Accurate Stereo Matching Based on 3D Labels in Driving Scenes.

We present a novel global stereo matching approach aimed at achieving accurate and smooth 3D label disparity estimation based on LocalExp. This innovative scheme can effectively assign an appropriate 3D label to each pixel from an infinite label space and generate a continuous disparity map with subpixel accuracy. To address the persistent challenge of weak texture, our method adopts a novel approach by utilizing calculated initial labels rather than random labels, which yields enhanced disparity maps. The process involves employing grid cells of the same size as the original image pair, allowing the reduced image pair to capture more detailed information in local regions. After enlarging the disparity map of reduced image pair and replacing the random initial labels with the refined ones, we achieve superior results in our method. Experimental results demonstrate the high matching accuracy of our method in driving scenes, and its generalizability is verified.