + 506, + 288 + ], + "score": 1.0, + "content": "[6] Jürgen Scherer, Saeed Yahyanejad, Samira Hayat, Evsen Yanmaz, Torsten Andre, Asif Khan, Vladimir", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 126, + 286, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 126, + 286, + 505, + 297 + ], + "score": 1.0, + "content": "Vukadinovic, Christian Bettstetter, Hermann Hellwagner, and Bernhard Rinner. An autonomous multi-uav", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 124, + 294, + 507, + 309 + ], + "spans": [ + { + "bbox": [ + 124, + 294, + 507, + 309 + ], + "score": 1.0, + "content": "system for search and rescue. In Proceedings of the First Workshop on Micro Aerial Vehicle Networks,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 125, + 306, + 358, + 319 + ], + "spans": [ + { + "bbox": [ + 125, + 306, + 358, + 319 + ], + "score": 1.0, + "content": "Systems, and Applications for Civilian Use, pages 33–38, 2015.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 110, + 324, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 110, + 324, + 505, + 335 + ], + "score": 1.0, + "content": "[7] Ebtehal Turki Alotaibi, Shahad Saleh Alqefari, and Anis Koubaa. Lsar: Multi-uav collaboration for search", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 125, + 333, + 337, + 345 + ], + "spans": [ + { + "bbox": [ + 125, + 333, + 337, + 345 + ], + "score": 1.0, + "content": "and rescue missions. IEEE Access, 7:55817–55832, 2019.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 109, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 109, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "[8] Yue Hu, Shaoheng Fang, Weidi Xie, and Siheng Chen. Aerial monocular 3d object detection. arXiv", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 123, + 360, + 251, + 373 + ], + "spans": [ + { + "bbox": [ + 123, + 360, + 251, + 373 + ], + "score": 1.0, + "content": "preprint arXiv:2208.03974, 2022.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 109, + 377, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 109, + 377, + 506, + 391 + ], + "score": 1.0, + "content": "[9] Yiming Li, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2X-Sim: A virtual", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 124, + 387, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 124, + 387, + 506, + 401 + ], + "score": 1.0, + "content": "collaborative perception dataset for autonomous driving. IEEE Robotics and Automation Letters, 7, 2022.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "score": 1.0, + "content": "[10] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Liu, and Jiaqi Ma. OPV2V: An open benchmark", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 124, + 414, + 469, + 428 + ], + "spans": [ + { + "bbox": [ + 124, + 414, + 469, + 428 + ], + "score": 1.0, + "content": "dataset and fusion pipeline for perception with vehicle-to-vehicle communication. ICRA, 2022.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 432, + 507, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 507, + 446 + ], + "score": 1.0, + "content": "[11] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 124, + 442, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 124, + 442, + 506, + 456 + ], + "score": 1.0, + "content": "Xing Hu, Jirui Yuan, et al. DAIR-V2X: A large-scale dataset for vehicle-infrastructure cooperative 3d", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 125, + 453, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 125, + 453, + 506, + 465 + ], + "score": 1.0, + "content": "object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 124, + 463, + 182, + 474 + ], + "spans": [ + { + "bbox": [ + 124, + 463, + 182, + 474 + ], + "score": 1.0, + "content": "(CVPR), 2022.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "[12] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 125, + 490, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 125, + 490, + 506, + 504 + ], + "score": 1.0, + "content": "communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 124, + 499, + 290, + 513 + ], + "spans": [ + { + "bbox": [ + 124, + 499, + 290, + 513 + ], + "score": 1.0, + "content": "pattern recognition, pages 4106–4115, 2020.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "[13] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com:", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 125, + 526, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 125, + 526, + 506, + 540 + ], + "score": 1.0, + "content": "Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 537, + 376, + 550 + ], + "spans": [ + { + "bbox": [ + 124, + 537, + 376, + 550 + ], + "score": 1.0, + "content": "on Robotics and Automation (ICRA), pages 6876–6883. IEEE, 2020.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "[14] Yang Zhou, Jiuhong Xiao, Yue Zhou, and Giuseppe Loianno. Multi-robot collaborative perception with", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 124, + 564, + 379, + 577 + ], + "spans": [ + { + "bbox": [ + 124, + 564, + 379, + 577 + ], + "score": 1.0, + "content": "graph neural networks. IEEE Robotics and Automation Letters, 2022.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "score": 1.0, + "content": "[15] Eduardo Arnold, Mehrdad Dianati, and Robert de Temple. Cooperative perception for 3d object detection", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 125, + 592, + 507, + 604 + ], + "spans": [ + { + "bbox": [ + 125, + 592, + 507, + 604 + ], + "score": 1.0, + "content": "in driving scenarios using infrastructure sensors. IEEE Transactions on Intelligent Transportation Systems,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 603, + 205, + 614 + ], + "spans": [ + { + "bbox": [ + 126, + 603, + 205, + 614 + ], + "score": 1.0, + "content": "23:1852–1864, 2022.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 618, + 507, + 632 + ], + "spans": [ + { + "bbox": [ + 104, + 618, + 507, + 632 + ], + "score": 1.0, + "content": "[16] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 126, + 630, + 177, + 641 + ], + "spans": [ + { + "bbox": [ + 126, + 630, + 177, + 641 + ], + "score": 1.0, + "content": "ECCV, 2022.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 646, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 506, + 659 + ], + "score": 1.0, + "content": "[17] Yiming Li, Juexiao Zhang, Dekun Ma, Yue Wang, and Chen Feng. Multi-robot scene completion: Towards", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 124, + 656, + 474, + 669 + ], + "spans": [ + { + "bbox": [ + 124, + 656, + 474, + 669 + ], + "score": 1.0, + "content": "task-agnostic collaborative perception. In Conference on Robot Learning (CoRL). PMLR, 2022.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "[18] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. CoBEVT: Cooperative", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 126, + 684, + 408, + 696 + ], + "spans": [ + { + "bbox": [ + 126, + 684, + 408, + 696 + ], + "score": 1.0, + "content": "bird’s eye view semantic segmentation with sparse transformers. CoRL, 2022.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 104, + 700, + 505, + 715 + ], + "spans": [ + { + "bbox": [ + 104, + 700, + 505, + 715 + ], + "score": 1.0, + "content": "[19] Sanbao Su, Yiming Li, Sihong He, Songyang Han, Chen Feng, Caiwen Ding, and Fei Miao. Uncertainty", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 124, + 710, + 354, + 725 + ], + "spans": [ + { + "bbox": [ + 124, + 710, + 354, + 725 + ], + "score": 1.0, + "content": "quantification of collaborative detection for self-driving, 2022.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 25 + } + ], + "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": "list", + "bbox": [ + 106, + 61, + 507, + 728 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 165, + 86 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 89, + 507, + 102 + ], + "spans": [ + { + "bbox": [ + 109, + 89, + 507, + 102 + ], + "score": 1.0, + "content": "[1] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun.", + "type": "text" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 97, + 506, + 112 + ], + "spans": [ + { + "bbox": [ + 124, + 97, + 506, + 112 + ], + "score": 1.0, + "content": "V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In European Conference", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 109, + 319, + 122 + ], + "spans": [ + { + "bbox": [ + 124, + 109, + 319, + 122 + ], + "score": 1.0, + "content": "on Computer Vision, pages 605–621. Springer, 2020.", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 126, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 109, + 126, + 506, + 138 + ], + "score": 1.0, + "content": "[2] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 136, + 507, + 150 + ], + "spans": [ + { + "bbox": [ + 124, + 136, + 507, + 150 + ], + "score": 1.0, + "content": "collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 146, + 150, + 159 + ], + "spans": [ + { + "bbox": [ + 125, + 146, + 150, + 159 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 163, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 109, + 163, + 506, + 177 + ], + "score": 1.0, + "content": "[3] Siheng Chen, Baoan Liu, Chen Feng, Carlos Vallespi-Gonzalez, and Carl K. Wellington. 3d point cloud", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 173, + 507, + 187 + ], + "spans": [ + { + "bbox": [ + 125, + 173, + 507, + 187 + ], + "score": 1.0, + "content": "processing and learning for autonomous driving: Impacting map creation, localization, and perception.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 184, + 316, + 196 + ], + "spans": [ + { + "bbox": [ + 125, + 184, + 316, + 196 + ], + "score": 1.0, + "content": "IEEE Signal Processing Magazine, 38:68–86, 2021.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 201, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 109, + 201, + 506, + 215 + ], + "score": 1.0, + "content": "[4] Zhi Li, Ali Vatankhah Barenji, Jiazhi Jiang, Ray Y Zhong, and Gangyan Xu. A mechanism for scheduling", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 209, + 507, + 225 + ], + "spans": [ + { + "bbox": [ + 124, + 209, + 507, + 225 + ], + "score": 1.0, + "content": "multi robot intelligent warehouse system face with dynamic demand. Journal of Intelligent Manufacturing,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 221, + 208, + 233 + ], + "spans": [ + { + "bbox": [ + 126, + 221, + 208, + 233 + ], + "score": 1.0, + "content": "31(2):469–480, 2020.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 239, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 110, + 239, + 505, + 250 + ], + "score": 1.0, + "content": "[5] Michela Zaccaria, Mikhail Giorgini, Riccardo Monica, and Jacopo Aleotti. Multi-robot multiple camera", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 248, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 124, + 248, + 506, + 262 + ], + "score": 1.0, + "content": "people detection and tracking in automated warehouses. In 2021 IEEE 19th International Conference on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 258, + 331, + 272 + ], + "spans": [ + { + "bbox": [ + 125, + 258, + 331, + 272 + ], + "score": 1.0, + "content": "Industrial Informatics (INDIN), pages 1–6. IEEE, 2021.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 276, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 110, + 276, + 506, + 288 + ], + "score": 1.0, + "content": "[6] Jürgen Scherer, Saeed Yahyanejad, Samira Hayat, Evsen Yanmaz, Torsten Andre, Asif Khan, Vladimir", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 286, + 505, + 297 + ], + "spans": [ + { + "bbox": [ + 126, + 286, + 505, + 297 + ], + "score": 1.0, + "content": "Vukadinovic, Christian Bettstetter, Hermann Hellwagner, and Bernhard Rinner. An autonomous multi-uav", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 124, + 294, + 507, + 309 + ], + "spans": [ + { + "bbox": [ + 124, + 294, + 507, + 309 + ], + "score": 1.0, + "content": "system for search and rescue. In Proceedings of the First Workshop on Micro Aerial Vehicle Networks,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 125, + 306, + 358, + 319 + ], + "spans": [ + { + "bbox": [ + 125, + 306, + 358, + 319 + ], + "score": 1.0, + "content": "Systems, and Applications for Civilian Use, pages 33–38, 2015.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 324, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 110, + 324, + 505, + 335 + ], + "score": 1.0, + "content": "[7] Ebtehal Turki Alotaibi, Shahad Saleh Alqefari, and Anis Koubaa. Lsar: Multi-uav collaboration for search", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 333, + 337, + 345 + ], + "spans": [ + { + "bbox": [ + 125, + 333, + 337, + 345 + ], + "score": 1.0, + "content": "and rescue missions. IEEE Access, 7:55817–55832, 2019.", + "type": "text" + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 350, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 109, + 350, + 506, + 363 + ], + "score": 1.0, + "content": "[8] Yue Hu, Shaoheng Fang, Weidi Xie, and Siheng Chen. Aerial monocular 3d object detection. arXiv", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 123, + 360, + 251, + 373 + ], + "spans": [ + { + "bbox": [ + 123, + 360, + 251, + 373 + ], + "score": 1.0, + "content": "preprint arXiv:2208.03974, 2022.", + "type": "text" + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 109, + 377, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 109, + 377, + 506, + 391 + ], + "score": 1.0, + "content": "[9] Yiming Li, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2X-Sim: A virtual", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 387, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 124, + 387, + 506, + 401 + ], + "score": 1.0, + "content": "collaborative perception dataset for autonomous driving. IEEE Robotics and Automation Letters, 7, 2022.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 417 + ], + "score": 1.0, + "content": "[10] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Liu, and Jiaqi Ma. OPV2V: An open benchmark", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 414, + 469, + 428 + ], + "spans": [ + { + "bbox": [ + 124, + 414, + 469, + 428 + ], + "score": 1.0, + "content": "dataset and fusion pipeline for perception with vehicle-to-vehicle communication. ICRA, 2022.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 432, + 507, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 507, + 446 + ], + "score": 1.0, + "content": "[11] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li,", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 442, + 506, + 456 + ], + "spans": [ + { + "bbox": [ + 124, + 442, + 506, + 456 + ], + "score": 1.0, + "content": "Xing Hu, Jirui Yuan, et al. DAIR-V2X: A large-scale dataset for vehicle-infrastructure cooperative 3d", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 125, + 453, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 125, + 453, + 506, + 465 + ], + "score": 1.0, + "content": "object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 124, + 463, + 182, + 474 + ], + "spans": [ + { + "bbox": [ + 124, + 463, + 182, + 474 + ], + "score": 1.0, + "content": "(CVPR), 2022.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "[12] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 490, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 125, + 490, + 506, + 504 + ], + "score": 1.0, + "content": "communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 124, + 499, + 290, + 513 + ], + "spans": [ + { + "bbox": [ + 124, + 499, + 290, + 513 + ], + "score": 1.0, + "content": "pattern recognition, pages 4106–4115, 2020.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 530 + ], + "score": 1.0, + "content": "[13] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com:", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 526, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 125, + 526, + 506, + 540 + ], + "score": 1.0, + "content": "Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 124, + 537, + 376, + 550 + ], + "spans": [ + { + "bbox": [ + 124, + 537, + 376, + 550 + ], + "score": 1.0, + "content": "on Robotics and Automation (ICRA), pages 6876–6883. IEEE, 2020.", + "type": "text" + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "[14] Yang Zhou, Jiuhong Xiao, Yue Zhou, and Giuseppe Loianno. Multi-robot collaborative perception with", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 564, + 379, + 577 + ], + "spans": [ + { + "bbox": [ + 124, + 564, + 379, + 577 + ], + "score": 1.0, + "content": "graph neural networks. IEEE Robotics and Automation Letters, 2022.", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 594 + ], + "score": 1.0, + "content": "[15] Eduardo Arnold, Mehrdad Dianati, and Robert de Temple. Cooperative perception for 3d object detection", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 592, + 507, + 604 + ], + "spans": [ + { + "bbox": [ + 125, + 592, + 507, + 604 + ], + "score": 1.0, + "content": "in driving scenarios using infrastructure sensors. IEEE Transactions on Intelligent Transportation Systems,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 603, + 205, + 614 + ], + "spans": [ + { + "bbox": [ + 126, + 603, + 205, + 614 + ], + "score": 1.0, + "content": "23:1852–1864, 2022.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 618, + 507, + 632 + ], + "spans": [ + { + "bbox": [ + 104, + 618, + 507, + 632 + ], + "score": 1.0, + "content": "[16] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception.", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 630, + 177, + 641 + ], + "spans": [ + { + "bbox": [ + 126, + 630, + 177, + 641 + ], + "score": 1.0, + "content": "ECCV, 2022.", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 646, + 506, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 506, + 659 + ], + "score": 1.0, + "content": "[17] Yiming Li, Juexiao Zhang, Dekun Ma, Yue Wang, and Chen Feng. Multi-robot scene completion: Towards", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 656, + 474, + 669 + ], + "spans": [ + { + "bbox": [ + 124, + 656, + 474, + 669 + ], + "score": 1.0, + "content": "task-agnostic collaborative perception. In Conference on Robot Learning (CoRL). PMLR, 2022.", + "type": "text" + } + ], + "index": 46, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 106, + 674, + 505, + 686 + ], + "score": 1.0, + "content": "[18] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. CoBEVT: Cooperative", + "type": "text" + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 684, + 408, + 696 + ], + "spans": [ + { + "bbox": [ + 126, + 684, + 408, + 696 + ], + "score": 1.0, + "content": "bird’s eye view semantic segmentation with sparse transformers. CoRL, 2022.", + "type": "text" + } + ], + "index": 48, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 700, + 505, + 715 + ], + "spans": [ + { + "bbox": [ + 104, + 700, + 505, + 715 + ], + "score": 1.0, + "content": "[19] Sanbao Su, Yiming Li, Sihong He, Songyang Han, Chen Feng, Caiwen Ding, and Fei Miao. Uncertainty", + "type": "text" + } + ], + "index": 49, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 710, + 354, + 725 + ], + "spans": [ + { + "bbox": [ + 124, + 710, + 354, + 725 + ], + "score": 1.0, + "content": "quantification of collaborative detection for self-driving, 2022.", + "type": "text" + } + ], + "index": 50, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 74, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 106, + 74, + 506, + 87 + ], + "score": 1.0, + "content": "[20] Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 83, + 341, + 97 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 341, + 97 + ], + "score": 1.0, + "content": "multiagent cooperative and competitive tasks. ICLR, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 102, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 107, + 102, + 505, + 114 + ], + "score": 1.0, + "content": "[21] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 111, + 397, + 125 + ], + "spans": [ + { + "bbox": [ + 125, + 111, + 397, + 125 + ], + "score": 1.0, + "content": "tenth international conference on machine learning, pages 330–337, 1993.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 129, + 506, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 506, + 142 + ], + "score": 1.0, + "content": "[22] Faisal Qureshi and Demetri Terzopoulos. Smart camera networks in virtual reality. Proceedings of the", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 139, + 244, + 151 + ], + "spans": [ + { + "bbox": [ + 125, + 139, + 244, + 151 + ], + "score": 1.0, + "content": "IEEE, 96(10):1640–1656, 2008.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 157, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 157, + 506, + 171 + ], + "score": 1.0, + "content": "[23] Yiming Li, Bir Bhanu, and Wei Lin. Auction protocol for camera active control. In 2010 IEEE International", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 167, + 365, + 180 + ], + "spans": [ + { + "bbox": [ + 125, + 167, + 365, + 180 + ], + "score": 1.0, + "content": "Conference on Image Processing, pages 4325–4328. IEEE, 2010.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 185, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 506, + 199 + ], + "score": 1.0, + "content": "[24] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 124, + 195, + 351, + 208 + ], + "spans": [ + { + "bbox": [ + 124, + 195, + 351, + 208 + ], + "score": 1.0, + "content": "Advances in neural information processing systems, 29, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 506, + 226 + ], + "score": 1.0, + "content": "[25] Yedid Hoshen. Vain: Attentional multi-agent predictive modeling. Advances in Neural Information", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 223, + 237, + 235 + ], + "spans": [ + { + "bbox": [ + 125, + 223, + 237, + 235 + ], + "score": 1.0, + "content": "Processing Systems, 30, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 241, + 507, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 507, + 254 + ], + "score": 1.0, + "content": "[26] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2X-ViT:", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 250, + 429, + 263 + ], + "spans": [ + { + "bbox": [ + 125, + 250, + 429, + 263 + ], + "score": 1.0, + "content": "Vehicle-to-everything cooperative perception with vision transformer. ECCV, 2022.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 270, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 506, + 281 + ], + "score": 1.0, + "content": "[27] Y Yuan and M Sester. Comap: A synthetic dataset for collective multi-agent perception of autonomous", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 279, + 507, + 293 + ], + "spans": [ + { + "bbox": [ + 125, + 279, + 507, + 293 + ], + "score": 1.0, + "content": "driving. The International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences,", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 289, + 196, + 301 + ], + "spans": [ + { + "bbox": [ + 125, + 289, + 196, + 301 + ], + "score": 1.0, + "content": "43:255–263, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "score": 1.0, + "content": "[28] Xingyi Zhou, Dequan Wang, and Philipp Krähenbühl. Objects as points. In arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 317, + 219, + 329 + ], + "spans": [ + { + "bbox": [ + 125, + 317, + 219, + 329 + ], + "score": 1.0, + "content": "arXiv:1904.07850, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "[29] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 343, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 125, + 343, + 482, + 357 + ], + "score": 1.0, + "content": "Proceedings of the 26th annual international conference on machine learning, pages 41–48, 2009.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "score": 1.0, + "content": "[30] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 373, + 433, + 385 + ], + "spans": [ + { + "bbox": [ + 126, + 373, + 433, + 385 + ], + "score": 1.0, + "content": "urban driving simulator. In Conference on robot learning, pages 1–16. PMLR, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 389, + 507, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 507, + 405 + ], + "score": 1.0, + "content": "[31] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical depth distribution network", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 400, + 303, + 413 + ], + "spans": [ + { + "bbox": [ + 125, + 400, + 303, + 413 + ], + "score": 1.0, + "content": "for monocular 3d object detection. CVPR, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "score": 1.0, + "content": "[32] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 126, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "applications of sumo-simulation of urban mobility. International journal on advances in systems and", + "type": "text", + "cross_page": true + } + ], + "index": 26 + }, + { + "bbox": [ + 125, + 439, + 235, + 451 + ], + "spans": [ + { + "bbox": [ + 125, + 439, + 235, + 451 + ], + "score": 1.0, + "content": "measurements, 5(3&4), 2012.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 455, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 470 + ], + "score": 1.0, + "content": "[33] Pengxiang Wu, Siheng Chen, and Dimitris N. Metaxas. Motionnet: Joint perception and motion prediction", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 125, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "for autonomous driving based on bird’s eye view maps. 2020 IEEE/CVF Conference on Computer Vision", + "type": "text", + "cross_page": true + } + ], + "index": 29 + }, + { + "bbox": [ + 124, + 476, + 348, + 488 + ], + "spans": [ + { + "bbox": [ + 124, + 476, + 348, + 488 + ], + "score": 1.0, + "content": "and Pattern Recognition (CVPR), pages 11382–11392, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 494, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 506, + 507 + ], + "score": 1.0, + "content": "[34] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 504, + 484, + 517 + ], + "spans": [ + { + "bbox": [ + 125, + 504, + 484, + 517 + ], + "score": 1.0, + "content": "simulation for autonomous vehicles. In Field and service robotics, pages 621–635. Springer, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 506, + 535 + ], + "score": 1.0, + "content": "[35] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars:", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 125, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "Fast encoders for object detection from point clouds. 2019 IEEE/CVF Conference on Computer Vision and", + "type": "text", + "cross_page": true + } + ], + "index": 34 + }, + { + "bbox": [ + 125, + 541, + 333, + 555 + ], + "spans": [ + { + "bbox": [ + 125, + 541, + 333, + 555 + ], + "score": 1.0, + "content": "Pattern Recognition (CVPR), pages 12689–12697, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_end_line": true + } + ], + "index": 25, + "bbox_fs": [ + 104, + 70, + 507, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 71, + 507, + 553 + ], + "lines": [ + { + "bbox": [ + 106, + 74, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 106, + 74, + 506, + 87 + ], + "score": 1.0, + "content": "[20] Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 83, + 341, + 97 + ], + "spans": [ + { + "bbox": [ + 125, + 83, + 341, + 97 + ], + "score": 1.0, + "content": "multiagent cooperative and competitive tasks. ICLR, 2019.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 102, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 107, + 102, + 505, + 114 + ], + "score": 1.0, + "content": "[21] Ming Tan. Multi-agent reinforcement learning: Independent vs. cooperative agents. In Proceedings of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 111, + 397, + 125 + ], + "spans": [ + { + "bbox": [ + 125, + 111, + 397, + 125 + ], + "score": 1.0, + "content": "tenth international conference on machine learning, pages 330–337, 1993.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 129, + 506, + 142 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 506, + 142 + ], + "score": 1.0, + "content": "[22] Faisal Qureshi and Demetri Terzopoulos. Smart camera networks in virtual reality. Proceedings of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 139, + 244, + 151 + ], + "spans": [ + { + "bbox": [ + 125, + 139, + 244, + 151 + ], + "score": 1.0, + "content": "IEEE, 96(10):1640–1656, 2008.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 157, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 157, + 506, + 171 + ], + "score": 1.0, + "content": "[23] Yiming Li, Bir Bhanu, and Wei Lin. Auction protocol for camera active control. In 2010 IEEE International", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 167, + 365, + 180 + ], + "spans": [ + { + "bbox": [ + 125, + 167, + 365, + 180 + ], + "score": 1.0, + "content": "Conference on Image Processing, pages 4325–4328. IEEE, 2010.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 185, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 185, + 506, + 199 + ], + "score": 1.0, + "content": "[24] Sainbayar Sukhbaatar, Rob Fergus, et al. Learning multiagent communication with backpropagation.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 124, + 195, + 351, + 208 + ], + "spans": [ + { + "bbox": [ + 124, + 195, + 351, + 208 + ], + "score": 1.0, + "content": "Advances in neural information processing systems, 29, 2016.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 213, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 506, + 226 + ], + "score": 1.0, + "content": "[25] Yedid Hoshen. Vain: Attentional multi-agent predictive modeling. Advances in Neural Information", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 125, + 223, + 237, + 235 + ], + "spans": [ + { + "bbox": [ + 125, + 223, + 237, + 235 + ], + "score": 1.0, + "content": "Processing Systems, 30, 2017.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 241, + 507, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 507, + 254 + ], + "score": 1.0, + "content": "[26] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2X-ViT:", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 125, + 250, + 429, + 263 + ], + "spans": [ + { + "bbox": [ + 125, + 250, + 429, + 263 + ], + "score": 1.0, + "content": "Vehicle-to-everything cooperative perception with vision transformer. ECCV, 2022.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 270, + 506, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 506, + 281 + ], + "score": 1.0, + "content": "[27] Y Yuan and M Sester. Comap: A synthetic dataset for collective multi-agent perception of autonomous", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 125, + 279, + 507, + 293 + ], + "spans": [ + { + "bbox": [ + 125, + 279, + 507, + 293 + ], + "score": 1.0, + "content": "driving. The International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 289, + 196, + 301 + ], + "spans": [ + { + "bbox": [ + 125, + 289, + 196, + 301 + ], + "score": 1.0, + "content": "43:255–263, 2021.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 320 + ], + "score": 1.0, + "content": "[28] Xingyi Zhou, Dequan Wang, and Philipp Krähenbühl. Objects as points. In arXiv preprint", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 125, + 317, + 219, + 329 + ], + "spans": [ + { + "bbox": [ + 125, + 317, + 219, + 329 + ], + "score": 1.0, + "content": "arXiv:1904.07850, 2019.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 506, + 347 + ], + "score": 1.0, + "content": "[29] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 125, + 343, + 482, + 357 + ], + "spans": [ + { + "bbox": [ + 125, + 343, + 482, + 357 + ], + "score": 1.0, + "content": "Proceedings of the 26th annual international conference on machine learning, pages 41–48, 2009.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "score": 1.0, + "content": "[30] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 373, + 433, + 385 + ], + "spans": [ + { + "bbox": [ + 126, + 373, + 433, + 385 + ], + "score": 1.0, + "content": "urban driving simulator. In Conference on robot learning, pages 1–16. PMLR, 2017.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 389, + 507, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 507, + 405 + ], + "score": 1.0, + "content": "[31] Cody Reading, Ali Harakeh, Julia Chae, and Steven L. Waslander. Categorical depth distribution network", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 125, + 400, + 303, + 413 + ], + "spans": [ + { + "bbox": [ + 125, + 400, + 303, + 413 + ], + "score": 1.0, + "content": "for monocular 3d object detection. CVPR, 2021.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 432 + ], + "score": 1.0, + "content": "[32] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 126, + 428, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 126, + 428, + 506, + 442 + ], + "score": 1.0, + "content": "applications of sumo-simulation of urban mobility. International journal on advances in systems and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 125, + 439, + 235, + 451 + ], + "spans": [ + { + "bbox": [ + 125, + 439, + 235, + 451 + ], + "score": 1.0, + "content": "measurements, 5(3&4), 2012.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 455, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 506, + 470 + ], + "score": 1.0, + "content": "[33] Pengxiang Wu, Siheng Chen, and Dimitris N. Metaxas. Motionnet: Joint perception and motion prediction", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 125, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 125, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "for autonomous driving based on bird’s eye view maps. 2020 IEEE/CVF Conference on Computer Vision", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 124, + 476, + 348, + 488 + ], + "spans": [ + { + "bbox": [ + 124, + 476, + 348, + 488 + ], + "score": 1.0, + "content": "and Pattern Recognition (CVPR), pages 11382–11392, 2020.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 494, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 506, + 507 + ], + "score": 1.0, + "content": "[34] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 125, + 504, + 484, + 517 + ], + "spans": [ + { + "bbox": [ + 125, + 504, + 484, + 517 + ], + "score": 1.0, + "content": "simulation for autonomous vehicles. In Field and service robotics, pages 621–635. Springer, 2018.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 522, + 506, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 506, + 535 + ], + "score": 1.0, + "content": "[35] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars:", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 125, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 125, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "Fast encoders for object detection from point clouds. 2019 IEEE/CVF Conference on Computer Vision and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 125, + 541, + 333, + 555 + ], + "spans": [ + { + "bbox": [ + 125, + 541, + 333, + 555 + ], + "score": 1.0, + "content": "Pattern Recognition (CVPR), pages 12689–12697, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 17.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 71, + 507, + 553 + ], + "lines": [], + "index": 17.5, + "bbox_fs": [ + 105, + 74, + 507, + 555 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 71, + 156, + 84 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 131, + 92, + 208, + 104 + ], + "lines": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "spans": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 146, + 107, + 505, + 180 + ], + "lines": [ + { + "bbox": [ + 145, + 106, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 145, + 106, + 505, + 120 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 118, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 161, + 118, + 288, + 131 + ], + "score": 1.0, + "content": "contributions and scope? [Yes]", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 131, + 376, + 144 + ], + "spans": [ + { + "bbox": [ + 145, + 131, + 376, + 144 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes]", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 145, + 144, + 470, + 156 + ], + "spans": [ + { + "bbox": [ + 145, + 144, + 470, + 156 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [N/A]", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 145, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 145, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 161, + 167, + 214, + 181 + ], + "spans": [ + { + "bbox": [ + 161, + 167, + 214, + 181 + ], + "score": 1.0, + "content": "them? [Yes]", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 131, + 183, + 302, + 194 + ], + "lines": [ + { + "bbox": [ + 129, + 182, + 304, + 196 + ], + "spans": [ + { + "bbox": [ + 129, + 182, + 304, + 196 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 146, + 197, + 452, + 222 + ], + "lines": [ + { + "bbox": [ + 145, + 197, + 453, + 210 + ], + "spans": [ + { + "bbox": [ + 145, + 197, + 453, + 210 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 145, + 210, + 424, + 223 + ], + "spans": [ + { + "bbox": [ + 145, + 210, + 424, + 223 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 130, + 226, + 241, + 237 + ], + "lines": [ + { + "bbox": [ + 128, + 223, + 243, + 239 + ], + "spans": [ + { + "bbox": [ + 128, + 223, + 243, + 239 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 146, + 241, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 145, + 240, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 145, + 240, + 507, + 253 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 162, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 162, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "mental results (either in the supplemental material or as a URL)? [Yes] See Section 5.1", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 162, + 262, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 162, + 262, + 287, + 274 + ], + "score": 1.0, + "content": "and the supplemental material.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 146, + 274, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 146, + 274, + 505, + 288 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 161, + 286, + 434, + 298 + ], + "spans": [ + { + "bbox": [ + 161, + 286, + 434, + 298 + ], + "score": 1.0, + "content": "were chosen)? [Yes] See Section 5.1 and the supplemental material.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 147, + 298, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 147, + 298, + 506, + 312 + ], + "score": 1.0, + "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 161, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 161, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "ments multiple times)? [No] We have not repeated experiments many times to get error", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 161, + 320, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 161, + 320, + 506, + 334 + ], + "score": 1.0, + "content": "bars since experiments of 3d object detection on large scale datasets is time-consuming.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 146, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 146, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g.,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 161, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 161, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "type of GPUs, internal cluster, or cloud provider)? [Yes] See Section 5.1 and the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 161, + 356, + 255, + 367 + ], + "spans": [ + { + "bbox": [ + 161, + 356, + 255, + 367 + ], + "score": 1.0, + "content": "supplemental material.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 133, + 370, + 505, + 382 + ], + "lines": [ + { + "bbox": [ + 129, + 368, + 507, + 385 + ], + "spans": [ + { + "bbox": [ + 129, + 368, + 507, + 385 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 146, + 385, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 145, + 385, + 491, + 398 + ], + "spans": [ + { + "bbox": [ + 145, + 385, + 491, + 398 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] See Section 5.1.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 146, + 398, + 423, + 410 + ], + "spans": [ + { + "bbox": [ + 146, + 398, + 423, + 410 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [Yes] See Section 5.1.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 146, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 146, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 162, + 421, + 351, + 434 + ], + "spans": [ + { + "bbox": [ + 162, + 421, + 351, + 434 + ], + "score": 1.0, + "content": "See Section 5.1 and the supplemental material.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 146, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 146, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 162, + 447, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 162, + 447, + 505, + 458 + ], + "score": 1.0, + "content": "using/curating? [N/A] We generate the data using the open-sourced tool and the owners", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 162, + 457, + 365, + 469 + ], + "spans": [ + { + "bbox": [ + 162, + 457, + 365, + 469 + ], + "score": 1.0, + "content": "consent to all the public use for research purposes.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 146, + 469, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 146, + 469, + 506, + 483 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 162, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 162, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "information or offensive content? [N/A] Our data is synthesized using the open-sourced", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 162, + 491, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 162, + 491, + 506, + 503 + ], + "score": 1.0, + "content": "tool, so there is no real-world personally identifiable information, nor offensive content.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 131, + 506, + 432, + 518 + ], + "lines": [ + { + "bbox": [ + 128, + 504, + 433, + 520 + ], + "spans": [ + { + "bbox": [ + 128, + 504, + 433, + 520 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 146, + 521, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 145, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 145, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 161, + 532, + 237, + 544 + ], + "spans": [ + { + "bbox": [ + 161, + 532, + 237, + 544 + ], + "score": 1.0, + "content": "applicable? [N/A]", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 145, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 145, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 162, + 555, + 342, + 568 + ], + "spans": [ + { + "bbox": [ + 162, + 555, + 342, + 568 + ], + "score": 1.0, + "content": "Board (IRB) approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 146, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 146, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 161, + 580, + 333, + 592 + ], + "spans": [ + { + "bbox": [ + 161, + 580, + 333, + 592 + ], + "score": 1.0, + "content": "spent on participant compensation? [N/A]", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + } + ], + "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": "title", + "bbox": [ + 107, + 71, + 156, + 84 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 158, + 85 + ], + "score": 1.0, + "content": "Checklist", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 131, + 92, + 208, + 104 + ], + "lines": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "spans": [ + { + "bbox": [ + 129, + 91, + 210, + 105 + ], + "score": 1.0, + "content": "1. For all authors...", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 129, + 91, + 210, + 105 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 107, + 505, + 180 + ], + "lines": [ + { + "bbox": [ + 145, + 106, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 145, + 106, + 505, + 120 + ], + "score": 1.0, + "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 118, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 161, + 118, + 288, + 131 + ], + "score": 1.0, + "content": "contributions and scope? [Yes]", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 131, + 376, + 144 + ], + "spans": [ + { + "bbox": [ + 145, + 131, + 376, + 144 + ], + "score": 1.0, + "content": "(b) Did you describe the limitations of your work? [Yes]", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 144, + 470, + 156 + ], + "spans": [ + { + "bbox": [ + 145, + 144, + 470, + 156 + ], + "score": 1.0, + "content": "(c) Did you discuss any potential negative societal impacts of your work? [N/A]", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 145, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 167, + 214, + 181 + ], + "spans": [ + { + "bbox": [ + 161, + 167, + 214, + 181 + ], + "score": 1.0, + "content": "them? [Yes]", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + } + ], + "index": 4.5, + "bbox_fs": [ + 145, + 106, + 505, + 181 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 183, + 302, + 194 + ], + "lines": [ + { + "bbox": [ + 129, + 182, + 304, + 196 + ], + "spans": [ + { + "bbox": [ + 129, + 182, + 304, + 196 + ], + "score": 1.0, + "content": "2. If you are including theoretical results...", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 129, + 182, + 304, + 196 + ] + }, + { + "type": "text", + "bbox": [ + 146, + 197, + 452, + 222 + ], + "lines": [ + { + "bbox": [ + 145, + 197, + 453, + 210 + ], + "spans": [ + { + "bbox": [ + 145, + 197, + 453, + 210 + ], + "score": 1.0, + "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 145, + 210, + 424, + 223 + ], + "spans": [ + { + "bbox": [ + 145, + 210, + 424, + 223 + ], + "score": 1.0, + "content": "(b) Did you include complete proofs of all theoretical results? [N/A]", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 145, + 197, + 453, + 223 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 226, + 241, + 237 + ], + "lines": [ + { + "bbox": [ + 128, + 223, + 243, + 239 + ], + "spans": [ + { + "bbox": [ + 128, + 223, + 243, + 239 + ], + "score": 1.0, + "content": "3. If you ran experiments...", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11, + "bbox_fs": [ + 128, + 223, + 243, + 239 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 241, + 505, + 366 + ], + "lines": [ + { + "bbox": [ + 145, + 240, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 145, + 240, + 507, + 253 + ], + "score": 1.0, + "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 251, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 162, + 251, + 505, + 263 + ], + "score": 1.0, + "content": "mental results (either in the supplemental material or as a URL)? [Yes] See Section 5.1", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 162, + 262, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 162, + 262, + 287, + 274 + ], + "score": 1.0, + "content": "and the supplemental material.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 274, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 146, + 274, + 505, + 288 + ], + "score": 1.0, + "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 286, + 434, + 298 + ], + "spans": [ + { + "bbox": [ + 161, + 286, + 434, + 298 + ], + "score": 1.0, + "content": "were chosen)? [Yes] See Section 5.1 and the supplemental material.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 147, + 298, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 147, + 298, + 506, + 312 + ], + "score": 1.0, + "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 309, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 161, + 309, + 506, + 322 + ], + "score": 1.0, + "content": "ments multiple times)? [No] We have not repeated experiments many times to get error", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 161, + 320, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 161, + 320, + 506, + 334 + ], + "score": 1.0, + "content": "bars since experiments of 3d object detection on large scale datasets is time-consuming.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 146, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 146, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "(d) Did you include the total amount of compute and the type of resources used (e.g.,", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 344, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 161, + 344, + 506, + 356 + ], + "score": 1.0, + "content": "type of GPUs, internal cluster, or cloud provider)? [Yes] See Section 5.1 and the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 161, + 356, + 255, + 367 + ], + "spans": [ + { + "bbox": [ + 161, + 356, + 255, + 367 + ], + "score": 1.0, + "content": "supplemental material.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + } + ], + "index": 17, + "bbox_fs": [ + 145, + 240, + 507, + 367 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 370, + 505, + 382 + ], + "lines": [ + { + "bbox": [ + 129, + 368, + 507, + 385 + ], + "spans": [ + { + "bbox": [ + 129, + 368, + 507, + 385 + ], + "score": 1.0, + "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 129, + 368, + 507, + 385 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 385, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 145, + 385, + 491, + 398 + ], + "spans": [ + { + "bbox": [ + 145, + 385, + 491, + 398 + ], + "score": 1.0, + "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] See Section 5.1.", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 398, + 423, + 410 + ], + "spans": [ + { + "bbox": [ + 146, + 398, + 423, + 410 + ], + "score": 1.0, + "content": "(b) Did you mention the license of the assets? [Yes] See Section 5.1.", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 146, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 421, + 351, + 434 + ], + "spans": [ + { + "bbox": [ + 162, + 421, + 351, + 434 + ], + "score": 1.0, + "content": "See Section 5.1 and the supplemental material.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 146, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 447, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 162, + 447, + 505, + 458 + ], + "score": 1.0, + "content": "using/curating? [N/A] We generate the data using the open-sourced tool and the owners", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 162, + 457, + 365, + 469 + ], + "spans": [ + { + "bbox": [ + 162, + 457, + 365, + 469 + ], + "score": 1.0, + "content": "consent to all the public use for research purposes.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 469, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 146, + 469, + 506, + 483 + ], + "score": 1.0, + "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 162, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "information or offensive content? [N/A] Our data is synthesized using the open-sourced", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 162, + 491, + 506, + 503 + ], + "spans": [ + { + "bbox": [ + 162, + 491, + 506, + 503 + ], + "score": 1.0, + "content": "tool, so there is no real-world personally identifiable information, nor offensive content.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 28.5, + "bbox_fs": [ + 145, + 385, + 506, + 503 + ] + }, + { + "type": "text", + "bbox": [ + 131, + 506, + 432, + 518 + ], + "lines": [ + { + "bbox": [ + 128, + 504, + 433, + 520 + ], + "spans": [ + { + "bbox": [ + 128, + 504, + 433, + 520 + ], + "score": 1.0, + "content": "5. If you used crowdsourcing or conducted research with human subjects...", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34, + "bbox_fs": [ + 128, + 504, + 433, + 520 + ] + }, + { + "type": "list", + "bbox": [ + 146, + 521, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 145, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 145, + 520, + 506, + 534 + ], + "score": 1.0, + "content": "(a) Did you include the full text of instructions given to participants and screenshots, if", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 532, + 237, + 544 + ], + "spans": [ + { + "bbox": [ + 161, + 532, + 237, + 544 + ], + "score": 1.0, + "content": "applicable? [N/A]", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 145, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 145, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "(b) Did you describe any potential participant risks, with links to Institutional Review", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 162, + 555, + 342, + 568 + ], + "spans": [ + { + "bbox": [ + 162, + 555, + 342, + 568 + ], + "score": 1.0, + "content": "Board (IRB) approvals, if applicable? [N/A]", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 146, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 146, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 161, + 580, + 333, + 592 + ], + "spans": [ + { + "bbox": [ + 161, + 580, + 333, + 592 + ], + "score": 1.0, + "content": "spent on participant compensation? [N/A]", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + } + ], + "index": 37.5, + "bbox_fs": [ + 145, + 520, + 506, + 592 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file