Title: SFMambaNet: A Spectral-Frequency Enhanced Selective State Space Model for Correspondence Pruning

Abstract:

The primary objective of correspondence pruning is to filter out outliers and retain inliers from an initial pool of matches. Traditional approaches utilizing Graph Neural Networks (GNNs) typically depend on geometric features derived from coarse Euclidean coordinates. These methods often fail to detect the nuanced geometric consistencies characteristic of inliers. Although Mamba-based architectures offer the advantages of global receptive fields and long-sequence modeling, they are prone to accumulating significant inconsistent features within their hidden state spaces, which complicates the differentiation between inliers and outliers.

To address these challenges, this study introduces frequency domain perception to the task for the first time, presenting SFMambaNet. This is a novel two-view correspondence pruning network that leverages a Spectral-Frequency enhanced Mamba architecture. The proposed method consists of two key components. First, we introduce the Local Spectral-Geometric Attention (LSGA) module. By integrating spectral positional encoding into local graph interactions and employing multi-scale Mamba processing, LSGA improves the discriminability of local features and better captures subtle geometric consistencies.

Building on this foundation, we develop the Spectral-Integrated Global Mamba (SIGM) block. This component embeds a frequency gating mechanism into the state space. It utilizes the frequency data supplied by LSGA to explicitly curtail the accumulation of high-frequency noise within hidden states, thereby reducing the propagation of inconsistent features. This design not only enhances the separability of inliers from outliers but also enables robust global context modeling with nearly linear computational complexity. Comprehensive experiments confirm that SFMambaNet surpasses current state-of-the-art methods across various challenging tasks. The source code is accessible at https://github.com/Kirito14IT/SFMambaNet.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC