EIMa: Efficient and robust local feature matching with interleaved Mamba
Junxi Liu, Yang Fang, Yu Tao, Bingbing Jiang, Uswah Khairuddin, Yeon Lee, Qilie Liu · Pattern Recognition · 2026
The trade-off between accuracy and efficiency remains a fundamental challenge in feature matching. Although recent Mamba-based detector-free matchers such as JamMa achieve remarkable performance in outdoor scenes with low inference latency, their lightweight parameterization and reliance on purely sequential modeling constrain generalization capabilities, leading to significant degradation in complex indoor environments. To address this limitation, we propose EIMa, a novel feature matching framework that effectively balances accuracy and computational efficiency. First, we propose an interleaved scan strategy that spatially interweaves tokens from both views at the feature level to promote high-frequency cross-view interactions. Second, we design a sparse omnidirectional scan scheme coupled with a lightweight global aggregator, which substantially reduces the scan sequence length while preserving a global receptive field. Finally, we introduce an aggregated attention mechanism that synergizes with these scan paradigms to strengthen cross-view global modeling, thereby significantly enhancing generalization in challenging scenarios. Extensive experiments demonstrate that EIMa achieves a superior performance-efficiency trade-off, providing an efficient and robust solution for lightweight local feature matching. Project page: https://github.com/ljxg00dj0b/EIMa .