ECCV 2020 Daily - Monday

3 Adam Harley 1 DAILY M o n d a y an ECCV paper from two years ago by Carl Vondrick called Tracking Emerges by Colorizing Videos . A key difference being that this method learns an unsupervised 3D tracker, while Vondrick’s learned on an unsupervised 2D tracker. Thinking about next steps, Adam would like to be able to develop a tracker that is just as robust but for deformable objects. A limitation of the work so far has been that the tracker assumes that the objects are rigid. “It would be really nice to get rid of the rigidity assumption,” he says hopefully. To find out more, visit Adam’s poster presentation [#1185] today (Monday) at 14:00 and tomorrow (Tuesday) at 00:00 (UTC +1). What computer vision techniques are involved? “There are some critical techniques,” Adam explains. “One is triangulation – we use that to be able to localize a stationary point from multiple viewpoints and know where it lands in those viewpoints. We also rely on metric learning, which allows us to learn a feature space that is correspondable and discriminative. We use the classic technique of RANSAC and this allows us to turn a set of noisy correspondences into a consensus estimate of the total object motion. We also utilize the classic technique of cross-correlation. This allows us to deliver heatmaps for where we think a point corresponds in the next time step.” Adam reveals this work was inspired by