DSN: A Fast Stereo Disparity Estimation Network based on Deformable Convolution

Abstract

In recent years, the stereo matching method based on the convolutional neural network has been greatly developed and achieved accurate disparity estimation results. However, the high precision stereo disparity estimation method is often slow in reasoning, so it cannot meet the requirements of real-time scene. Moreover, this type of method has a large number of parameters and is therefore not friendly to resource-constrained embedded devices. In this paper, a novel 2D cost aggregation module based on deformable convolution is constructed based on AnyNet, and a lightweight stereo matching network named DSN is constructed accordingly. We measured performance on SceneFlow and KTTTI2015 dataset. Experiments show that our method achieves the approximate inference time and parameter quantity of AnyNet, and the accuracy of disparity estimation is much better than AnyNet, achieving a better tradeoff between performance and time.