±à¼ÍƼö: |
±¾ÎÄÀ´×Ôcsdn£¬±¾ÎÄÖ÷Òª½éÉÜÁËÄ¿±ê¼ì²âËã·¨ºÍÎïÌ幨¼üµã¼ì²âµÄÓ¦Óó¡¾°ÒÔ¼°Î»Öüì²âµÄËã·¨ÌØµã¡£ |
|
Ä¿±ê¼ì²âËã·¨
Ä¿±ê¼ì²â¸ÅÄî
Ä¿±ê¼ì²âÕâÀï²ûÊöÁ½¸öÓ¦Óó¡¾°£¬1 ΪÎïÌåλÖüì²â£¬2 ΪÎïÌ幨¼üµã¼ì²â¡£
1 ÎïÌåλÖüì²â
Ïà±ÈÓëͼƬ·ÖÀ࣬Ŀ±ê¼ì²âËã·¨½á¹ûÒªÇó²»½öʶ±ð³öͼƬÖеÄÎïÀíÀà±ð²¢ÇÒÊä³öÎïÌåµÄλÖòÎÊý¡£
ÎïÌåµÄλÖÃͨ¹ýbounding box±íʾ¡£bounding boxÖ¸µÄÊÇÄܹ»¿ò³öÎïÌåµÄ¾ØÐοòÔÚͼƬÖеĶÔÓ¦²ÎÊý£¬ÈçÏÂͼËùʾ¡£

ͼÖкìÉ«¿ò£¬¾Í³ÆÎªÒ»¸öbounding box£¬bounding
boxÓɸö²ÎÊýÀ´ÃèÊö[bx,by,bh,bw] [b_x, b_y, b_h, b_w][b x
,b y ,b h ,b w ]£¬ ÆäÖÐ(bx,by b_x, b_yb x ,b y )¶ÔÓ¦¾ØÐοòµÄÖÐÐĵã×ø±ê£¬bh
b_hb h Ôò¶ÔÓ¦¾ØÐοòµÄ¸ß¶È£¬bw b_wb w Ôò¶ÔÒ»¸ö¾ØÐοòµÄ¿í¶È¡£ÈçÏÂͼËùʾ¡£

´Ó¶ø bx=0.5,by=0.7,bh=0.3,bw=0.4 b_x=0.5, b_y=0.7, b_h=0.3,
b_w=0.4b
½«Í¼Æ¬×÷ΪģÐ͵ÄÊäÈ룬 Êä³öΪÎïÌåµÄÀà±ðºÍbounding box²ÎÊý¡£ÈçÏÂͼËùʾ¡£

ÕâÀïÎÒÃǼÙÉèµÄÊÇÒ»¸öͼƬÖÐÖ¸¶¨ÎªÒ»¸öÎïÌåµÄÀà±ð¼°Î»Öá£
Îâ¶÷´ïÀÏʦµÄ¿Î³ÌÖÐÕâÑù¶¨ÒåÄ£ÐÍÊä³öy=[pc,bx,by,bh,bw,c1,c2,c3]T \hat{y}=[p_c,
b_x, b_y, b_h, b_w, c_1, c_2, c_3]^T
y=[pc,bx,by,bh,bw,c1,c2,c3]T \hat{y}=[p_c, b_x, b_y,
b_h, b_w, c_1, c_2, c_3]^T
ÆäÖУ¬pc p_cp ΪͼÖÐÓÐÎïÌåµÄ¸ÅÂÊ;
[bx,by,bh,bw] [b_x, b_y, b_h, b_w][b
]ÔòÊǶÔÓ¦ÎïÌåµÄλÖÃ(bounding box)²ÎÊý;
c1,c2,c3 c_1, c_2, c_3c ÔòÊǶÔÓ¦ÄĸöÎïÌåµÄclass¡£ÉÙÒ»¸öc4
c_4c ÊÇÒòΪc1,c2,c3 c_1, c_2, c_3c ¾ùΪ0ʱ£¬Ôòc4=1 c_4=1c
2 ÎïÌ幨¼üµã¼ì²â(landmark detection)
ÎïÌ幨¼üµã¼ì²âµÄÓ¦Óó¡¾°ÓУ¬ÈËÁ³¹Ø¼üµã(ÑÛ¾¦¡¢±Ç×Ó¡¢×ì¼°Á³µÄ±ßÔµµã)µÄ¼ì²â£¬ÈËÌå×Ë̬¼ì²â(Í·£¬ ¸ì²²¼°Íȵĸ÷¸ö¹Ø¼üµãµÄ¼ì²â)£¬»¹ÓбÈÈçÌì³ØµÄÒ·þ¹Ø¼üµã¼ì²âµÈ¶¼ÊÇÊôÓÚÕâ¸ö·¶³ëµÄÓ¦Óó¡¾°¡£
ÊäÈëΪͼƬ£¬Êä³öÔòÊÇÕâһϵÁйؼüµãµÄ×ø±êλÖÃ[[l1x,l1y],[l2x,l2y],...,[lnx,lny]]
[[l_{1x},l_{1y}], [l_{2x},l_{2y}], ... , [l_{nx},l_{ny}]][[l
¼´landmark¡£


