µ¼¶Á£ºÔÚCVPR 2020ÉÏ£¬£¬£¬£¬£¬ÉÌÌÀÖǻ۽»Í¨²úÆ·ÏßÍŶÓÌá³öµÄ×Ô¾ÀÕý×Ô˳ӦÅòÕÍÂʼÆÊýÍøÂ磬£¬£¬£¬£¬Õë¶Ô¼ÆÊýʹÃüÖеã±êעλÖ÷×ÆçÖºÍ͸ÊÓÕ÷ÏóÔì³ÉÖØ´óµÄ±ê׼ת±äµÄÎÊÌâÌá³öÁËÓÐÓõÄÍøÂçÉè¼ÆºÍ¼àÊÓÒªÁì¡£¡£¡£¡£¡£ÔÚ¼àÊÓ·½·¨·½Ã棬£¬£¬£¬£¬ADSCNetʹÓÃÍøÂçѧϰµÄЧ¹ûÀ´¾ÀÕý·×ÆçÖµÄÈ˹¤±ê×¢´Ó¶ø¸üÓÐÓõÄѵÁ·£»£»£»£»£»£»£»ÔÚÍøÂçÉè¼Æ·½Ã棬£¬£¬£¬£¬ADSCNetÌá³ö×Ô˳ӦÅòÕÍÂʵľí»ý½á¹¹£¬£¬£¬£¬£¬²î±ðλÖýÓÄɲî±ðµÄÅòÕÍÂÊÀ´Ë³Ó¦±ê×¼µÄת±ä¡£¡£¡£¡£¡£ADSCNetÔÚËĸö¹ûÕæÊý¾Ý¼¯ÉϾùÓÐÏÔÖøµÄÌáÉý¡£¡£¡£¡£¡£

ÂÛÎÄÃû³Æ: Adaptive Dilated Network with Self-Correction Supervision for Counting

ÎÊÌâºÍÌôÕ½

Ä¿µÄ¼ÆÊý×÷ΪÅÌËã»úÊÓ¾õµÄÒ»¸öÖ÷ҪƫÏò¡£¡£¡£¡£¡£ÔÚ¹¤Òµ½çÓÐ×ÅÆÕ±éµÄÓ¦Ó㬣¬£¬£¬£¬ÀýÈ罻ͨ³¡¾°ÏµÄÓµ¶ÂÅжϣ¬£¬£¬£¬£¬ÊÓƵ¼àÊÓϵÄÁ÷Á¿Í³¼ÆÒÔ¼°µØÌúÖеÄÈËÁ÷ÆÊÎöµÈ¡£¡£¡£¡£¡£½üÄêÀ´£¬£¬£¬£¬£¬Ê¹Óþí»ýÉñ¾­ÍøÂ磨CNN£©µÄÒªÁìÈ¡µÃÁËÏÔÖøµÄÏ£Íû¡£¡£¡£¡£¡£¿ÉÊÇ£¬£¬£¬£¬£¬ÕâÏîʹÃüÈÔÈ»¾ßÓÐÌôÕ½£º

a. ÓÉÓÚ÷缯µÄ³¡¾°£¬£¬£¬£¬£¬¹ØÓÚÄ¿µÄ¶à½ÓÄɵã±ê×¢µÄ·½·¨£¬£¬£¬£¬£¬Õâ¾Í´øÀ´±êעλÖ÷×ÆçÖµÄÎÊÌ⣬£¬£¬£¬£¬ÈçÏÂͼ(a)µÄ»Æµã£¬£¬£¬£¬£¬µãµÄλÖÿÉÄÜÔÚ×ìÉÏ£¬£¬£¬£¬£¬ÑÛ¾¦£¬£¬£¬£¬£¬¶ú¶äµÈ¡£¡£¡£¡£¡£ÄÇôÊÂʵÄÇÀï²ÅÊǸüÓÐÀûÓÚÍøÂçѧϰµÄλÖÃÄØ£¿£¿£¿£¿£¿£¿£¿

b. ÈçÏÂͼ(b)ÔÚ¼à¿ØµÄ÷缯µÄ³¡¾°Ï£¬£¬£¬£¬£¬²»µ«ÔÚ²î±ðµÄ³¡¾°ÖÐÄ¿µÄµÄ±ê×¼²î±ð´ó£¬£¬£¬£¬£¬²¢ÇÒÔÚͳһÕÅͼÖÐÒ²ÓÐÓÉÓÚ͸ÊÓÕ÷ÏóÔì³ÉÄ¿µÄ»áÓÐÖØ´óµÄ±ê׼ת±ä¡£¡£¡£¡£¡£

