P3SNet: Parallel Pyramid Pooling Stereo Network

In autonomous driving and advanced driver assistance systems (ADAS), stereo matching is a challenging research topic. Recent work has shown that high-accuracy disparity maps can be obtained with end-to-end training with the help of deep convolutional neural networks from stereo images. However, many of these methods suffer from long run-time for real-time studies. Therefore, in this paper, we introduce P3SNet, which can generate both real-time results and competitive disparity maps to the state-of-the-art. P3SNet architecture consists of two main modules: parallel pyramid pooling and hierarchical disparity aggregation. The parallel pyramid pooling structure makes it possible to obtain local and global information intensively from its multi-scale features. The hierarchical disparity aggregation provides multi-scale disparity maps by using a coarse-to-fine training strategy with the help of the costs obtained from multi-scale features. The proposed approach was evaluated on several benchmark datasets. The results on all datasets showed that the proposed P3SNet achieved better or competitive results while having lower runtime. The code is available at https://github.com/aemlek/P3SNet.

Automated summary

This paper introduces P3SNet, which can generate real-time and competitive disparity maps by using deep convolutional neural networks for end-to-end training from stereo images. The P3SNet architecture consists of two main modules: parallel pyramid pooling and hierarchical disparity aggregation. The parallel pyramid pooling structure allows for intensive extraction of local and global information from multi-scale features, while the hierarchical disparity aggregation provides multi-scale disparity maps using a coarse-to-fine training strategy with the help of costs obtained from multi-scale features.