Video Motion (ICCV 2013)S. Ricco and C. Tomasi. Video motion for every visible point. ICCV 2013. [pdf] Video motion describes motion in an image sequence by computing sequence-length paths for a dense sampling of the visible surfaces in a scene. A single path lists the location where a scene point would appear if it were visible in every frame of the sequence. Each path is coupled with a vector of binary visibility flags for the associated point that identify the frames in which the tracked point is unoccluded. Please see the paper for details on the process of inferring video motion paths for a given sequence. The videos below compare the quality of the motion representation we extract with LDOF trajectories [1] and LME paths [2]. Each video shows the input video sequence in the leftmost panel, followed by three panels with each frame warped to align with the first, middle, or last frame of the sequence, respectively. With perfect motion estimates, the three warps should appear to be static replicas of the target frame. The only moving portion should be the white region that we use to flag pixels that have been marked as occluded. These results are discussed in more detail in Susanna Ricco's Ph.D. thesis. [pdf]
[1] N. Sundaram et al. Dense point trajectories by GPU-acclerated large
displacement optical flow. ECCV 2010. View results for: Flowerbed, Truck, Marple1, Marple7, Marple8. Flowerbed LDOF trajectories [1]: LME paths [2]: Video motion (new): |