A Lightweight Target-Regularization Network of Stereo Matching for Inland Waterways

Abstract

Stereo matching for inland waterways is one of the key technologies for the autonomous navigation of Unmanned Surface Vehicles (USVs), which involves pixel-level matching of reference images and target images. However, inland waterways exhibit a combination of man-made and natural objects, resulting in chaotic geometric features that are difficult to match. Furthermore, USVs constrain the parameter budget of the employed algorithms. Due to these difficulties, high-precision learning-based algorithms have not yet been applied in this field. To tackle the aforementioned challenges, we introduce a lightweight, target-regularization stereo matching network, LTNet, that balances accuracy and parameter efficiency. LTNet leverages the inherent perspective information in the target image to effectively regularize the features extracted from stereo pairs, thereby improving the efficiency of feature matching. First, we design a lightweight 4D cost volume, named Geometry Target Volume (GTV), for encoding the geometric features of the target image. The GTV is engineered to minimize extraneous information and to construct a simplified, distinct initial matching volume. Subsequently, we introduce a Left-Right Consistency Refinement (LRCR) module to integrate soft constraints from the target perspective into disparity estimation, effectively mitigating inaccuracies in matching caused by ambiguous features. Furthermore, we employ knowledge distillation to boost the generalization performance of our lightweight model across the USVInland dataset. Experimental results indicate that our proposed LTNet exhibits competitive performance, with a parameter count of merely 3.7 million. The code is available at https://github.com/tustAilab/LTNet.

Keywords: Inland waterways, Stereo matching, Lightweight Network, Cost volume, Left-right consistency