Mamba-Stereo: Mamba Regularization for Stereo matching

Abstract

Stereo matching networks are essential for applications such as autonomous driving, robotic navigation, and augmented reality (AR). Traditional cost volumes, built using image features, often fail to capture global geometric information, leading to detail loss, edge blurring, and errors in textureless regions. To address these issues, we introduce a novel cost volume enhancement module based on the Mamba mechanism, termed Mamba Regularization. Mamba Regularization flattens the high-dimensional cost volume into one-dimensional features, utilizing the Spatial Mamba Block to capture long-range dependencies within whole-volume features at every scale. Additionally, we design the Spatial Residual Convolution (SRC) module to compensate for spatial information loss during flattening. Experimental results on the SceneFlow and KITTI benchmarks demonstrate that Mamba Regularization can serve as a plug-and-play module, significantly enhancing the performance of various stereo matching networks. While it introduces a slight increase in inference time, it achieves a favorable trade-off between accuracy and computational efficiency. Moreover, the proposed module exhibits strong cross-dataset generalization and maintains high inference efficiency, further validating its effectiveness in real-world applications. Code is available at https://github.com/S1aoXuan/Mamba-Regularization.