Ëæ×ÅÉî¶ÈѧϰµÄÒ»Ö±Éú³¤£¬£¬£¬£¬£¬£¬£¬Ä¿µÄ¼ì²â¾«¶ÈµÄÌáÉýÓú¼ÓÄÑÌ⣬£¬£¬£¬£¬£¬£¬Ò»Ñùƽ³£¶øÑÔÄÜÌáÉý1~2%µÄmAP(ƽ¾ù׼ȷÂÊ)±ãËãÊǺܴóµÄÍ»ÆÆÁË¡£¡£¡£¡£¡£

¿ÉÊÇ£¬£¬£¬£¬£¬£¬£¬ÉÌÌÀÑо¿ÍŶӽÒÏþµÄTSDËã·¨ÌáÉýÁËÄÑÒÔÖÃÐŵÄ3~5% mAP!²¢ÇÒÊÇÔÚÐÐÒµÄѶÈ×î¸ßµÄÊý¾Ý¼¯Open ImagesºÍCOCO¡£¡£¡£¡£¡£ÒÔ¸ÃËã·¨×÷Ϊ½¹µãµÄ½â¾ö¼Æ»®(Á´½Ó£º

https://arxiv.org/pdf/2003.07557.pdf)ÈÙ»ñOpen Images Object Detection Challenge 2019 ¹Ú¾ü¡£¡£¡£¡£¡£

Ê×´ÎÌá³ö·ÖÀàºÍ»Ø¹éʹÃü»®·Öѧϰ

»úеÊìϤһ¸öÎïÌåÐèÒª´ÓÁ½¸öά¶È£º·ÖÀàºÍ»Ø¹é¡£¡£¡£¡£¡£·ÖÀàÊÇָʶ±ðÎïÌåµÄÊôÐÔ£¬£¬£¬£¬£¬£¬£¬ÀýÈçÅжÏÒ»·ùͼƬÉϵĶ¯ÎïÊÇһֻèÕÕ¾ÉÒ»Ö»¹·;»Ø¹éÊÇָʶ±ðÎïÌåµÄÇøÓòºÍλÖᣡ£¡£¡£¡£

¹Å°åÄ¿µÄ¼ì²âÒªÁì(ºÃ±ÈFaster RCNN)ͨ³£ÊÇ·ÖÀàºÍ»Ø¹éÒ»Æðѧϰ£¬£¬£¬£¬£¬£¬£¬¹²ÏíProposal(ÎïÌåDZÔÚ±£´æµÄÇøÓò¿ò)ºÍSibling head(ÌØÕ÷ÌáÈ¡Æ÷)¡£¡£¡£¡£¡£

ËüÊÇÔõÑùµÄÀú³ÌÄØ?Èçͼ2µÄÏ°벿·Ö£¬£¬£¬£¬£¬£¬£¬·ÖÀàºÍ»Ø¹é¹²ÏíÒ»¸öÊäÈëProposal P(¼´ºìÉ«¿ò)£¬£¬£¬£¬£¬£¬£¬ÒÔ¼°Í³Ò»¸öÌØÕ÷ÌáÈ¡Æ÷f(¡¤)£¬£¬£¬£¬£¬£¬£¬Êä³öЧ¹ûW/o TSD¡£¡£¡£¡£¡£

¿ÉÊÇÎÒÃÇ·¢Ã÷£¬£¬£¬£¬£¬£¬£¬×îÖÕÊä³öµÄͼƬ¿òÆä·ÖÀàÖÃÐŶȺͼì²â¿òµÄ׼ȷ¶ÈÊÇ·×Æçֵģ¬£¬£¬£¬£¬£¬£¬ÈçÏÂͼ¼ì²â¿òµÄÖÃÐŶȺܸߣ¬£¬£¬£¬£¬£¬£¬¿ÉÊÇûÓпòס·É»ú»úÒí£¬£¬£¬£¬£¬£¬£¬¼´ÇøÓò¿òÊǽû¾øÈ·µÄ¡£¡£¡£¡£¡£

