Learning Optical Flow with Adaptive Graph Reasoning

Estimating per-pixel motion between video frames, known as optical flow, is a long-standing problem in understanding and analysis. Most contemporary flow techniques largely focus on addressing the cross-image matching with feature similarity, few methods considering how to explicitly reason over given scene for achieving holistic understanding. In this work, taking fresh perspective, we introduce novel graph-based approach, called adaptive graph reasoning (AGFlow), emphasize value of scene/context information flow. Our key idea decouple context from procedure, exploit effectively assist estimation by learning graph. The proposed AGFlow can incorporate it within producing more robust accurate results. On both Sintel clean final passes, our achieves best accuracy EPE 1.43 2.47 pixels, outperforming state-of-the-art approaches 11.2% 13.6%, respectively. Code publicly available at https://github.com/megvii-research/AGFlow.