ÒªÁìÏÈÈÝ

Õë¶ÔÒÔÉÏÌá³öµÄÎÊÌ⣬£¬£¬£¬£¬ÎÒÃÇÌá³öÁËÒ»¸öÐÂÓ±µÄ¼ÆÊý¿ò¼Ü£¬£¬£¬£¬£¬ÈçÏÂͼËùʾ¡£¡£¡£¡£¡£ËüÓÉ×Ô˳ӦÅòÕ;í»ýÍøÂçºÍ×ÔУÕý¼àÊÓ×é³É¡£¡£¡£¡£¡£ÔÚÕâÒ»²¿·Ö£¬£¬£¬£¬£¬ÎÒÃÇÊ×ÏÈ»á´Ó¸ß˹»ìÏýÄ£×Ó£¨GMM£©µÄ½Ç¶ÈÃ÷È·¹Å°åµÄÄ¿µÄÃܶÈͼ£¬£¬£¬£¬£¬È»ºóÎÒÃǽ«ÏÈÈÝÔõÑùʹÓÃÒ»ÖÖÆÚÍû×î´ó»¯(EM)µÄ·½·¨¾ÙÐÐ×Ô¾ÀÕý¸üбêÇ©£¬£¬£¬£¬£¬×îºó½«ÏÈÈÝ×Ô˳ӦÅòÕÍÂʾí»ýµÄÍøÂç½á¹¹ºÍʵÏÖϸ½Ú¡£¡£¡£¡£¡£

1.    ×Ô¾ÀÕýµÄ¼àÊÓ·½·¨

ÄîÍ·£ºËæ×ÅÄ£×ÓµÄѵÁ·µÄ¾ÙÐУ¬£¬£¬£¬£¬·×ÆçÖµĵã±ê×¢»áÓ°ÏìÍøÂçѧϰµÄÉÏÏÞ¡£¡£¡£¡£¡£Í¨¹ýÊӲ췢Ã÷ѧϰһ׼ʱ¼äÒԺ󣬣¬£¬£¬£¬ÍøÂçÕ¹ÍûµÄÃܶÈÔÚÏìӦλÖÃÒ»ÖÂÐÔÉϺÃÓÚÈ˹¤±ê×¢¡£¡£¡£¡£¡£ÒÔÊÇÎÒÃÇÏ£Íûͨ¹ýʹÓÃÍøÂçµÄÕ¹ÍûÀ´¾ÀÕý±ê×¢µÄλÖ㬣¬£¬£¬£¬´Ó¶ø»ñµÃ¸üÒ»ÖÂͬʱ¸üÓÐÀûÓÚÍøÂçѧϰµÄÃܶÈͼ±êÇ©¡£¡£¡£¡£¡£

ÒªÁ죺

½«¸ß˹ÃܶÈͼ¿´×÷Ò»¸ö¸ß˹»ìÏýÄ£×Ó(GMM):