ͼ2£º¹Å°åÄ¿µÄ¼ì²âÒªÁìºÍÉÌÌÀTSDÒªÁì½ÏÁ¿

ÄÇôΪʲô»áÔì³ÉÕâÑùµÄÎÊÌâ?ÉÌÌÀÑо¿ÍŶÓͨ¹ýÒ»¸öʵÑéÀ´Ì½Ñ°£¬£¬£¬£¬£¬£¬£¬·¢Ã÷·ÖÀàºÍ»Ø¹éÀú³ÌµÄÌص㣬£¬£¬£¬£¬£¬£¬ÈçÏÂͼ£º

ÆäÖÐÌåÏÖ¹ØÓÚµÚyÀàµÄÕ¹ÍûÖÃÐŶÈ£¬£¬£¬£¬£¬£¬£¬ÊÇÔ¤¶¨µÄMargin¡£¡£¡£¡£¡£

¹ØÓڻعéʹÃü£¬£¬£¬£¬£¬£¬£¬Í¨¹ýÒÔϹ«Ê½°ü¹Ü»Ø¹éʹÃü׼ȷ¶È±È¹Å°åÒªÁì¸ü¸ß£º

ÊÇԭʼ¼ì²âÆ÷Í·²¿»ñµÃµÄ¼ì²â¿ò£¬£¬£¬£¬£¬£¬£¬ÊÇTSDËã·¨»ñµÃµÄ¼ì²â¿ò¡£¡£¡£¡£¡£ÈôÊÇÄ¿½ñµÄProposalÊÇÒ»¸ö¸ºÑù±¾£¬£¬£¬£¬£¬£¬£¬ÄÇô¸ÃLoss»á±»Ignore¡£¡£¡£¡£¡£

ÔÚÍÆÀí½×¶Î£¬£¬£¬£¬£¬£¬£¬Ô­Ê¼µÄ¼ì²âÆ÷Í·²¿²»ÔÙʹÓᣡ£¡£¡£¡£Classical LossÊÇԭʼËðʧº¯Êý£¬£¬£¬£¬£¬£¬£¬TSD LossÊDZ¾ÎÄÌá³öµÄËðʧº¯Êý¡£¡£¡£¡£¡£

ʵÑéЧ¹û

1¡¢Óë²î±ðÍøÂç½âñî·½·¨½ÏÁ¿£¬£¬£¬£¬£¬£¬£¬TSDЧ¹û×îºÃ

Ê×ÏȽ«TSDÓëÔÚ²î±ðµÄÍøÂç²ã¾ÙÐÐʹÃü¼ä½âñîµÄ½á¹¹¾ÙÐнÏÁ¿£¬£¬£¬£¬£¬£¬£¬ÈçÉÏͼËùʾ¡£¡£¡£¡£¡£

±í1£ºÖÖÖÖ½âñî·½·¨ÓëTSD½âñî·½·¨½ÏÁ¿

×ۺϲÎÊýÄ¿ºÍÐÔÄÜÌáÉý£¬£¬£¬£¬£¬£¬£¬TSDÏà±ÈÓÚÆäËû½âñî·½·¨£¬£¬£¬£¬£¬£¬£¬¾ß±¸ÏÔ×ŵÄÓÅÊÆ¡£¡£¡£¡£¡£

2¡¢TSDÓëSibling headÍŽáѵÁ·Ð§¹û¸üºÃ

ÔÚTSDÕûÌåµÄѵÁ·ÖУ¬£¬£¬£¬£¬£¬£¬¹Å°åµÄSibling headÈÔÈ»¿ÉÒÔ¾ÙÐÐÍŽáѵÁ·À´ÓÅ»¯Backbone£¬£¬£¬£¬£¬£¬£¬Í¬Ê±½¥½øÔ¼ÊøËðʧ(PC)¿ÉÒÔ½øÒ»²½ÌáÉýTSDµÄÐÔÄÜ¡£¡£¡£¡£¡£

