Graph-Based 3D Multi-Person Pose Estimation Using Multi-View Images

Size Wu^1,3  Sheng Jin^2,3  Wentao Liu³  Lei Bai⁴  Chen Qian³  Dong Liu¹  Wanli Ouyang^{4  1}University of Science and Technology of China  ²The University of Hong Kong  ³SenseTime Research and Tetras.AI  ⁴The University of Sydney

Wsz32741010@mail.ustc.edu.cn     {jinsheng, liuwentao, qianchen}@sensetime.com      baisanshi@gmail.com     dongeliu@ustc.edu.cn     wanly.ouyang@uni.sydney.edu.au

Part 1 ÄîÍ·ºÍÅä¾°

ͼ 1 ±¾ÎÄËã·¨µÄ¿ÉÊÓ»¯Õ¹Ê¾

¶àÊӽǵĶàÈË3D×Ë̬Ԥ¼Æ£¬£¬£¬£¬£¬£¬Ö¸´Ó¶à¸öÏà»ú£¨Òѱ궨£©ÅÄÉãµÄͼÏñÖУ¬£¬£¬£¬£¬£¬»Ö¸´³ö¸Ã³¡¾°µÄËùÓеÄÈËÌå¹Ç¼Ü¡£¡£¡£¡£¡£¡£¡£¡£ËüÊÇÊÓƵÐж¯ÆÊÎöºÍ¸ß¼¶ÈË»ú½»»¥µÄ»ù´¡£¡£¡£¡£¡£¡£¡£¡£¬£¬£¬£¬£¬£¬ÔÚÔ˶¯ÆÊÎö¡¢Ó°ÊÓÌØЧµÈÁìÓò¾ßÓÐÖ÷ÒªµÄÓ¦ÓÃÔ¶¾°¡£¡£¡£¡£¡£¡£¡£¡£

ÏÖÔÚ¶àÊÓ½Ç3DÈËÌå×Ë̬Ԥ¼ÆÓÐÈçÏÂÁ½ÖÖÖ÷Á÷¼Æ»®£º

ͼ 2 Æ¥Åä+Èý½Ç»¯ÖØÐÞËã·¨µÄʾÒâͼ

(a) Æ¥Åä+Èý½Ç»¯ÖØÐÞ£¨Triangulation£©£ºÊ×ÏÈ£¬£¬£¬£¬£¬£¬¼ì²â¸÷¸öÊÓ½ÇµÄ 2D ÈËÌå¹Ç¼Ü£»£»£»£»£»£»£»ÔÙʹÓü¸ºÎ¹ØϵÅÌËã¾àÀ룬£¬£¬£¬£¬£¬¾ÙÐжàÊÓ½ÇÈËÌåµÄÆ¥Åä£»£»£»£»£»£»£»×îÖÕ£¬£¬£¬£¬£¬£¬Ê¹ÓöàÊӽǵÄ2D×ø±êºÍÏà»ú²ÎÊý£¬£¬£¬£¬£¬£¬ÅÌËã³ö3D×ø±ê¡£¡£¡£¡£¡£¡£¡£¡£

´ú±íÒªÁ죺MVPose[1]

Èõµã£º£¨1£©ÒªÁìµÄ¾«¶ÈÊ®·ÖÒÀÀµ 2D ¼ì²âЧ¹û¡£¡£¡£¡£¡£¡£¡£¡£2D ¼ì²âµÄÎó²î»áÓ°ÏìÆ¥ÅäµÄЧ¹û£¬£¬£¬£¬£¬£¬¹ýʧµÄÆ¥Åä½øÒ»²½µ¼ÖÂÒì³£µÄÖØÐÞ£»£»£»£»£»£»£»2D µÄÎó²îÒ²»áÓ°Ïì 3D ÖØÐ޵ľ«¶È¡£¡£¡£¡£¡£¡£¡£¡££¨2£©Æ¥Å䲿·ÖºÍÈý½ÇÖØÐÞ²¿·Ö²¢²»ÊÇÊý¾ÝÇý¶¯µÄ£¬£¬£¬£¬£¬£¬Ã»ÓмàÊÓѵÁ·ºÍËðʧ·´´«¡£¡£¡£¡£¡£¡£¡£¡£