ÎïÌåλÖüì²âËã·¨
±¾ÎÄÖ»ÊÇ´óÖ½éÉÜλÖüì²âµÄËã·¨ÌØµã¡£
ÆäʵRCNN, Fast-rcnn, SPP-net, Faster-rcnn¶¼¿ÉÒÔ·ÖΪÁ½¸öÖ÷Òª²¿·Ö£º1.
region proposal ÌáÈ¡regions£»¶øYOLO¼°SSDÔòͨ¹ý·Ö¸ñ¼°anchorµÄ·½·¨À´´úÌæ´óÅúÁ¿µÄregions£¬Ïà±ÈYOLO¼°SSD¼ÆËãÁ¿Ð¡£¬´Ó¶ø¿ìһЩ¡£
½ÓÏÂÀ´Áгöÿ¸öÍøÂçµÄÖ÷ÒªÌØµã¡£
RCNN½éÉÜ
Ïà¹Ø¸ÅÄî½éÉÜ£º
regions¾ÍÊÇ´ÓÔͼÏñÉϽØÈ¡µÄÓпÉÄܰüº¬ÎïÌåµÄͼÏñÇøÓò¼¯ºÏ¡£
bounding boxÊÇÖ¸°üº¬ÎïÌåµÄ¾ØÐοò¡£ÎÄÖÐÇ°ÃæÓнéÉÜ¡£
IOU(Intersection over Union)ÓÃÀ´ºâÁ¿Á½¸öregionsÖ®¼äµÄÖØµþ¶È¡£ÉÏͼ¡£

