Learning Optical Flow From Still Images

This document reports additional details concerning CVPR 2021 paper -"Learning optical flow from still images". Section 1 shows a further comparison between RAFT models trained on depthstilled data with models trained on real images with proxy labels obtained by a hand-made flow algorithm, while most of the remaining material concerns visualizations of the data generated by our depthstillation pipeline (Sections 2-6), as well as qualitative results of RAFT models on standard datasets (Sintel in Section 7, KITTI 2012 and 2015 in Section 8) and videos freely available online (Section 9).