Group-Based Atrous Convolution Stereo Matching Network

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