Hierarchical Feature Fusion and Multi-scale Cost Aggregation for Stereo Matching

Abstract

To further improve the accuracy of disparity estimation in ill-posed regions and weak texture regions, in this paper we propose HFMANet: which is a stereo matching method based on hierarchical feature fusion and multi-scale cost aggregation. Specifically, we first propose a hierarchical feature fusion module, which innovatively fuses low-level features and high-level features to obtain rich semantic information while retaining the edge information of the image. Secondly, we propose a multi-scale cost aggregation module to extract rich global context information. At the same time, the layer-by-layer fusion optimization helps increase the receptive field to capture more structural information, reduce the dependence on local information, and help the disparity estimation of ill-posed regions and weak-textured regions. Comprehensive experiments are conducted on the SceneFlow and KITTI datasets, and achieve competitive results, which proves the effectiveness of the proposed method.