Stereo confidence measurement based on linearity of disparity profile for monocular depth estimation

Abstract

In stereo matching, one of the fundamental problems in computer vision fields, errors within occlusions, object boundaries, reflective surfaces, textureless regions, and repeated pattern regions remain critical problems to be solved. Stereo confidence estimation plays a role in mitigating the aforementioned problems by detecting unreliable pixels in disparity obtained through stereo-matching and being used to refine the results of those pixels. However, in the case of previous learning-based methods, there is a limitation in that a separate training process is needed. In this thesis, we propose a stereo confidence measurement method without extra network and training by defining an ideal disparity profile according to the disparity plane sweep based on the linearity of the disparity profile. Experimental results in self-supervised monocular depth estimation problems, where stereo confidence estimation is utilized, demonstrate the validity of the proposed method.