Mamba-based multi-branch cost aggregation for stereo matching

Open Webpage

Xingyuan Lu, Yanbing Xue, Leida Li, Shiyin Li, Zan Gao

Published: 2025, Last Modified: 05 Jan 2026Appl. Soft Comput. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•We introduce the lightweight aggregation network, Mamba-based multi-branch cost aggregation (MMBA), consisting of the Mamba Aggregation Branch, Edge Aggregation Branch, and Spatial Aggregation Branch. These branches work together to efficiently aggregate information, learn edge details, and preserve spatial details, significantly enhancing the network’s aggregation capability.•We propose the AggMamba Block, a lightweight cost aggregation block that combines three-dimensional State Space Models (SSMs) with the Feature Select-MLP technique. The SSMs improve information flow efficiency and reduce computational complexity, while Feature Select-MLP filters key feature information, ensuring high computational performance and effective disparity integration.•Evaluations on the widely-used KITTI 2015 and KITTI 2012 benchmark datasets show that the proposed framework outperforms other real-time frameworks in terms of accuracy while maintaining strong performance.