Learning Features by Watching Objects Move
lastmod 2026-11-10created 2018-12-052 edits
Learning Features by Watching Objects Move is a self-supervised learning paper that learns visual features without labels by using the separation of moving objects in video as the supervisory signal. It trains a network to predict, from a single frame, motion-based segmentation masks obtained from optical flow, and shows that the features learned this way transfer to downstream tasks such as object detection.
Documentation
- Learning Features by Watching Objects Move
- https://people.eecs.berkeley.edu/~pathak/unsupervised_video/
-
Learning_Features_by_Watching_Objects_Move.pdf