ÆäÖÐDÌåÏÖ¸ß˹ÃܶÈͼ£¬£¬£¬£¬£¬KÌåÏÖÄ¿µÄ¸öÊý£¬£¬£¬£¬£¬xÌåÏÖͼÖеÄλÖÃ

ÕâÀï¿ÉÒÔÓÃÈ˹¤±ê×¢µÄµã×÷Ϊ¾ùÖµ£¬£¬£¬£¬£¬Àο¿ÖµÎª·½²î£¬£¬£¬£¬£¬ÌìÉú¸ß˹»ìÏýÄ£×ӵijõʼÂþÑÜ£¬£¬£¬£¬£¬¶øÍøÂçÕ¹ÍûµÄÃܶÈͼ¿ÉÒÔ½üËÆ¿´×÷ÍøÂçƾ֤ͼÏñÌØÕ÷Õ¹ÍûµÄÒ»¸öÄ¿µÄÂþÑÜ¡£¡£¡£¡£¡£¹þ¹þ(haha)ÌåÓýÒªÁì¾ÍÊÇʹÓÃÍøÂçÕ¹ÍûÀ´ÒÔÒ»ÖÖÀàËÆÆÚÍûÖµ×î´ó»¯£¨EM£©µÄ·½·¨¸üиß˹»ìÏýÄ£×Ó´Ó¶ø»ñµÃÊʺϵıêÇ©¡£¡£¡£¡£¡£

Ïêϸ·½·¨ÈçÏ£º

E°ì·¨£º

ÎÒÃÇʹÓúóÑé¸ÅÂÊÀ´ÆÀ¹ÀµÚλÖú͵ڸß˹ÂþÑܵġ°ÔðÈΡ±¡£¡£¡£¡£¡£

È»ºóÓÃÄ£×ÓÕ¹ÍûµÄÃܶÈͼ¾ÀÕý¡°ÔðÈΡ±¡£¡£¡£¡£¡£

M°ì·¨£º

ÖØÐÂÔ¤¼Æ¸ß˹»ìÏýÄ£×ӵIJÎÊý

ÆäÖÐXÌåÏÖλÖþØÕóËæ×ÅEºÍM½»ÌæÖ´ÐУ¬£¬£¬£¬£¬ÎÒÃÇ»á»ñµÃ¸üÒ»ÖÂÏìÓ¦µÄ±êÇ©¡£¡£¡£¡£¡£ÔÚ¸üеÄÀú³Ì£¬£¬£¬£¬£¬ÓÉÓÚÒÑ֪ÿ¸öÄ¿µÄ¹ØÓÚÕûÌåµÄÂþÑÜÊÇÏàͬµÄ£¬£¬£¬£¬£¬ÒÔÊǹØÓÚÖØÐÂÔ¤¼ÆµÄȨÖØϵÊý¦ÐÎÒÃÇ»áÀο¿Îª1/K.

×Ô¾ÀÕýËðʧº¯Êý£º

Ìá³öµÄ×Ô¾ÀÕýËðʧº¯Êý°üÀ¨Á½¸ö²¿·Ö£¬£¬£¬£¬£¬Ò»¸ö²¿·ÖÊÇÖ±½ÓȫͼºÍ¾ÀÕýºóÃܶÈͼ½ÏÁ¿L1¾àÀ룬£¬£¬£¬£¬Õⲿ·Ö¹Ø×¢ÕûͼÊýÄ¿ÉϵÄÎó²î£¬£¬£¬£¬£¬µÚ¶þ²¿·ÖΪȨÖØϵÊýµÄ¼àÊÓ£¬£¬£¬£¬£¬Ö÷Òª¹Ø×¢¸öÌ壬£¬£¬£¬£¬°ü¹Ü¹ØÓÚÕûÌåµÄТ˳һÖÂ

