MambaGaze-Stereo: Enhanced Depth Estimation for Autonomous Driving via Selective Attention and State Space Models

This is the authors' abstract. We don't add key points for this paper.

In autonomous driving, accurate and real-time depth estimation is essential for precise environmental perception and safe decision-making. However, traditional stereo matching approaches often struggle with capturing long-range dependencies, preserving fine details, and achieving robust disparity estimation in complex scenes. To address these challenges, we introduce MambaGaze-Stereo, a stereo matching framework based on State Space Models (SSM). Our framework includes a MobileNetV4-based feature extraction module, which significantly enhances multi-scale feature representation. We further incorporate a FlashAttention-based Feature Correlation Enhancement Module (FACEM), which utilizes an efficient attention mechanism to improve feature correlation and disparity estimation accuracy. Additionally, we propose the Mamba Selective Multi-Update Block (MSMUB), leveraging bidirectional selective attention to enable adaptive feature modeling across various scales, meeting the spatial information requirements of complex driving scenarios. Evaluated on the KITTI2012 dataset, MambaGaze-Stereo achieves an Out-Noc error rate of 0.84% in non-occluded regions and an Out-All error rate of 1.09% across all regions, outperforming existing state-of-the-art models and securing the top position. Our method ofers a compelling solution for high-precision stereo matching tasks in autonomous driving applications.