¾ØÐοòA¡¢BµÄÒ»¸öÖØºÏ¶ÈIOU¼ÆË㹫ʽΪ£º
IOU=(A¡ÉB)/(A¡ÈB)
¾ÍÊǾØÐοòA¡¢BµÄÖØµþÃæ»ýÕ¼A¡¢B²¢¼¯µÄÃæ»ý±ÈÀý:
IOU=SI/(SA+SB-SI)
NMS(¼«´óÖµÒÖÖÆ)ÊÇÌô³ö²»Í¬ÀàÖеÄ×î´óÖµ£¬ÉáÆúÏà½üµÄµ«ÊDz»ÊǼ«´óµÄÖµ¡£
ÔÚÄ¿±êʶ±ðÖУ¬ÓÃÓÚ×îºóµÄbounding boxɸѡ¡£¾¹ýSVMÅбðµÄbounding boxes£¬°üº¬ÁËËùÓÐÀà±ðµÄbounding
boxes£¬Í¨¹ýNMSʵÏÖ£¬×îºóÁôϲ»Í¬ÀàµÄ×îÓŵÄbounding box¡£ÏêϸµÄʵÏÖÔÀí²Î¿¼²©¿Í
RCNN½á¹¹Ö÷Òª²½Ö裺
ͼÏñ¶ÔÓ¦µÄºòÑ¡ÇøÓò(1K~2K¸ö)Éú³É£»
ºòÑ¡ÇøÓòµÄÌØÕ÷ÌáÈ¡£»
ÌØÕ÷·ÖÀà(SVM)£»Áôϰüº¬ÎïÌåµÄregions£¬´Ó¶ø×÷Ϊbounding boxes¡£
bounding boxɸѡ£¬²¢ÐÞÕý¡£
¸÷»·½ÚʹÓü¼Êõ£º
ºòÑ¡ÇøÓòÉú³É
ʹÓÃSelective SearchÉú³ÉͼÏñµÄºòÑ¡ÇøÓò¡£
´óÌåÔÀí£º
ÊÇÒ»ÖÖͼÏñ·Ö¸ñµÄÊֶΣ¬Í¨¹ýºÏ²¢ÏàËÆÑÕÉ«»òÎÄÀíÇÒÏàÁÚµÄÇøÓò£¬ÊµÏÖͼÏñ·Ö¸î£¬½«×îºóµÄÇøÓòÈ¥ÖØÊä³ö¡£Selective
SearchÔ´ÂëµØÖ·¡£
ÌØÕ÷ÌáÈ¡
ʹÓÃѵÁ·ºÃµÄÉñ¾ÍøÂçÌáÈ¡regionsÌØÕ÷¡£
RCNNѵÁ·ÌØÕ÷Ö÷Òª¾¹ý2¸ö½×¶Î£º
Pre-train£º
ʹÓÃILVCR 2012Êý¾Ý¼¯¼°¼ò»¯°æµÄHinton 2012ÄêÔÚImage NetÉϵķÖÀàÍøÂçÀ´½øÐÐԤѵÁ·¡£(È«Á¬½Ó²ãÌáÈ¡ÌØÕ÷4096ά£¬ÔÙʹÓÃÈ«Á¬½Ó(4096->1000)ʵÏÖ1000Àà·ÖÀà)¡£
Fine-tune:
Ìæ»»Pre-trainµÄ×îºóÊä³ö²ã£¬»»Îª(4096->21)21·ÖÀàµÄÊä³ö²ã£¬Ê¹ÓÃÊý¾Ý¼¯PASCAL
VOC 2007À´ÑµÁ·ÍøÂç¡£´Ë´¦ÑµÁ·µÄÕý¸ºÑù±¾µÄ±ê¶¨£ºIOU>0.5ÔòΪÕýÑù±¾¡£
ÌØÕ÷·ÖÀà
ʹÓÃһϵÁÐSVMÀ´·ÖÀྐྵýÍøÂçÌáÈ¡µ½µÄÌØÕ÷£¬SVM¸öÊýµÈÓÚÒª±ê³öµÄÎïÖÖÊý¡£²¢ÇÒʹÓÃÁËhard negative
mining¡£
SVMѵÁ·Ê±£¬Õý¸ºÑù±¾±ê¶¨£ºIOU>0.3ÔòΪÕýÑù±¾¡£
ÔÚ²âÊÔʱ£¬SVMÊä³öΪÕýµÄregions×÷Ϊ´ýɸѡµÄbounding boxes¡£
bounding boxɸѡ
ʹÓü«´óÖµÒÖÖÆ(NMS)¡£É¸Ñ¡µÃµ½Ã¿Ò»Àà¶ÔÓ¦µÄ×îÓŵÄbounding box¡£
×îºóͨ¹ýbounding box»Ø¹é¾«Ï¸ÓÅ»¯bounding box¡£
bounding box»Ø¹é¾ÍÊÇͨ¹ýÉñ¾ÍøÂçÄâºÏÆ½ÒÆºÍ·ÅËõÁ½¸öº¯Êý£¬Ê¹ÓÃÕâÁ½¸öº¯Êýʹ¸ü¼ÓÄâºÏÕæÊµµÄλÖÃÇøÓò¡£loss¾ÍÊÇÔʼboxÓëÄ¿±êboxÖ®¼äµÄ¾àÀë¡£
RCNNÍøÂçµÄȱµã
ѵÁ·ÐèÒª¼¸¸ö½×¶Î£»
SSµÃµ½µÄregionsÊDz»Í¬³ß´çµÄ¡£²»Í¬³ß´çµÄregionsÐèÒª¾¹ý£¬²Ã¼ô¼°·ÅËõµ½ÏàͬµÄ³ß´ç£¬ÕâÑù»áʹһЩÄÚÈݶªÊ§¡£
ÿ¸öregionµÄ¶¼ÐèÒª¾¹ýÄ£ÐÍÈ¥ÌáÈ¡£¬²¢´æ·ÅÖÁ´ÅÅÌ£»
SPP-net½éÉÜ
SPP-net(Spatial Pyramid Pooling) Ìá³öÖ÷Òª½á¹¹£º
Ìá³ö½ð×ÖËþ½á¹¹£¬Ê¹Ä£ÐÍ¿ÉÒÔ´¦Àí²»Í¬³ß´çµÄÊäÈë¡£
ʹÓÃregionsµ½ÌØÕ÷²ãÖ®¼äµÄÓ³É䣬´Ó¶øÊµÏÖÖ»Ìáȡһ´ÎÍê³ÉµÄͼÏñ£¬regionsÌØÕ÷ͨ¹ýÓ³Éä¹ØÏµÀ´»ñÈ¡¡£
SPP-netµÄȱµã£º
½ð×ÖËþ½á¹¹ËäȻʹµÃÄ£ÐÍ´¦Àí²»Í¬³ß´çÊäÈëµÄÎÊÌ⣬µ«Í¬Ê±Ê¹Ä£ÐͲ»Í¬Í¬Ê±ÑµÁ·½ð×ÖËþ×ó²àµÄ¾í»ý²ã¡£ÖÁÓÚÔÒò£¬×÷Õ߸ø³öµÄÔÒòÊǽð×ÖËþµÄpooling²ã¶ÔÓ¦µÄ¸ÐÊÜÒ°Ì«´ó¡£±¾È˾õµÃ¸ÐÊÜÒ°Ì«´ó£¬»áʹÌݶȴ«µÝ²»Îȶ¨¡£
Ïêϸ½á¹¹²»×÷½éÉÜ¡£
ѵÁ·ÒÀ¾ÉÐèÒª¼¸¸ö½×¶Î¡£ÒÀ¾É²ÉÓÃSVMʵÏÖ·ÖÀà¡£
SPP-netÖм¼Êõµã½²½â

