Video Object Segmentation via Cellular Automata Refinement

We present a robust algorithm for unsupervised video object segmentation using foreground prior estimated from optical flow. Optical flow is an important cue for predicting the region of foreground object in a video. However, the estimation of flow is inherently inaccurate near the occluded object boundaries. We show that, even though the foreground prior might be unreliable due to the inaccurately estimated flow, Cellular Automata can be used to refine the foreground prior and thus is helpful to define the energy function for better segmentation accuracy. The experiments on the recently proposed DAVIS dataset show that our method performs favorably against the existing ones.