实验室交流课件.ppt

SMART SMART Overview * Moving Object Proposals(MOPs) Compute the optical flow Brox T, Malik J. Large displacement optical flow: descriptor matching in variational motion estimation[J]. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2011, 33(3): 500-513. * Compute optical flow boundries Dollár P, Zitnick C L. Structured forests for fast edge detection[C] //Computer Vision (ICCV), 2013 IEEE International Conference on. IEEE, 2013: 1841-1848. Moving Object Proposals(MOPs) Compute multiple figure-ground segmentations Kr?henbühl P, Koltun V. Geodesic object proposals[M]// Computer Vision–ECCV 2014. Springer International Publishing, 2014: 725-739. * Moving Object Proposals(MOPs) * Overview * Moving Objectness Detector(MOD) CNN – a dual-pathway architecture Krizhevsky A, Sutskever I, Hinton G E. Imagenet classification with deep convolutional neural networks[C] //Advances in neural information processing systems. 2012: 1097-1105. * Moving Objectness Detector(MOD) CNN – Initialization Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C] //Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on. IEEE, 2014: 580-587. * CNN – Training using “Caffe” Jia Y, Shelhamer E, Donahue J, et al. Caffe: Convolutional architecture for fast feature embedding[C]//Proceedings of the ACM International Conference on Multimedia. ACM, 2014: 675-678. Overview * Tube proposal generation Per frame MOPs to spatio-temporal tubes Grady L. Random walks for image segmentation[J]. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2006, 28(11): 1768-1783. * Tube proposal generation * Experiments * Moseg - Inaccurate mapping of trajectory clusters to pixel tubes often slightly leak to the background, especially for animals with thin limbs Discussion – Failure cases * VSB100 - Large motion or full object occlusion an additional linking step, where similarly looking tubes are linked a