±í2£º¹Å°åSibling headÓëTSDÍŽáѵÁ·¿ÉÌáÉýÐÔÄÜ

±í3£º¼ÓÈëPCËðʧº¯Êý¹ØÓÚ·ÖÀàºÍ»Ø¹é׼ȷÂÊÓ°Ïì

´Ó±í2¿ÉÒÔ¿´³öÁ½¸öHeadÍŽáѵÁ·¿ÉÒÔ½øÒ»²½´øÀ´ÐÔÄܵÄÌáÉý¡£¡£¡£¡£¡£´Ó±í3¿ÉÒÔ¿´³ö£¬£¬£¬£¬£¬£¬£¬¼ÓÈëËðʧº¯ÊýPCÖ®ºó£¬£¬£¬£¬£¬£¬£¬ÎÞÂÛ¹ØÓÚ·ÖÀàÕվɹØÓڻع飬£¬£¬£¬£¬£¬£¬PC¶¼¿ÉÒÔ½øÒ»²½´øÀ´×¼È·ÂÊÌáÉý£¬£¬£¬£¬£¬£¬£¬Á½Õ߶¼ÓÃPC´øÀ´µÄЧ¹ûÊÇ×îºÃµÄ¡£¡£¡£¡£¡£

3¡¢TSDÔÚCOCO¡¢Open ImagesºÍÍøÂç½á¹¹ÉϾùÄÜ´ó·ùÌáÉý׼ȷÂÊ

´Ó±í4¿ÉÒÔ¿´³ö£¬£¬£¬£¬£¬£¬£¬ÔÚ²î±ðÍøÂç½á¹¹ÉÏ(°üÀ¨ResNet-50¡¢ResNet-101¡¢ResNet-152µÈ)£¬£¬£¬£¬£¬£¬£¬±¾ÎÄÌá³öµÄTSDÒªÁìÔÚ׼ȷ¶ÈÉϾùÓÐÏÔ×ÅÌáÉý¡£¡£¡£¡£¡£ÇÒÍÆÀíʱ¼äÔöÌí²»Áè¼Ý10%¡£¡£¡£¡£¡£

±í4£ºÔÚ²î±ðÍøÂç½á¹¹ÉÏ£¬£¬£¬£¬£¬£¬£¬¹Å°åÒªÁìÓëTSDÒªÁì׼ȷ¶È±ÈÕÕ

±í5¿ÉÒÔ¿´³ö£¬£¬£¬£¬£¬£¬£¬ÔڹȸèÌá³öµÄOpen Images´ó¹æÄ£Êý¾Ý¼¯ÉÏ£¬£¬£¬£¬£¬£¬£¬TSDÒªÁì¾ùÓÐÐÔÄÜÉϵÄÏÔÖøÌáÉý¡£¡£¡£¡£¡£

±í5£ºÔÚOpen ImagesÊý¾Ý¼¯ÉÏ£¬£¬£¬£¬£¬£¬£¬

¹Å°åÒªÁìÓëTSDÒªÁì±ÈÕÕ

±í6£ºÔÚCOCOÖ§½âʹÃüÉÏ£¬£¬£¬£¬£¬£¬£¬¹Å°åÒªÁìÓëTSDÒªÁì±ÈÕÕ

ÔÚ±í6µÄCOCOÖ§½âʹÃüÉÏ£¬£¬£¬£¬£¬£¬£¬TSDÒªÁìÒ²ÄÜ´øÀ´ÏÔ×ÅÐÔÄÜÌáÉý£¬£¬£¬£¬£¬£¬£¬²¢ÇÒÌáÉý·ù¶È½Ï´ó¡£¡£¡£¡£¡£

4¡¢ÓëÄ¿½ñÐÐÒµÁìÏÈËã·¨½ÏÁ¿

»ùÓÚResNet-101µÄBackbone£¬£¬£¬£¬£¬£¬£¬TSDËã·¨µÖ´ïÁËеÄstate-of-the-artµÄÐÔÄÜ£¬£¬£¬£¬£¬£¬£¬²¢ÇÒÔÚSENet154-DCNµÄ»ù´¡½á¹¹Ï£¬£¬£¬£¬£¬£¬£¬COCOµÄÐÔÄִܵïÁË51.2 map¡£¡£¡£¡£¡£

