Self-supervised Learning of PSMNet via Generative Adversarial Networks

Stereo matching is an essential method to estimate the disparity between the left and the right images to obtain the depth information. With the advancement of deep learning technology, learning-based stereo matching methods continue to make breakthroughs in accuracy. However, current learning-based stereo matching models rely on disparity labels during the training. The high cost of acquiring data with real disparity information in real-life scenarios impedes the widespread application of these models. Therefore, we explore self-supervised methods for stereo matching. Through this paradigm, the model can be trained solely on binocular data and output high-precision disparity. This paper addresses the challenge of self-supervised stereo matching within the Generative Adversarial Networks framework. Experiments show that our model outperforms most self-supervised models, and the D1 metric reaches 6.10%.