Abstract
Recently, end-to-end convolutional neural networks have achieved remarkable success in disparity estimation tasks. However, these neural networks usually have difficulty in finding correct correspondences in ill-posed regions, such as texture-less areas, edge details, and small objects. This paper proposes an atrous granular multi-scale network based on depth edge subnetwork(Dedge-AGMNet) to overcome the difficulty above. This work has the following contributions. On one hand, the devised depth edge subnetwork provides the geometric knowledge and depth edge constraints. To incorporate the depth edge cues efficiently, the depth edge-spatial pyramid pooling(Dedge-SPP) module fuses the depth edge features to the disparity estimation branch. And the loss functions are proposed respectively for supervised and unsupervised tasks, which can improve the adaptability of the depth edge auxiliary network. On the other, the designed granular convolution is very suitable for constructing the atrous granular multi-scale (AGM) module. AGM module could capture multi-scale context information that requires fewer parameters and consumes fewer computational resources. In summary,the depth edge cues and multi-scale context information are both beneficial to explore potential corresponding points in ill-posed regions. Integrating the ranks of different stereo datasets, our network outperforms other stereo matching networks and shows very strong robustness for different environments. The Dedge-AGMNet advances state-of-the-art performances on the Sceneflow, KITTI 2012 and KITTI 2015 benchmark datasets.
Abstract (translated)
URL
https://arxiv.org/abs/1908.09346