ͼ 3 3D¿Õ¼äÌåËØ»¯Ëã·¨µÄʾÒâͼ

(b) 3D¿Õ¼äÌåËØ»¯£º½« 3D ¿Õ¼äµÈ¾àµØ»®·ÖΪһ¸ö¸öСÍø¸ñ£¬£¬£¬£¬£¬£¬Í¨¹ý¸ÅÂÊÄ£×Ó»òÕß 3D ¾í»ýÉñ¾­ÍøÂ磨CNN£©¼ì²âÒªº¦µã¡£¡£¡£¡£¡£¡£¡£¡£

´ú±íÒªÁ죺VoxelPose[2]

Èõµã£º£¨1£©¿Õ¼äÌåËØ»¯µÄ¾«¶È»áÊܵ½Íø¸ñ¾ÞϸµÄÖÆÔ¼£¬£¬£¬£¬£¬£¬»á±¬·¢Á¿»¯Îó²î£»£»£»£»£»£»£»£¨2£©¿Õ¼äÌåËØ»¯ÔÚÏàͬµÄ¾«¶ÈÏ£¬£¬£¬£¬£¬£¬ÅÌËãÖØƯºóËæ¿Õ¼ä¾ÞϸÈý´Î·½ÔöÌí£¬£¬£¬£¬£¬£¬ÎÞ·¨Ó¦ÓÃÓڽϴ󳡾°¡£¡£¡£¡£¡£¡£¡£¡£

ͼ 4 ±¾ÎÄËã·¨µÄʾÒâͼ

(c) ±¾ÎÄÒªÁ죺ÎÒÃÇÍŽáÁ½ÕßµÄÓÅÊÆ£¬£¬£¬£¬£¬£¬Ìá³öÁËÒ»ÖÖ»ùÓÚͼ¾í»ýÉñ¾­ÍøÂçµÄ£¬£¬£¬£¬£¬£¬×Ô¶¥ÏòÏ£¨Top-down£©µÄÁ½½×¶ÎËã·¨¡£¡£¡£¡£¡£¡£¡£¡£¹þ¹þ(haha)ÌåÓýÕûÌåËã·¨Á÷³Ì£¬£¬£¬£¬£¬£¬·ÖΪÁ½¸ö½×¶Î£º3DÈËÌåÖÐÐĵ㶨λ+3DÈËÌå×Ë̬Ԥ¼Æ¡£¡£¡£¡£¡£¡£¡£¡£

ͼ 5 ±¾ÎÄÒªÁìµÄÕûÌå¿ò¼Üͼ

1. Õë¶Ô3DÈËÌåÖÐÐĵ㶨룬£¬£¬£¬£¬£¬ÎÒÃǽÓÄÉÁËproposal-refine "ÓÉ´Öµ½¾«"µÄÕ½ÂÔ¡£¡£¡£¡£¡£¡£¡£¡£Ê×ÏȾÙÐС±´Öɸѡ¡±£¬£¬£¬£¬£¬£¬Õ¹ÍûһϵÁкòÑ¡ÖÐÐĵãλÖã¨MMG£©£»£»£»£»£»£»£»ÔÙÔÚºòÑ¡ÖÐÐĵãÖÜΧ¡°Ï¸Ä塱ËÑË÷£¨CRG£©£¬£¬£¬£¬£¬£¬»ñµÃ׼ȷµÄÈËÌåÖÐÐĵ㶨λЧ¹û¡£¡£¡£¡£¡£¡£¡£¡£

