This paper presents a novel approach to designing and analyzing an algorithm for the unsupervised recurrent all-pairs field transform of optical flow. Optical flow estimation algorithms attempt the prediction of motion flow for two consecutive frames in a video. The challenge is to accurately model the motion between frames without any prior knowledge and an efficient algorithm to achieve optical flow estimation has to take into consideration the physical constraints of the scene. The proposed method combines two-frame optical flow estimation with recurrent all-pairs field transform to produce an accurate and efficient solution. First, the frames are sampled to determine the motion field in both images, then the optical flow is estimated through the recurrent all-pairs field transform of optical flow. Afterward, the parameters of the all-pairs field transform are determined to minimize the spatial error in the motion field. Finally, the optical flow is verified using a predictive error metric and improved via a deep learning optimization procedure. Results demonstrate that the proposed algorithm is capable of providing accurate estimations of optical flow in a number of video sequence datasets.

Self-supervised learning, Multi-frame, Full-Image Warping