Efficient Stereo Matching With Adaptive Disparity Range Scaling and Feature Calibration

Abstract

Cascaded cost volume-based stereo matching has gained significant attention for its ability to produce high-resolution depth maps with efficient hardware utilization. However, previous methods often perform uniform offset sampling within a predicted disparity range, which can easily overlook the ground-truth disparity, leading to error propagation and accumulation. Moreover, these methods lack specific designs to tackle spatial feature misalignment, which limits their effectiveness in real-time applications. In this article, we introduce a lightweight cascaded stereo matching framework that features a disparity range scaling (DRS) module and a spatial feature calibration (SFC) module. The DRS module adaptively adjusts the disparity sampling range and constructs a new cost volume that incorporates finer semantic features, ensuring that computational and memory resources are focused on more critical regions. The SFC module performs selective sampling and calibration to extract the most informative and representative features at each position, effectively mitigating spatial feature misalignment. Through rigorous experiments across SceneFlow, KITTI 2012, and KITTI 2015, our model’s effectiveness and efficiency have been confirmed, demonstrating that it outperforms other speed-based algorithms while remaining competitive with state-of-the-art approaches.