Disparity Distribution Equalization: An Effective Data Enhancement for Stereo Matching

With the continuous development of convolutional neural networks(CNN), stereo matching algorithms have made great achievements. However, existing studies mainly focus on network model structure, ignoring the dataset itself and related data augmentations. In this paper, we focus on the impact of the inhomogeneity of dataset disparity distribution in the field of binocular depth estimation. We performed quantitative analyses on multiple datasets and discovered that an imbalanced disparity distribution resulted in models failing to effectively cover low interval of disparity distribution. Based on this observation, we propose the method of disparity distribution equalization and provide two approaches for adjusting the disparity distribution: scaling transformation and random translation transformation. On this basis, we retrain some representative state-of-the-art algorithms(e.g., PSMNet, GA-Net, AANet). Experiments demonstrate that through the disparity distribution equalization, those algorithms can significantly improved on the test set without introducing additional parameters.