Multi-path Feature Mining Network for Stereo Matching

Many existing networks for disparity estimation have attained satisfactory performance, but it still is a challenge to gain high-quality disparities for the inherently ill-posed regions and improve detail in disparity map. To tackle this problem, we propose a Multi-path Feature Mining Network(MFMNet), which contains two sub-networks: initial disparity estimation network and disparity refinement network. Specially, to generate high-quality disparity with rich detail, we specially design a initial disparity estimation network with multi-scale feature fusion and richer feature extraction module to capture context information and gain more reliable features. Furthermore, we propose a disparity refinement network with atrous spatial pyramid pooling to enhance disparity refinement performance. In this work, we adopt one-dimension single direction(1SD) correlation to calculate matching cost rather than the 1D correlation used widely. During training, we adopt a novel loss function strategy to perform online hard point mining, which enhance the overall network performance. Experimentation shows that our multi-path feature mining network provides state-of-the-art performance on FlyingThings3D and KITTI datasets.