التدفق البصري

Classic optical flow often assumes brightness constancy and spatial smoothness. RAFT uses feature correlations across image pairs and a recurrent update operator to refine dense flow. Occlusion, illumination change, motion boundaries, textureless areas, and large displacement can make correspondence difficult. A flow vector is image-plane motion and does not by itself provide metric depth or object identity. In frame interpolation, estimated optical flow can guide synthesis of intermediate frames between two observed frames; this application depends on the quality of the motion estimate. An occlusion occurs when a scene point is hidden in one frame or newly revealed in another, leaving no valid pixel correspondence in the hidden frame. Flow can therefore be uncertain at occlusion boundaries.