Reproduce “JamMa: Ultra-lightweight Local Feature Matching with Joint Mamba” on OpenResearch

About this paper

Xiaoyong Lu, Songlin Du

JamMa is an ultra-lightweight feature matcher that leverages the Mamba State Space Model and a novel JEGO strategy to achieve competitive matching performance with significantly reduced computational overhead. This approach addresses the limitations of high-cost Transformer-based methods by offering linear computational complexity, enabling efficient processing of high-resolution images.

Key insights

• Transformer-based feature matchers achieve high accuracy but incur significant computational costs due to quadratic complexity, limiting their use in resource-constrained environments.
• State Space Models (SSMs), particularly Mamba, offer linear computational complexity and global receptive fields, making them suitable for efficient sequence modeling.
• The JEGO (Joint, Efficient, Global, Omnidirectional) strategy, with its Joint Scan and JEGO Merge components, effectively adapts Mamba for feature matching by enabling frequent mutual interaction between image features, achieving omnidirectional representation, and aggregating global information.

Summary by alphaXiv.