Multitasking Correlation Network for Depth Information Reconstruction

Abstract

In this paper, we propose a novel multi-tasking network for stereo matching. The proposed network is trained to approximate similarity functions in statistics and linear algebra such as correlation coefficient, distance correlation and cosine similarity. By doing this, the proposed method decreases the amount of time needed to calculate the disparity map by using CNN's ability to calculate multiple pairs of image patches at the same time. We then compare the execution time and overall accuracy between the traditional method using functions and our method. The results show the model's ability to mimic the traditional method's performance while taking considerably less time to perform the task.