Construct CNNs which are capable of solving the optical flow estimation problem as a supervised learning task.
Train CNN end-to-end to learn predicting the optical flow field from a pair of images
Propose and compare two architectures: a genetic architecture and another one including a layer that correlates feature vectors at different image locations.
数据库： KITTI Sintel
While optical flow estimation needs precise per-pixel localization, it also requires finding correspondences between two input images.
Not only learning image feature representations, but also learn to match them at different locations in the two images.