深度学习及其应用_ 课件 22-卷积网络视觉应用下.pptxVIP

ComputervisiontasksDeepLearninganditsApplicationSJTUDeepLearningLecture.1

SJTUDeepLearningLecture.2R-CNN:ProblemsAdhoctrainingobjectivesFine-tunenetworkwithsoftmaxclassifier(logloss)Trainpost-hoclinearSVMs(hingeloss)Trainpost-hocbounding-boxregressions(leastsquares)Trainingisslow(84h),takesalotofdiskspaceInference(detection)isslow47s/imagewithVGG16[SimonyanZisserman.ICLR15]FixedbySPP-net[Heetal.ECCV14]Girshicketal,“Richfeaturehierarchiesforaccurateobjectdetectionandsemanticsegmentation”,CVPR2014.SlidecopyrightRossGirsh