Óŵ㣺£¨1£©ÏÖÓÐÒªÁìÔÚÈں϶àÊӽǵÄÌØÕ÷ʱ£¬£¬£¬£¬£¬£¬Í¨³£Ö±½Ó¼¶Áª»ò³Ø»¯¸÷¸öÊӽǵÄÌØÕ÷ÏòÁ¿£¬£¬£¬£¬£¬£¬ºöÊÓÁËÊÓ½ÇÖ®¼äµÄÏ໥¹Øϵ¡£¡£¡£¡£¡£¡£¡£¡£ÎÒÃÇʹÓÃͼ¾í»ýÍøÂ磬£¬£¬£¬£¬£¬Äܹ»ÓÐÓÃʹÓÿçÊӽǵÄÌØÕ÷¡£¡£¡£¡£¡£¡£¡£¡££¨2£©¹þ¹þ(haha)ÌåÓý"ÓÉ´Öµ½¾«"Õ½ÂÔ£¬£¬£¬£¬£¬£¬Äܹ»ÓÐÓÃÌáÉýÄ£×ÓµÄÕ¹Íû¾«¶È¡£¡£¡£¡£¡£¡£¡£¡££¨3£©ËÑË÷¿Õ¼ä±»ÏÞÖÆÔÚÁ˺òÑ¡ÖÐÐĵãÖÜΧ£¬£¬£¬£¬£¬£¬ÓëÏÖʵµÄ¿Õ¼ä¾ÞϸÎ޹أ¬£¬£¬£¬£¬£¬´ó´ó½µµÍÁËÔËËãÖØƯºó¡£¡£¡£¡£¡£¡£¡£¡£

2. Õë¶Ô3DÈËÌå×Ë̬Ԥ¼Æ£¬£¬£¬£¬£¬£¬ÎÒÃÇͬÑù½ÓÄÉ"ÓÉ´Öµ½¾«"µÄÕ½ÂÔ£¬£¬£¬£¬£¬£¬Ê×ÏÈÓÃÄ£×ÓÕ¹Íû³õʼ3DÈËÌå×Ë̬£¬£¬£¬£¬£¬£¬ÔÙÅäºÏʹÓÃÈËÌå¹Ç¼Ü½á¹¹ÐÅÏ¢ºÍ¶àÊÓ½ÇÌØÕ÷£¬£¬£¬£¬£¬£¬À´ÓÅ»¯Õ¹ÍûµÄЧ¹û£¨PRG£©¡£¡£¡£¡£¡£¡£¡£¡£

Óŵ㣺ÎÒÃǽÓÄÉÁË"ÓÉ´Öµ½¾«"Õ½ÂÔ¡£¡£¡£¡£¡£¡£¡£¡£Í¼¾í»ýÉñ¾­ÍøÂçPRG£¬£¬£¬£¬£¬£¬Äܹ»Ö±½Ó¶ÔÈËÌå¹Ç¼Ü½á¹¹¾ÙÐн¨Ä££¬£¬£¬£¬£¬£¬Ê¹ÓùǼÜÔ¼Êø£¬£¬£¬£¬£¬£¬ÓÅ»¯ÁËÄ£×ÓµÄÕ¹ÍûЧ¹û¡£¡£¡£¡£¡£¡£¡£¡£

Part 2: ÒªÁì

2.1 ÖÐÐĵãºòÑ¡ÇøÓòÌìÉú£ºMulti-view Matching Graph(MMG)

ÎÒÃÇÉè¼ÆÁ˶àÊÓ½ÇÆ¥ÅäͼÉñ¾­ÍøÂ磨MMG£©£¬£¬£¬£¬£¬£¬Åжϲî±ðÊӽǵÄÁ½Á½2DÖÐÐĵ㣬£¬£¬£¬£¬£¬ÊÇ·ñÊôÓÚͳһСÎÒ˽¼Ò¡£¡£¡£¡£¡£¡£¡£¡£Ëæºó£¬£¬£¬£¬£¬£¬¹ØÓÚÊôÓÚͳһСÎÒ˽¼ÒµÄÒ»¶ÔÒªº¦µã£¬£¬£¬£¬£¬£¬Á½Á½ÖØÐÞ³öÒ»¸ö3D×ø±ê×÷ΪºòÑ¡ÖÐÐĵã¡£¡£¡£¡£¡£¡£¡£¡£

ͼ 6 MMGÄ£¿£¿£¿£¿£¿£¿£¿éʾÒâͼ

½¨Í¼£¨Graph Construction£©£ºÎÒÃÇʹÓø÷¸öÊӽǼì²â»ñµÃµÄÈËÌå2DÖÐÐĵ㣬£¬£¬£¬£¬£¬À´½á¹¹¿çÊÓ½Çͼģ×Ó¡£¡£¡£¡£¡£¡£¡£¡£Í¼Ä£×ӵġ°½Úµã¡±Îª ¸÷¸öÊӽǼì²âµ½µÄ2DÈËÌåÖÐÐĵã£»£»£»£»£»£»£»½ÚµãµÄÌØÕ÷Ϊ£º2DÈËÌåÖÐÐĵãλÖõÄͼÏñÌØÕ÷¡£¡£¡£¡£¡£¡£¡£¡£Í¼Ä£×ӵġ°±ß¡±£¬£¬£¬£¬£¬£¬Á½Á½ÅþÁ¬²î±ðÊӽǵĽڵ㣬£¬£¬£¬£¬£¬Á½¸ö½Úµã¶ÔÓ¦µÄ2DÖÐÐĵãµÄ¼«Ïß¾àÀë×÷Ϊ±ßÌØÕ÷¡£¡£¡£¡£¡£¡£¡£¡£

ÐÅϢת´ï£¨Message Passing£©£ºÊ¹ÓÃGNNÀ´¾ÙÐÐÏà¹ØÐÔÌØÕ÷µÄѧϰ¡£¡£¡£¡£¡£¡£¡£¡£ÎÒÃÇʹÓÃEdgeConvÀ´´î½¨Í¼¾í»ýÉñ¾­ÍøÂçÄ£×Ó£¬£¬£¬£¬£¬£¬¶ÔËù½á¹¹ºÃµÄGraph ¾ÙÐоí»ý£¬£¬£¬£¬£¬£¬Ò»Ö±¸üнڵãµÄÌØÕ÷£»£»£»£»£»£»£»Í¨¹ýͼµÄ±íÕ÷£¬£¬£¬£¬£¬£¬Ä£×ÓͬʱʹÓÃÁ˼¸ºÎÐÅÏ¢£¨±ßÌØÕ÷£©ºÍͼÏñÐÅÏ¢£¨½ÚµãÌØÕ÷£©£¬£¬£¬£¬£¬£¬¸ßЧµÄÈں϶àÊÓ½ÇÌØÕ÷£¬£¬£¬£¬£¬£¬Æ¥Å侫¶ÈÔ¶¸ßÓÚÖ±½ÓʹÓü«ÏßÆ¥Åä¡£¡£¡£¡£¡£¡£¡£¡£

ÅжϱߵÄÊôÐÔ£ºÑµÁ·Ò»¸ö±ßÅбðÆ÷£¨Edge Discriminator£©£¬£¬£¬£¬£¬£¬¶Ôÿһ¶ÔÖÐÐĵ㣨¼´Ò»¸ö±ß£©¾ÙÐÐÅб𣬣¬£¬£¬£¬£¬ÅжÏÕâÒ»¶ÔÖÐÐĵãÊÇ·ñÊôÓÚͳһСÎÒ˽¼Ò¡£¡£¡£¡£¡£¡£¡£¡£

Ìá³öºòÑ¡µã£ºÃ¿Ò»¶Ô±»ÅжÏΪͳһСÎÒ˽¼ÒµÄÖÐÐĵ㣬£¬£¬£¬£¬£¬Í¨¹ýÈý½Ç»¯ÖØÐÞ³öÒ»¸ö3DºòÑ¡µã¡£¡£¡£¡£¡£¡£¡£¡£

2. 2 ÖÐÐĵã×ø±êÓÅ»¯£ºCenter Refinement Graph(CRG)

ÓÐÁ˺òÑ¡ÖÐÐĵãºó£¬£¬£¬£¬£¬£¬ÎÒÃÇÒÔºòÑ¡µãΪÇòÐĵÄÇòÐιæÄ£×÷ΪËÑË÷¿Õ¼ä£¬£¬£¬£¬£¬£¬ÎÞаµØÔÚºòÑ¡ÇøÓò²ÉÑù¡£¡£¡£¡£¡£¡£¡£¡£¹ØÓÚÿ¸ö²ÉÑùµã£¬£¬£¬£¬£¬£¬½«²ÉÑùµãͶӰµ½¸÷¸öÊӽDz¢ÔÚÏìӦλÖÃÌáÈ¡ÌØÕ÷¡£¡£¡£¡£¡£¡£¡£¡£½Ó×ÅʹÓÃÖÐÐĵãÓÅ»¯Í¼Ä£×Ó£¨CRG£©£¬£¬£¬£¬£¬£¬Í¨¹ý¶à¸öÊӽǽڵãµÄÅþÁ¬£¬£¬£¬£¬£¬£¬ÊµÏÖÁ˸ßЧµÄ¶àÊÓ½ÇÌØÕ÷Èںϣ¬£¬£¬£¬£¬£¬¿ÉÒÔ׼ȷµØÅжϲÉÑùµãÊÇ·ñΪÈËÌåÖÐÐĵã¡£¡£¡£¡£¡£¡£¡£¡£

ͼ7 CRGÄ£¿£¿£¿£¿£¿£¿£¿éʾÒâͼ

ËÑË÷¿Õ¼ä£ºÒÔºòÑ¡µãΪÇòÐĵÄÇòÐιæÄ£×÷ΪËÑË÷¿Õ¼ä£¬£¬£¬£¬£¬£¬ËùÓÐËÑË÷¿Õ¼äµÄ²¢¼¯×÷Ϊ×ܵÄËÑË÷¿Õ¼ä¡£¡£¡£¡£¡£¡£¡£¡£

²ÉÑù£ºÎÒÃÇ¿ÉÒÔ¾ÙÐÐ˳ӦÐԵIJÉÑù£¬£¬£¬£¬£¬£¬ÏÈÔÚËÑË÷¿Õ¼äÖеȾà²ÉÑù£¬£¬£¬£¬£¬£¬ÓÃÖÐÐĵãͼ¾í»ýÍøÂç¼ì²âÖÐÐĵ㣬£¬£¬£¬£¬£¬ÔÚ¼ì²âµ½µÄÖÐÐĵãÖÜΧ½øÒ»²½Ï¸Äå²ÉÑù£¬£¬£¬£¬£¬£¬ÒÔ»ñµÃ¸ü׼ȷµÄλÖᣡ£¡£¡£¡£¡£¡£¡£

½¨Í¼£¨Graph Construction£©£º¶Ôÿһ¸ö3D ²ÉÑùµã£¬£¬£¬£¬£¬£¬¹¹½¨Ò»¸öͼģ×Ó¡£¡£¡£¡£¡£¡£¡£¡£ÆäÖУ¬£¬£¬£¬£¬£¬½ÚµãºÍ½ÚµãÌØÕ÷»®·Ö¶ÔÓ¦ 3D²ÉÑùµãͶӰµ½¸÷¸öÊӽǺóµÄ2DλÖà ÒÔ¼° ¸ÃλÖõÄͼÏñÌØÕ÷¡£¡£¡£¡£¡£¡£¡£¡£Í¼Öеıߣ¬£¬£¬£¬£¬£¬¶Ô¸÷¸ö½Úµã¾ÙÐÐÈ«ÅþÁ¬¡£¡£¡£¡£¡£¡£¡£¡£

ÐÅϢת´ï£¨Message Passing£©£ºÊ¹ÓÃGCNÀ´¾ÙÐÐÏà¹ØÐÔÌØÕ÷µÄѧϰ¡£¡£¡£¡£¡£¡£¡£¡£ÎÒÃÇʹÓÃEdgeConvÀ´´î½¨Í¼¾í»ýÉñ¾­ÍøÂçÄ£×Ó£¬£¬£¬£¬£¬£¬Ê¹Óöà²ãͼ¾í»ýÒ»Ö±¸üнڵãµÄÌØÕ÷¡£¡£¡£¡£¡£¡£¡£¡£

ÅжÏͼµÄÊôÐÔ£ºÊ×ÏȶÔͼÖнڵã¾ÙÐÐÈ«¾Ö³Ø»¯£¬£¬£¬£¬£¬£¬»ñµÃͼµÄÌØÕ÷£¬£¬£¬£¬£¬£¬ÔÙѵÁ·Ò»¸ö¶à²ã¸ÐÖª»úMLPs£¬£¬£¬£¬£¬£¬ÅжÏͼµÄÊôÐÔ£ºÅжϲÉÑùµãÊÇ·ñΪÈËÌåÖÐÐÄ£¬£¬£¬£¬£¬£¬¼´Êä³öÿ¸ö²ÉÑùµãΪÈËÌåÖÐÐĵÄÖÃÐŶÈ¡£¡£¡£¡£¡£¡£¡£¡£

·Ç¼«´óÖµÒÖÖÆ (NMS)£º»ñµÃÿ¸ö²ÉÑùµãµÄÖÃÐŶȺ󣬣¬£¬£¬£¬£¬Í¨¹ýNMS²Ù×÷£¬£¬£¬£¬£¬£¬»ñµÃÓÅ»¯ºóµÄÈËÌåÖÐÐĵã×ø±ê¡£¡£¡£¡£¡£¡£¡£¡£

2. 3 ÈËÌå×Ë̬ÓÅ»¯£ºPose Regression Graph(PRG)

ͼ8 PRGÄ£¿£¿£¿£¿£¿£¿£¿éʾÒâͼ

ÎÒÃǽÓÄÉ"ÓÉ´Öµ½¾«"µÄÕ½ÂÔ£¬£¬£¬£¬£¬£¬À´Ô¤¼ÆÈËÌå3D×Ë̬¡£¡£¡£¡£¡£¡£¡£¡£Ê×ÏÈ£¬£¬£¬£¬£¬£¬ÎÒÃǽÓÄÉÏÖÓеÄ×Ë̬Ԥ¼ÆÒªÁì[2]£¬£¬£¬£¬£¬£¬»ñµÃ³õʼµÄ3DÈËÌå×Ë̬£¨Initial 3D Pose£©£»£»£»£»£»£»£»ÎªÁËÌáÉýÈËÌåÒªº¦µãµÄÕ¹Íû¾«¶È£¬£¬£¬£¬£¬£¬±¾ÎÄÌá³öÁËÈËÌå×Ë̬»Ø¹éͼģ×Ó£¨Pose Regression Graph, PRG£©£¬£¬£¬£¬£¬£¬Ê¹ÓÃͼ¾í»ý£¬£¬£¬£¬£¬£¬¸ßЧµØÈں϶àÊӽǵÄÌØÕ÷ºÍÈËÌåµÄÍØÆ˽ṹÐÅÏ¢£¨Í·ºÍ²±×ÓÏàÁ¬£¬£¬£¬£¬£¬£¬Êֺ͸첲ÏàÁ¬µÈ£©£¬£¬£¬£¬£¬£¬»Ø¹é³öÿ¸öÒªº¦µã×ø±êµÄÐÞÕýÖµ¡£¡£¡£¡£¡£¡£¡£¡£

½¨Í¼£¨Graph Construction£©£ºÊ×ÏȽ«³õʼ3DÈËÌå×Ë̬£¬£¬£¬£¬£¬£¬Í¶Ó°µ½¸÷¸öÏà»úÊӽǣ¬£¬£¬£¬£¬£¬»ñµÃ¸÷¸öÊӽǵÄ2DÈËÌå×Ë̬¡£¡£¡£¡£¡£¡£¡£¡£Ê¹Óø÷¸öÊÓ½ÇϵÄ2DÈËÌåÒªº¦µã£¬£¬£¬£¬£¬£¬½á¹¹¿çÊÓ½Ç×Ë̬ͼ£¨Multi-view Pose Graph£©¡£¡£¡£¡£¡£¡£¡£¡£

ͼģ×ÓµÄÿ¸ö¡°½Úµã¡±´ú±í¸÷¸öÊӽǵÄÿ¸ö2DÒªº¦µã¡£¡£¡£¡£¡£¡£¡£¡£½ÚµãµÄÌØÕ÷°üÀ¨£ºÒªº¦µãµÄÖÖ±ðÐÅÏ¢£¬£¬£¬£¬£¬£¬2DÒªº¦µãλÖô¦µÄͼÏñÌØÕ÷£¬£¬£¬£¬£¬£¬ÒÔ¼°³õʼ3DÈËÌå×Ë̬ÖжÔÓ¦Òªº¦µãµÄÖÃÐŶÈ¡£¡£¡£¡£¡£¡£¡£¡£

ͼģ×ӵġ°±ß¡±´ú±í½ÚµãÖ®¼äµÄ¹Øϵ£¬£¬£¬£¬£¬£¬¹²°üÀ¨Á½ÖÖÀàÐ͵ıߣº¿çÊÓ½ÇÇÒÏàͬÀàÐÍÒªº¦µã¼äµÄÅþÁ¬±ß ºÍ µ¥ÊÓ½ÇÇÒ²î±ðÀàÐÍÒªº¦µã¼äµÄÅþÁ¬±ß¡£¡£¡£¡£¡£¡£¡£¡£

ÐÅϢת´ï£¨Message Passing£©£ºÊ¹ÓÃGCNÀ´¾ÙÐÐÏà¹ØÐÔÌØÕ÷µÄѧϰ¡£¡£¡£¡£¡£¡£¡£¡£ÎÒÃÇʹÓÃEdgeConvÀ´´î½¨Í¼¾í»ýÉñ¾­ÍøÂçÄ£×Ó£¬£¬£¬£¬£¬£¬¶ÔËù½á¹¹ºÃµÄGraph ¾ÙÐоí»ý£¬£¬£¬£¬£¬£¬Ò»Ö±¸üнڵãµÄÌØÕ÷£»£»£»£»£»£»£»Íê³É¶àÊÓ½ÇÌØÕ÷µÄ¸üкÍÈںϺ󣬣¬£¬£¬£¬£¬¶ÔÏàͬÀàÐ͵ÄÒªº¦µãÌØÕ÷¾ÙÐÐ×î´óÖµ³Ø»¯£¬£¬£¬£¬£¬£¬»ñµÃÒ»¸±ÈËÌå¹Ç¼Ü¡£¡£¡£¡£¡£¡£¡£¡£

»Ø¹éÐÞÕýÖµ£¨Regression£©£ºÊ¹ÓûعéÄ£×Ó£¬£¬£¬£¬£¬£¬Õ¹Íû³ö Nx3 άµÄÆ«ÒÆÏòÁ¿£¨ÆäÖУ¬£¬£¬£¬£¬£¬N´ú±íÒªº¦µã¸öÊý£©£¬£¬£¬£¬£¬£¬´ú±íÏà¹ØÓÚ³õʼ3DÈËÌå×Ë̬µÄÐÞÕýÖµ¡£¡£¡£¡£¡£¡£¡£¡£

Part 3: ʵÑéЧ¹û

ÎÒÃÇÔÚ CMU Panoptic ºÍ Shelf Á½¸öÖ÷Á÷µÄÊý¾Ý¼¯ÉÏ×öÁËʵÑé¡£¡£¡£¡£¡£¡£¡£¡£¶¨Á¿ÊµÑéÅú×¢£¬£¬£¬£¬£¬£¬ÎÒÃÇÌá³öµÄËã·¨£¬£¬£¬£¬£¬£¬È¡µÃÁË×îÓŵľ«¶È¡£¡£¡£¡£¡£¡£¡£¡£²¢ÇÒÔÚÅÌËãÁ¿ºÍºÄʱ·½Ã棬£¬£¬£¬£¬£¬¹þ¹þ(haha)ÌåÓýÒªÁìÏà±È֮ǰµÄSOTAÒ²ÓÐÏÔ×ÅÓÅÊÆ¡£¡£¡£¡£¡£¡£¡£¡£

±í1 CMU Panoptic Êý¾Ý¼¯µÄ¶¨Á¿Ð§¹û±ÈÕÕ

±í2 Shelf Êý¾Ý¼¯µÄ¶¨Á¿Ð§¹û±ÈÕÕ

Part 4: ×ܽáÓëÕ¹Íû

ÔÚ±¾ÎÄÖУ¬£¬£¬£¬£¬£¬ÎÒÃÇÌá³öÁËÒ»Ì××Ô¶¥ÏòϵĶàÊÓ½Ç3DÈËÌå×Ë̬Ԥ¼Æ½â¾ö¼Æ»®¡£¡£¡£¡£¡£¡£¡£¡£ÎÒÃÇÕë¶Ô¸ÃʹÃü£¬£¬£¬£¬£¬£¬È«ÐÄÉè¼ÆÁËÖÖÖÖ¡°¶àÊӽǡ±Í¼¾í»ýÉñ¾­ÍøÂ磨MMG, CRG, PRG£©À´ÌáÈ¡ÈËÌå½á¹¹ÐÔÌØÕ÷¡£¡£¡£¡£¡£¡£¡£¡£ÎÒÃÇÔÚ¸÷Êý¾Ý¼¯ÉϵÄʵÑ飬£¬£¬£¬£¬£¬Ò²³ä·Ö֤ʵÎúÎÒÃÇËã·¨µÄÓÐÓÃÐÔ¡£¡£¡£¡£¡£¡£¡£¡£¹ØÓÚ¶ÔδÀ´µÄÕ¹Íû£¬£¬£¬£¬£¬£¬ÎÒÃǽ«¼ÌÐøÑо¿°ÑËã·¨À©Õ¹µ½Ê±Ðò£¬£¬£¬£¬£¬£¬ÊµÏÖ¸ü¸ßЧµÄ¶àÊÓ½ÇÈËÌå×Ë̬¸ú×Ù¡£¡£¡£¡£¡£¡£¡£¡£ÔÚÒªÁì²ãÃ棬£¬£¬£¬£¬£¬ÔõÑùÔ½·¢ºÏÀíµØʹÓÃÏà»úµÄ¼¸ºÎÐÅÏ¢£¬£¬£¬£¬£¬£¬Éè¼Æ¸ü¸ßЧµÄͼ¾í»ýÉñ¾­ÍøÂ磬£¬£¬£¬£¬£¬ÊÇÒ»¸öÖ÷ÒªµÄË¢ÐÂÆ«Ïò¡£¡£¡£¡£¡£¡£¡£¡£

Reference:

[1] Dong, J., Jiang, W., Huang, Q., Bao, H., & Zhou, X. (2019). Fast and robust multi-person 3d pose estimation from multiple views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 7792-7801).

[2] Tu, H., Wang, C., & Zeng, W. (2020). Voxelpose: Towards multi-camera 3d human pose estimation in wild environment. In Computer Vision¨CECCV 2020: 16th European Conference, Glasgow, UK, August 23¨C28, 2020, Proceedings, Part I 16 (pp. 197-212). Springer International Publishing.