HomoMamba for Self-supervised Homography Estimation

Homography is a global projective transformation that can convert points in one image taken from different perspectives into corresponding points in another image. The computational complexity of the currently popular Trans-former-based method is quadratically related to the spatial resolution of the features, and the computational cost is too high. However, Mamba performs well in modeling long-distance dependencies of linear complexity, providing a solution to the above dilemma. Therefore, in this paper, we propose a homography estimation method HomoMamba that uses Mamba to optimize feature fusion. It is based on the Mamba module and uses a special cascade method to establish connections between different features at low cost. In addition, we also propose 2D-Homography-Scan (HS2D) and HomoAttention modules. The former improves the scanning method of the 2D-Selective-Scan (SS2D) module, which can scan two images at the same time to better learn the feature relationship between images; the latter is a special attention module that uses a two-stage approach to enhance the difference and correlation information between features. Experimental results demonstrate that HomoMamba outperforms current state-of-the-art Transformer-based methods for homography estimation.