×ÜÌå¶øÑÔ£¬£¬£¬£¬£¬Ìá³öµÄ×Ô¾ÀÕý¼àÊÓ¾ßÓÐÐí¶àÀíÏëµÄÊôÐÔ¡£¡£¡£¡£¡£ Ê×ÏÈ£¬£¬£¬£¬£¬ËüÄܹ»ÈÝÈ̱ê×¢Îó²î¡£¡£¡£¡£¡£ ¶¯Ì¬¸üÐÂÄ¿µÄÃܶÈͼ¿ÉÒÔ¾ÀÕýijЩ±êÇ©µÄÎó²î£¬£¬£¬£¬£¬ÒÔ×ÊÖúÄ£×Óѧϰµ½Ò»ÖµÄÌØÕ÷±í´ï¡£¡£¡£¡£¡£ Æä´Î£¬£¬£¬£¬£¬¹ØÓÚ·½²îµÄת±äÊdz°ôµÄ¡£¡£¡£¡£¡£ ¿ÉÒÔƾ֤ͼÏñÌØÕ÷½ÓÄɵü´úµ÷½â·½²îÒÔ˳ӦÏìÓ¦ÇøÓò¡£¡£¡£¡£¡£ µÚÈý£¬£¬£¬£¬£¬Ëü¹¤¾ßÊýÄ¿µÄת±äºÜÃô¸Ð¡£¡£¡£¡£¡£ »ìÏýϵÊýµÄ²¨¶¯ÓÐÓõط´Ó¦ÁË©¼ìºÍÎó¼ì¡£¡£¡£¡£¡£ ÏÂÃæչʾÁËÃܶÈͼԤ¼ÆÖеÄËÄÖÖ³£¼ûÇéÐΣ¨²ü¶¶£¬£¬£¬£¬£¬·½²îÔöÌíºÍ¸ß˹ºËµÄת±ä£©×Ô¾ÀÕýµÄ±ÈÕÕ¡£¡£¡£¡£¡£

2.    ×Ô˳ӦÅòÕÍÂÊÍøÂç½á¹¹

ÎÒÃÇ´ÓÁ½¸ö½Ç¶ÈÉè¼ÆÁË×Ô˳ӦÅòÕ;í»ý

1)      ´Ó±ê׼ת±ä·½Ã棬£¬£¬£¬£¬ÎÒÃÇʹÓÃÒ»Á¬µÄ¸ÐÊÜÒ°Ò²À´Æ¥ÅäÒ»Á¬µÄ±ê׼ת±ä¡£¡£¡£¡£¡£

2)      ΪÁËѧϰ¿Õ¼ä¸ÐÖª£¬£¬£¬£¬£¬²î±ðµÄλÖûؽÓÄɲî±ðµÄÅòÕÍÂÊÀ´¾ÙÐвÉÑù¡£¡£¡£¡£¡£

ÏÂͼΪ¹þ¹þ(haha)ÌåÓý×Ô˳ӦÅòÕ;í»ýµÄÀú³Ì£º

°ì·¨1£ºÒÔÏàͬÌØÕ÷ΪÊäÈ룬£¬£¬£¬£¬Í¨¹ý±ê×¼3¡Á3¾í»ý²ã»ñµÃÒ»ÕÅÓëԭͼÏàͬ¾ÞϸµÄµ¥Í¨µÀµÄÅòÕÍÂÊͼ¡£¡£¡£¡£¡£ÌØÊâµØ£¬£¬£¬£¬£¬ÎÒÃÇÌí¼ÓÁËÒ»¸öReLU²ãÀ´°ü¹ÜÅòÕÍÂÊͼÉÏÖµ¶¼Îª·Ç¸ºÊý¡£¡£¡£¡£¡£

°ì·¨2£º¶ÔÌØÕ÷¾ÙÐÐ×Ô˳Ӧ¸ÐÊÜÒ°µÄ²ÉÑù£¬£¬£¬£¬£¬²î±ðλÖõIJÉÑùÍø¸ñ¾ÞϸΪÅòÕÍÂÊͼ¶ÔӦλÖõÄÖµ£¬£¬£¬£¬£¬Õâ¸öÖµ¿ÉÄÜ»áÊÇСÊý£¬£¬£¬£¬£¬ÕâÀïÎÒÃǽÓÄÉÁËË«ÏßÐÔ²åÖµ¾ÙÐвÉÑù

