ECCV 2016 Daily - Tuesday

Karel Lebeda, Centre for Vision, Speech and Signal Processing, University of Surrey - UK Direct-from-Video: Unsupervised NRSfM 8 ECCV Daily: Tuesday Presentations ECCV Daily: Karel, what is your work about? Karel Lebeda: It’s about non-rigid structure from motion directly from video. It’s on performing simultaneous tracking of an object and reconstructing its non-rigid shape. I prepared it during my PhD with my supervisors, Simon Hadfield, Richard Bowden, both from the University of Surrey. It’s on non-rigid structure from motion, reconstructing and time pairing the shape of an object from a video sequence. It’s a bit different from standard, non-rigid structure from motion in that it works directly on the video. It doesn’t use any kind of template or any kind of prior knowledge or class specific knowledge. It only works with the video and the bounding box in the first frame which selects what the object of interest is. The implementation of the algorithm I have used will be publicly available. That is the main novelty of the project. ECCV Daily: What is so challenging? Karel: The problem is that it’s difficult to track the object or get point tracks and trajectories without knowing where the object is. To know where the object is, you need to know these trajectories in advance. The special thing is doing both of these things simultaneously. Not only that the tracking helps the reconstruction or the reconstruction helps the tracking, but as they work together, they perform better than finding the results of one and then the other. ECCV Daily: How did you solve this difficulty? Karel: It’s a bit similar to SLAM. You have a cloud of points which you are trying to reconstruct. We track these points and get a 3D point “ Blendshapes are not used enough in computer vision ”