Intensive matching optimization method and device based on frame-level twin structure

• Main Authors: GUAN KAI, LIU BING, DONG XIAOHUAN, NIU ZEXUAN, JIN FEI, ZHANG HAO, YANG FAN, MIAO YUZHE, XU GUOQING, TONG FURONG
• Format: Patent
• Language: Chinese; English
• Published: 05.07.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; IMAGE DATA PROCESSING OR GENERATION, IN GENERAL; PHYSICS

The invention belongs to the technical field of computer vision, and particularly relates to a dense matching optimization method and device based on a frame-level twin structure, and the method comprises the steps: inputting a left image and a right image of a target data set into a PSMNetMS network left parallax branch, and generating a left parallax image; inputting a horizontal flipped graph of the original left and right graphs to a right parallax branch of the PSMNetMS network to generate a right parallax graph; the loss of the left disparity map and the loss of the right disparity map are calculated through a loss function; and only updating the network parameters of the right parallax branch of the PSMNetMS network by using back propagation, synchronously updating the left parallax branch of the PSMNetMS network through weight sharing, continuously iterating until the network converges, and ending the training. The time consumed for model training is shortened by nearly 50%, meanwhile, the precision o