°ì·¨3£º¶Ô²ÉÑùÖµ¾ÙÐмÓȨÇóºÍ»ñµÃеÄÌØÕ÷

¹þ¹þ(haha)ÌåÓý×Ô˳ӦÅòÕÍÂʾí»ý²»ÐèÒªÌØÁíÍâ±ê×¼±êÇ©£¬£¬£¬£¬£¬Ö»ÐèÒª×îºóµÄÃܶȱêÇ©¾Í¿ÉÒÔÈÃÍøÂç×Ô¼ºÑ§Ï°Ë³Ó¦²î±ð±ê×¼µÄÄ¿µÄ¡£¡£¡£¡£¡£Í¬Ê±Ïà½ÏÁ¿Ðαä¾í»ý[1]£¬£¬£¬£¬£¬¹þ¹þ(haha)ÌåÓý²ÉÑùÍø¸ñÖÐÊÇÍêÈ«¶Ô³ÆµÄ£¬£¬£¬£¬£¬²ÉÑùµÄÌØÕ÷²»»áÓÐÏà¶ÔÄ¿µÄλÖÃÉϵÄÎó²î£¬£¬£¬£¬£¬ºÍ×îÖÕÄ¿µÄµÄλÖÃÓиüºÃµÄÒ»ÖÂÐÔ£¬£¬£¬£¬£¬Ô½·¢ÊʺϼÆÊýÕâÖÖλÖÃÃô¸ÐÐÔµÄʹÃü¡£¡£¡£¡£¡£

ʵÑéЧ¹û

ÏÂͼΪ¿ÉÊÓЧ¹ûµÄ±ÈÕÕ£¬£¬£¬£¬£¬¿ÉÒÔ¿´³ö£¬£¬£¬£¬£¬ADSCNetÏà½ÏÁ¿¹Å°åµÄ¼àÊÓÕ¹ÍûµÄÃܶÈͼÖ÷ÒªÓÐÁ½·½ÃæµÄÓÅÊÆ£º1.²î±ðÄ¿µÄ¸üÒ»ÖµÄÏìӦǿ¶È 2.²î±ðÄ¿µÄÏìÓ¦µÄλÖÃÔ½·¢Ò»Ö¡£¡£¡£¡£¡£ÏìÓ¦µãÖ÷Òª¼¯ÖÐÍ·²¿µÄ×óÉϽÇÂÖÀª´¦£¬£¬£¬£¬£¬Åú×¢ÎúÏà¹ØÓÚÈ˹¤±ê×¢µÄÑÛ¾¦£¬£¬£¬£¬£¬±Ç×ӵȣ¬£¬£¬£¬£¬Í·²¿ÂÖÀªÊÇÏà¶Ô¸ü½ûÖ¹Ò×ÕÚµ²£¬£¬£¬£¬£¬¸üÊʺϼÆÊýʹÃüµÄÌØÕ÷µã¡£¡£¡£¡£¡£Í¨¹ýÏÂͼµÚËÄÁпÉÒÔ¿´µ½ÕûÌåÉÏ´óµÄÄ¿µÄÐèÒª¸ü´ó¸ÐÊÜÒ°£¬£¬£¬£¬£¬Ò»Ð©ÓÐÓïÒåµÄÅä¾°Ä¿µÄÒ²ÐèÒª¸ü´óµÄ¸ÐÊÜÒ°È¥Çø·Ö¡£¡£¡£¡£¡£

ͬʱÎÒÃÇÒ²¾ÙÐÐÏûÈÚʵÑéµÄ±ÈÕÕ£¬£¬£¬£¬£¬Ê×ÏÈÎÒʵÑéÁËÓÐÓõÄÊý¾ÝÔöÌí·½·¨£¬£¬£¬£¬£¬¼ÓÈëBNºÍÔö´óbatchsizeÀ´È·Á¢ÐµÄBaseline¡£¡£¡£¡£¡£ÎÒÃÇÕâÀ︴ÏÖÁËCSRNet[2]ºÍMCNN[3]×÷ΪBaselineÒªÁì¾ÙÐнÏÁ¿£¬£¬£¬£¬£¬ÈçÏÂͼÊ×ÏÈÊÇ×Ô¾ÀÕý¼àÊÓµÄЧ¹û¡£¡£¡£¡£¡£×Ô˳Ӧ¼àÊÓÔÚÈý¸öbaselineÉÏÈ¡µÃÁËÒ»ÖµÄÌáÉý¡£¡£¡£¡£¡£ ËûÃÇÏà¶ÔµÄMAEÌáÉý»®·ÖΪ6.19£¥£¬£¬£¬£¬£¬8.57£¥£¬£¬£¬£¬£¬8.72£¥¡£¡£¡£¡£¡£