½áÂÛ

±¾ÎÄÕë¶ÔͨÓÃÎïÌå¼ì²âËã·¨ÖзÖÀàʹÃüºÍ»Ø¹éʹÃüÖ®¼äÓÉÓÚÓÅ»¯Ä¿µÄ·×ÆçÖ´øÀ´µÄDZÔÚ³åÍ»£¬£¬£¬£¬£¬£¬£¬Ìá³öÁË»ùÓÚʹÃü¼ä¿Õ¼ä×Ô˳Ӧ½âñîµÄ¼ì²âËã·¨TSD¡£¡£¡£¡£¡£

ÔÚ¼ì²âÆ÷Í·²¿Í¨¹ýÌض¨Éè¼ÆµÄÆ«ÒÆÁ¿ÌìÉúÕ½ÂÔÒÔ¼°ÍŽáѵÁ·ÓÅ»¯½¥½øËðʧÀ´ÓÐÓõÄÌáÉý¼ì²âÆ÷ÐÔÄÜ£¬£¬£¬£¬£¬£¬£¬ÔÚÌØÁíÍâÍÆÀíʱ¼äÏûºÄ²»Áè¼Ý10%µÄÇéÐÎÏ£¬£¬£¬£¬£¬£¬£¬Äܹ»ÎȹÌÌáÉý3%~5%µÄ¼ì²âÐÔÄÜ£¬£¬£¬£¬£¬£¬£¬²¢ÇÒ³ÉΪOpen Images 2019 Object detection challenge¶á¹ÚµÄ½¹µãËã·¨¡£¡£¡£¡£¡£

´ó×ÚʵÑé֤ʵ£¬£¬£¬£¬£¬£¬£¬TSD¿ÉÒÔ½ÏÁ¿ÈÝÒ׵ĴîÅäÖݪֲî±ðµÄÍøÂç½á¹¹ÒÔ¼°Anchor-basedµÄ¼ì²â¿ò¼ÜÀ´¾ÙÐÐʹÓᣡ£¡£¡£¡£

¸ü¶à¾«²ÊÄÚÈÝ¿ÉÉó²éÔ­ÎÄ£¬£¬£¬£¬£¬£¬£¬Á´½Ó£ºhttps://arxiv.org/pdf/2003.07540.pdf

ÕÐƸ

ÉÌÌÀÑо¿Ôº-»ù´¡ÊÖÒÕÍŶÓÕÐƸÑо¿Ô±\¼ûÏ°Ñо¿Ô±£º

1.Ç¿»¯Ñ§Ï°Æ«Ïò£ºÌ½Ë÷ÔÚÖØ´óÇéÐÎ(ÈçÓÎÏ·AI¡¢ÏµÍ³ÓÅ»¯)ÖеÄÇ¿»¯Ñ§Ï°ÒªÁ죬£¬£¬£¬£¬£¬£¬Ñо¿°üÀ¨µ«²»ÏÞÓÚmulti agent RL¡¢imitation learning¡¢model-free³¡¾°ÏÂsampling efficiencyµÈ»ù´¡ÎÊÌâ;¼ÓÈëÏÖʵÏîÄ¿ÖеÄRLÒªÁìÉè¼ÆÓë̽Ë÷¡£¡£¡£¡£¡£

2.¾ÞÐÍÍøÂçÓÅ»¯Æ«Ïò£ºÑо¿ÔÚbillion levelÊý¾Ýϵij¬´óÐÍÍøÂç½á¹¹Éè¼ÆÓëËÑË÷£¬£¬£¬£¬£¬£¬£¬°üÀ¨µ«²»ÏÞÓÚÈËÁ³Ê¶±ðÄ£×ÓÓÅ»¯¡¢sampling based NAS£¬£¬£¬£¬£¬£¬£¬optimizerÉè¼ÆÓëËÑË÷µÈ¡£¡£¡£¡£¡£

ÓÐÐËȤµÄСͬ°éÇëͶµÝ¼òÀúÖÁliuyu@sensetime.com¡£¡£¡£¡£¡£