CONVERGED CLASSIFICATION NETWORK FOR MATCHING COST COMPUTATION

Abstract

Stereoscopic vision lets us identify the world around us in 3D by incorporating data from depth signals into a clear visual model of the world. The stereo matching algorithm capable of producing the disparity or depth map in computer. This map is crucial for many applications such as 3D reconstruction, robotics and autonomous driving.The disparity map also prone to errors such as noises in the region which contains object occlusions, reflective regions, and repetitive patterns.So we propose this stereo matching algorithm to produce a disparity map and to reduce the errors by incorporating a deep learning approach. This paper focused on matching cost computation step as an initial step to produce the disparity or depth map. The proposed convolutional neural network designed with the output neurons in the classification part scaled-downin converging style. The raw cost generated aggregated by the normalized box filter. Then the disparity map computed using Winner Take All approach. The final disparity map refined using Weighted Median Filter. Overall quantitative results for the proposed work performed competitively compared to other established stereo matching algorithm based on the Middlebury standard benchmark online system.