Jon Barron: We tried everything under the sun and identified a simple model for learning *unsupervised* optical flow: you give it videos and it learns to produce flow, no labels needed. Results are on par with supervised techniques!
7 replies, 635 likes
Kosta Derpanis: Rico Jonschkowski, Austin Stone, Jonathan T. Barron, Ariel Gordon, Kurt Konolige, Anelia Angelova, What Matters in Unsupervised Optical Flow, #ECCV2020
0 replies, 40 likes
roadrunner01: What Matters in Unsupervised Optical Flow
github: https://github.com/google-research/google-research/tree/master/uflow https://t.co/nHh00CJN7j
0 replies, 32 likes
Daisuke Okanohara: They identified important components for unsupervised optical flow; cost volume normalization, smoothness at flow resolution, range map-based occlusion estimation, self-supervision, rectangular resizing, census loss, level dropout. UFlow achieves new SOTA. https://arxiv.org/abs/2006.04902
0 replies, 6 likes
Jia-Bin Huang: Great insight from careful ablation study!
It is, however, increasingly difficult to carry out such works in the universities (eg getting one row of results may take up to one week).
1 replies, 4 likes
HotComputerScience: Most popular computer science paper of the day:
"What Matters in Unsupervised Optical Flow"
0 replies, 3 likes
arXiv CS-CV: What Matters in Unsupervised Optical Flow http://arxiv.org/abs/2006.04902
0 replies, 1 likes
Brundage Bot: What Matters in Unsupervised Optical Flow. Rico Jonschkowski, Austin Stone, Jonathan T. Barron, Ariel Gordon, Kurt Konolige, and Anelia Angelova http://arxiv.org/abs/2006.04902
1 replies, 0 likes
Found on Aug 18 2020 at https://arxiv.org/pdf/2006.04902.pdf