Diffusion based Cost Aggregation Method for Stereo Matching Algorithm

Stereo vision technology is currently a prominent area of research in computer vision.Essential applications such as 3D reconstruction, robotics, and autonomous driving widely use stereo vision for depth estimation.Errors in the disparity map from the stereo matching (SM) algorithm can also occur from low texture and repetitive patterns in stereo images.The raw cost volume, derived from the matching cost computation step in the first stage of stereo matching, still contains errors.This paper proposed a diffusion-based method, Multiscale directional iterative diffusion (MDID), to aggregate the initial cost volume as the second stage of the SM algorithm.The aim is to reduce the error generated by smoothing the cost volume during the matching cost computation stage and preserve the edges of the object.Then the disparity map is computed using the Winner Take All (WTA) approach.The final disparity map was refined using a multistep refinement method: Left Right Consistency check, hole-filing and Weighted Median filter.The accuracy of the proposed SM algorithm has been evaluated using the Middlebury v3 stereo benchmarking database systems, qualitatively and quantitatively.Overall results for the proposed complete SM algorithm using the MDID performed competitively compared to other published methods such as GINet, GANet+ADL, Z2ZNCC and it has proven it effectiveness to cater the low texture region to enhance the disparity map quantitative and qualitatively with average error of 5.09% and 8.11% for Middlebury 2014 dataset.