ABSTRACT
This study presents a dynamic programming approach to solving variational problems in computer vision, focusing on stereo matching, optical flow, and image segmentation. Our method improves recursion and backtracking efficiency, enhancing speed and accuracy. Evaluations on benchmark datasets such as Middlebury Stereo, KITTI Optical Flow, and BSDS500 show higher accuracy, faster processing, and comparable memory use versus leading methods like SGM, Farneback, and GrabCut. This approach suits real-time applications, addressing modern computer vision challenges effectively.
