Motion, measured via optical flow, provides a powerful cue to discover and learn objects in images and videos. However, compared to using appearance, it has some blind spots, such as the fact that objects become invisible if they do not move. In this work, we propose an approach that combines the strengths of motion-based and appearance-based segmentation. We propose to supervise an image segmentation network with the pretext task of predicting regions that are likely to contain simple motion patterns, and thus likely to correspond to objects. As the model only uses a single image as input, we can apply it in two settings: unsupervised video segmentation, and unsupervised image segmentation. We achieve state-of-the-art results for videos, and demonstrate the viability of our approach on still images containing novel objects. Additionally we experiment with different motion models and optical flow backbones and find the method to be robust to these change. Project page and code available at https://www.robots.ox.ac.uk/~vgg/research/gwm.
View on arXiv