Disparity Computation With Low Intensity Quantization on Stereo Image Pairs

This is the authors' abstract. We don't add key points for this paper.

From variate bit-rate stereo matching, it is observed that the image pair with a low intensity quantization level is still capable of providing good disparity maps. In this paper, a mathematical model representing the level of disparity discontinuity is proposed to formulate the mismatching prediction based on the intensity quantization. It is used to derive the minimum quantization level for quality-assured stereo matching. Due to the high computational cost of stereo processing, the reduction of image data usage will benefit both network training and real-time inference. The formulation presented in this work is investigated extensively for various types of scenes for disparity computation. In the experiments, the feasibility of our approach is validated with Middlebury, KITTI 2015, and synthetic Scene Flow datasets. Code is available athttps://github.com/qqwweerr987/On-Stereo-Image-Intensity-Quantization-and-Disparity-Computation.