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

Researchers introduce SFMambaNet, a novel deep learning architecture that combines spectral-frequency analysis with Mamba-based state space models to improve correspondence pruning—the task of filtering accurate feature matches from noisy initial sets. The method outperforms existing Graph Neural Network approaches by integrating frequency domain perception to better distinguish valid correspondences from outliers.

SFMambaNet addresses a fundamental challenge in computer vision: distinguishing genuine feature correspondences from false positives. Traditional GNN-based methods struggle because they rely solely on geometric features derived from Euclidean coordinates, lacking the sophistication to capture subtle geometric consistencies. The researchers' innovation lies in combining two complementary approaches: incorporating spectral information into local processing and deploying frequency-based gating mechanisms in global context modeling.

The architecture's novelty stems from applying frequency domain analysis—a technique borrowed from signal processing—to a computer vision problem. Most existing correspondence pruning methods operate purely in geometric space, blind to the frequency characteristics of noise patterns. By introducing Local Spectral-Geometric Attention blocks that embed spectral positional encoding, the system gains enhanced discriminative power at local scales. The Spectral-Integrated Global Mamba component then uses frequency information as a gating signal to suppress accumulated noise in hidden states, effectively filtering out inconsistent features before they propagate through the network.

This work matters for multiple applications requiring precise feature matching: 3D reconstruction, image registration, and structure-from-motion pipelines all depend on robust correspondence pruning. The near-linear computational complexity makes the approach practical for large-scale deployments. The open-sourced implementation accelerates adoption within the research community and downstream computer vision applications.

Future developments likely involve extending spectral-frequency approaches to other vision tasks beyond correspondence pruning, and exploring how frequency-aware mechanisms could enhance other state space models in deep learning.

• →SFMambaNet integrates frequency domain perception with Mamba state space models for improved feature correspondence filtering
• →Local Spectral-Geometric Attention blocks use spectral positional encoding to enhance detection of subtle geometric consistencies
• →Frequency gating mechanisms suppress noise accumulation in hidden states, improving inlier-outlier separation
• →The method achieves near-linear computational complexity while outperforming current state-of-the-art approaches
• →Open-source implementation enables rapid adoption in computer vision and 3D reconstruction applications