Stereo Matching Using Gabor Convolutional Neural Network

Abstract

Stereo matching is important to many robot applications. Recent work has shown that we can use convolutional neural networks (CNNs) to compute disparity map from a stereo image pair directly. However, current methods don’t take the problem of geometric deformation caused by perspective projection into account. To address the problem, we design a new model that utilizes Gabor convolutional neural networks (Gabor CNNs) in the feature extraction part of the network architecture. In Gabor CNNs, the learned filters from ordinary CNNs are modulated by the Gabor filters, which can enhance their abilities to address the problem of geometric deformations. Finally, we test our model on KITTI and Scene Flow datasets. The results show that our model outperforms the ordinary model that uses ordinary CNNs by a large margin.