Group-Based Atrous Convolution Stereo Matching Network

Stereo matching is the key technology in stereo vision. Given a pair of rectified images, stereo matching determines correspondences between the pair images and estimate depth by obtaining disparity between corresponding pixels. The current work has shown that depth estimation from a stereo pair of images can be formulated as a supervised learning task with an end-to-end frame based on convolutional neural networks (CNNs). However, 3D CNN puts a great burden on memory storage and computation, which further leads to the significantly increased computation time. To alleviate this issue, atrous convolution was proposed to reduce the number of convolutional operations via a relatively sparse receptive field. However, this sparse receptive field makes it difficult to find reliable corresponding points in fuzzy areas, e.g., occluded areas and untextured areas, owing to the loss of rich contextual information. To address this problem, we propose the Group-based Atrous Convolution Spatial Pyramid Pooling (GASPP) to robustly segment objects at multiple scales with affordable computing resources. The main feature of the GASPP module is to set convolutional layers with continuous dilation rate in each group, so that it can reduce the impact of holes introduced by atrous convolution on network performance. Moreover, we introduce a tailored cascade cost volume in a pyramid form to reduce memory, so as to meet real-time performance. The group-based atrous convolution stereo matching network is evaluated on the street scene benchmark KITTI 2015 and Scene Flow and achieves state-of-the-art performance.