A Depth Multiple Hierarchical StereoNet by Using Thermodynamic Color Guidance

Abstract

Thermodynamics is devoted to the study of energy or materials, and few researchers have combined the laws of thermodynamics with stereoscopic vision. Thus, we propose Depth Multiple Hierarchical StereoNet (DMH-Net) by using thermodynamic color guidance, which present end-to-end deep architecture for real-time stereo matching, realize sub-pixel precision and produce high-quality refinement disparity on dataset benchmarks. At the same time, we use color thermodynamics to guide the image output of an accurate Hill disparity image. Our DMH-Net consists of four modules: the first module uses a deep spatial pooling method to extract multiple features that work on the initial left and right images. The second module use color guided to produce high-quality cost volume. We use the ahead of two modules to calculate depth disparity precision, which can segment coarse to fine to obtain background and edge of objects information. Thirdly module integrates the information about 3D cost volume and color channel to frame a 4D hierarchical cost volume. The fourth module is designed by depth hierarchical refinement to regress the final disparity. Finally, we achieve sub-pixel precision and real-time performance on benchmarks, which can produce high-quality thermodynamic disparity regression maps.