Dynamic programming with adaptive and self-adjusting penalty for real-time accurate stereo matching

Dense disparity map extraction is one of the most active research areas in computer vision. It tries to recover three-dimensional information from a stereo image pair. A large variety of algorithms has been developed to solve stereo matching problems. This paper proposes a new stereo matching algorithm, capable of generating the disparity map in real-time and with high accuracy. A novel stereo matching approach is based on per-pixel difference adjustment for the absolute differences, gradient matching and rank transform. The selected cost metrics are aggregated using guided filter. The disparity calculation is performed using dynamic programming with self-adjusting and adaptive penalties to improve disparity map accuracy. Our approach exploits mean-shift image segmentation and refinement technique to reach higher accuracy. In addition, a parallel high-performance graphics hardware based on Compute Unified Device Architecture is used to implement this method. Our algorithm runs at 36 frames per second on 640 x 480 video with 64 disparity levels. Over 707 million disparity evaluations per second (MDE/s) are achieved in our current implementation. In terms of accuracy and runtime, our algorithm ranks the third place on Middlebury stereo benchmark in quarter resolution up to the submitting.

Automated summary

This study introduces a new stereo matching algorithm that utilizes pixel-level difference adjustment, gradient matching, and rank transform, combined with guided filter aggregation of cost metrics and dynamic programming for disparity calculation to improve accuracy and runtime of the disparity map. Mean-shift image segmentation and refinement techniques are also used to enhance accuracy. The algorithm achieves high disparity evaluation speed and ranks third in accuracy and runtime on the Middlebury stereo benchmark.