EGH-Net: Energy-Guided Hypergraph for Two-View Correspondence Learning

Xiaojie Li, Xu Cheng, Jingtao Wang, Xin Jiang, Zechao Li

Keywords: Correspondence Learning, Image Feature Matching, Camera Pose Estimation

Abstract:

Learning reliable correspondences (inliers) and removing unreliable correspondences (outliers) is a fundamental task in computer vision. However, previous works based on local neighborhood graphs fail to effectively capture high-order constraints among nodes.
To address these, we propose an Energy-Guided Hypergraph Network (EGH-Net), which leverages energy functions to guide the hypergraph to accurately capture higher-order constraints, thereby achieving more effective outlier rejection.
Specifically, we first construct the hypergraph to capture group-wise relations, and then design the intra-graph energy function to compute feature differences among multiple nodes within subgraphs and to model the consistency constraints within hyperedges.
Then, we design the inter-graph energy function to capture structural similarity across subgraphs,
and implement it through the proposed Graph Kernel (GK) module using multi-scale feature decomposition.
Finally, we optimize both intra- and inter-graph energy terms via stochastic gradient descent (SGD) to dynamically update feature representations, so as to improve the local geometric consistency of node features and effectively achieve structural alignment across subgraphs.
Extensive experiments demonstrate that EGH-Net achieves superior performance compared to state-of-the-art methods across various visual tasks. The code will be released when the paper is accepted.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 10063