¶ø×Ô˳ӦÅòÕ;í»ý·½Ã棬£¬£¬£¬£¬ÎÒ±ÈÕÕÁ˲î±ðµÄÀο¿ÅòÕÍÂʺͶàÁÐÍøÂç×éºÏÒÔ¼°Ðαä¾í»ýµÄЧ¹û¡£¡£¡£¡£¡£Ïà½ÏÁ¿Àο¿µÄÅòÕÍÂÊ£¬£¬£¬£¬£¬ÎÒÃÇÖ»ÔöÌíÁËÓÐÏÞµÄÔËË㣬£¬£¬£¬£¬È´È¡µÃÁËÏÔ×ŵÄÌáÉý¡£¡£¡£¡£¡£

×îºóºÍÄ¿½ñSOTAµÄ±ÈÕÕ£¬£¬£¬£¬£¬ADSCNetÔÚËĸö¹ûÕæÊý¾Ý¼¯È¡µÃ¸üÓŵÄÌåÏÖ£¬£¬£¬£¬£¬²¢ÓÐ×ÅÏÔ×ŵÄÌáÉý£¬£¬£¬£¬£¬Åú×¢ÎúÎÒÃÇÒªÁìµÄÓÐÓÃÐÔ¡£¡£¡£¡£¡£

½áÓï

ÔÚ±¾ÎÄÖУ¬£¬£¬£¬£¬ÎÒÃÇΪ¼ÆÊýÎÊÌâÌá³öÁËÒ»ÖÖÐÂÓ±µÄ¼àÊÓѧϰ¿ò¼Ü¡£¡£¡£¡£¡£ËüʹÓÃÄ£×ÓÔ¤¼ÆÀ´µü´úµØ¾ÀÕýGT£¬£¬£¬£¬£¬²¢Ìá³ö×Ô¾ÀÕýËðʧº¯Êýͬʱ¼àÊÓÕûÌåµÄÊýÄ¿ºÍ¸öÌåµÄÂþÑÜ¡£¡£¡£¡£¡£Í¬Ê±ÕâÖÖÒªÁì¿ÉÒÔÓ¦Óõ½ËùÓлùÓÚCNNµÄÒªÁìÖС£¡£¡£¡£¡£ÁíÒ»·½Ã棬£¬£¬£¬£¬ÎÒÃÇÌá³öÁË×Ô˳ӦÅòÕ;í»ý£¬£¬£¬£¬£¬Ëüͨ¹ýÿ¸öλÖõĶ¯Ì¬µØѧϰ²î±ðµÄÅòÕÍÂÊÒÔ˳ӦĿµÄÖØ´óµÄ±ê׼ת±ä¡£¡£¡£¡£¡£ÔÚËĸöÊý¾Ý¼¯ÉϾÙÐеÄʵÑéÅú×¢£¬£¬£¬£¬£¬Ëü¿ÉÒÔÏÔÖøÌáÉý¼ÆÊýÍøÂçµÄÐÔÄÜ¡£¡£¡£¡£¡£ ͬʱҲ˵Ã÷ÎúʹÓÃÄ£×Ó´ÓͼÏñÌØÕ÷ÉÏѧϰµÄÐÅÏ¢Äܹ»±»ÓÃÓÚ¾ÀÕý±ê×¢À´ÌáÉýÐÔÄÜ¡£¡£¡£¡£¡£

References

[1] Dai, Jifeng, et al. Deformable convolutional networks. In ICCV, 2017.

[2] Li, Yuhong, Xiaofan Zhang, and Deming Chen. Csrnet: Dilated convolutional neural networks for understanding the highly congested scenes. In CVPR, 2018.

[3] Zhang, Yingying, et al. Single-image crowd counting via multi-column convolutional neural network. In CVPR, 2