1. ½ð×ÖËþ½²½â

ʵ¼ÊÔÀí¾ÍÊÇ£¬¶ÔÿͨµÀµÄÌØÕ÷½øÐзֿ飬·Ö³É16*16, 4*4, 1*1¿é£¬Ìô³öÿ¿éµÄ×î´óÖµ£¬µÃµ½16*16,
4*4, 1*1ÌØÕ÷£¬Õ¹¿ª³É16*16 + 4*4 + 1*1ΪÏòÁ¿£¬Êä¸øÈ«Á¬½Ó²ã¡£
2. regionsµ½ÌØÕ÷²ãÖ®¼äÊÇÈçºÎÓ³Éä

ͼÖÐÊäÈë²ãÀ¶É«¼ÓÉ·Ö¶ÔӦÿһ²ãµÄÇøÓò£¬Ò»¸öÇøÓò¶ÔÓ¦¾¹ý¾í»ý/poolingºóµÄ¶ÔÓ¦ÇøÓò³ß´ç¼ÆËã·½·¨£¬ÓëÔÍ¼ÌØÕ÷¾¹ý¾í»ý/poolingºóµÄ¶ÔÓ¦³ß´çÏàͬ¡£×¢£º½ö½öÊǽèÓÃÕâ¸öͼ±íʾһÏÂregionsÊäÈëÓëÆäËû²ãÖ®¼äµÄ¶ÔÓ¦¹ØÏµ¡£¸ÃͼÊÇÔÒâÖ¼ÔÚ˵ÀûÓÃ1x1¾í»ýºËʵÏÖÈ«¾í»ý¡£
Fast-RCNN½éÉÜ£º
Fast-RCNNÓÅ»¯¼¼Êõµã

Ìá³öROI pooling layer£¬²ÉÓõ¥²ã½ð×ÖËþ½á¹¹£¬Æäʵ¾ÍÊÇÔÚÌØÕ÷²ãֻʹÓÃÒ»¸ö½ð×ÖËþmax-pooling£¬½øÒ»²½¼ò»¯ÁËregions¶ÔÓ¦ÌØÕ÷²ãµÄÓ³Éä¹ØÏµ¡£
Ìá³öÌݶȴ«µÝ·½·¨£¬ÊµÏÖÕû¸öÍøÂçÍøÂç½á¹¹µÄÈ«²¿ÑµÁ·¡£
ÔÚÁ½²ãÈ«Á¬½ÓÖмÓÈëSVD½µÎ¬£¬¼Ó¿ìѵÁ·ËÙ¶È¡£
Êä³öʹÓÃÁ½¸ösoftmax£¬Ò»¸öÓÃÓÚclass·ÖÀ࣬һ¸öÓÃÓÚbounding box»Ø¹é¡£
Fast-RCNNÓÅ»¯¼¼Êõµã½²½â
ROI pooling layer½²½â£º
POI pooling layer ²ãÊÇλÓÚ¾í»ýºó£¬È«Á¬½Ó֮ǰµÄlayer¡£
ÊäÈëΪ¾í»ý²ãÊä³ö(H*W*N)¼°R¸öROIÔª×飬R±íʾregionsµÄ¸öÊý¡£Ã¿¸öROIÊÇÒ»¸öÔª×é(n,
r, c, h, w)£¬nÊÇÌØÕ÷Ó³ÉäµÄË÷Òý£¬n¡Ê{0, ¡ ,N-1}£¬(r, c)ÊÇRoI×óÉϽǵÄ×ø±ê£¬(h,
w)ÊǸßÓë¿í¡£
Êä³öÊÇmax-pool¹ýµÄÌØÕ÷Ó³É䣬H¡¯ x W¡¯ x CµÄ´óС£¬H¡¯¡ÜH£¬W¡¯¡ÜW¡£½ð×ÖËþµÄµÄÇиî²ÎÊý£ºbin-size
~ h/H¡¯ x w/W¡¯£¬ÕâÑù¾ÍÓÐH¡¯ x W¡¯¸ö£¬binµÄ´óСÊÇ×ÔÊÊÓ¦µÄ£¬È¡¾öÓÚRoIµÄ´óС¡£
Faster-RCNN½éÉÜ£º
Faster-RCNNÓÅ»¯¼¼Êõµã
Ìá³öRPN(Region Proposal Networks)ÇøÓòÉú³ÉÍøÂ磬ʹÓÃÉñ¾ÍøÂçÉú³Éregions£¬´úÌæRCNNÖеÄSelective
Search·½·¨¡£½ÚÊ¡regions proposalµÄʱ¼ä¡£»ù±¾ÊµÏÖend to endѵÁ·¡£

RPN´óÖÂÔÀí½éÉÜ£º
ÔÚ×îºóµÄ¾í»ý²ã(¼´ÌØÕ÷ÌáÈ¡²ã)ÉÏʹÓû¬¶¯´°¿ÚÀ´Ô¤²â¡£Ê¹ÓÃ3x3µÄ¾í»ýºË¶ÔµÃµ½µÄ256άµÄÌØÕ÷ͼ½øÐ묶¯¾í»ý£¬·ÖÁ½Â··Ö±ðʹÓÃÀûÓÃ1*1/¾í»ý£¬×îºóһ·Êä³öËùÓÐanchorsµÄÄ¿±êºÍ·ÇÄ¿±ê(±³¾°)µÄ¸ÅÂÊ£¬Áíһ·Êä³öanchors
boxÏà¹ØµÄËĸö²ÎÊý£¬°üÀ¨boxµÄÖÐÐÄ×ø±êxºÍy£¬box¿íwºÍ³¤h¡£Ã¿´Î»¬¶¯¾í»ý£¬Êä³ök¸öanchorsµÄÊÇ·ñ°üº¬ÎïÌåÒÔ¼°Î»ÖÃÐÅÏ¢¡£Òò´Ë×îºóµÄRPNµÄÊä³öÊÇһ·Ϊ2k¶ÔÓ¦·ÖÀà(ÊÇ·ñ°üº¬ÎïÌå)£¬Áíһ·Ϊ4k¶ÔÓ¦anchorsµÄλÖÃÐÅÏ¢¡£

×÷Õß²ÉÓÃËIJ½ÑµÁ··¨£º
1£© µ¥¶ÀѵÁ·RPNÍøÂç£¬ÍøÂç²ÎÊýÓÉԤѵÁ·Ä£ÐÍÔØÈ룻
2£© µ¥¶ÀѵÁ·Fast-RCNNÍøÂ磬½«µÚÒ»²½RPNµÄÊä³öºòÑ¡ÇøÓò×÷Ϊ¼ì²âÍøÂçµÄÊäÈë¡£¾ßÌå¶øÑÔ£¬RPNÊä³öÒ»¸öºòÑ¡¿ò£¬Í¨¹ýºòÑ¡¿ò½ØÈ¡ÔͼÏñ£¬²¢½«½ØÈ¡ºóµÄͼÏñͨ¹ý¼¸´Îconv-pool£¬È»ºóÔÙͨ¹ýroi-poolingºÍfcÔÙÊä³öÁ½Ìõ֧·£¬Ò»ÌõÊÇÄ¿±ê·ÖÀàsoftmax£¬ÁíÒ»ÌõÊÇbbox»Ø¹é¡£½ØÖ¹µ½ÏÖÔÚ£¬Á½¸öÍøÂ粢ûÓй²Ïí²ÎÊý£¬Ö»ÊÇ·Ö¿ªÑµÁ·ÁË£»
3£© ÔÙ´ÎѵÁ·RPN£¬´Ëʱ¹Ì¶¨ÍøÂ繫¹²²¿·ÖµÄ²ÎÊý£¬Ö»¸üÐÂRPN¶ÀÓв¿·ÖµÄ²ÎÊý£»
4£© ÄÇRPNµÄ½á¹ûÔÙ´Î΢µ÷Fast-RCNNÍøÂ磬¹Ì¶¨ÍøÂ繫¹²²¿·ÖµÄ²ÎÊý£¬Ö»¸üÐÂFast-RCNN¶ÀÓв¿·ÖµÄ²ÎÊý¡£
YOLOËã·¨½éÉÜ
YOLO¼¼Êõ´óÖÂÔÀí
½«Ò»·ùͼÏñ·Ö³ÉSxS¸öÍø¸ñ(grid cell)£¬Èç¹ûij¸öobjectµÄÖÐÐÄ
ÂäÔÚÕâ¸öÍø¸ñÖУ¬ÔòÕâ¸öÍø¸ñ¾Í¸ºÔðÔ¤²âÕâ¸öobject¡£

ÿ¸öÍø¸ñÒªÔ¤²âÂäÔÚ¸ÃÍø¸ñµÄÎïÌåÀà±ð¼°B¸öbounding boxÐÅÏ¢£¬Ã¿¸öbounding boxÔ¤²âÐÅÏ¢°üº¬×ÔÉíλÖÃÐÅÏ¢(Ëĸö±äÁ¿)ºÍconfidenceÖµ¡£
confidence´ú±íÁËËùÔ¤²âµÄboxÖк¬ÓÐobjectµÄÖÃÐŶȺÍÕâ¸öboxÔ¤²âµÄÓжà×¼(Óë±ê×¢µÄground
regionµÄIOU)Á½ÖØÐÅÏ¢¡£

ÆäÖÐÈç¹ûÓÐobjectÂäÔÚÒ»¸ögrid cellÀµÚÒ»ÏîÈ¡1£¬·ñÔòÈ¡0¡£ µÚ¶þÏîÊÇÔ¤²âµÄbounding
boxºÍʵ¼ÊµÄgroundtruthÖ®¼äµÄIOUÖµ¡£
ÿ¸öbounding boxÒªÔ¤²â(x, y, w, h)ºÍconfidence¹²5¸öÖµ£¬Ã¿¸öÍø¸ñ»¹ÒªÔ¤²âÒ»¸öÀà±ðÐÅÏ¢£¬¼ÇΪCÀà¡£ÔòSxS¸öÍø¸ñ£¬Ã¿¸öÍø¸ñÒªÔ¤²âB¸öbounding
box»¹ÒªÔ¤²âC¸öcategories¡£Êä³ö¾ÍÊÇS x S x (5*B+C)µÄÒ»¸ötensor¡£
×¢Ò⣺classÐÅÏ¢ÊÇÕë¶Ôÿ¸öÍø¸ñµÄ£¬confidenceÐÅÏ¢ÊÇÕë¶Ôÿ¸öbounding boxµÄ¡£
¾ÙÀý˵Ã÷: ÔÚPASCAL VOCÖУ¬Í¼ÏñÊäÈëΪ448x448£¬È¡S=7£¬B=2£¬Ò»¹²ÓÐ20¸öÀà±ð(C=20)¡£ÔòÊä³ö¾ÍÊÇ7x7x30µÄÒ»¸ötensor¡£
ÔÚtestµÄʱºò£¬Ã¿¸öÍø¸ñÔ¤²âµÄclassÐÅÏ¢ºÍbounding
boxÔ¤²âµÄconfidenceÐÅÏ¢Ïà³Ë£¬¾ÍµÃµ½Ã¿¸öbounding boxµÄclass-specific
confidence score:

|