Reproduce “CorrMoE: Mixture of Experts with De-stylization Learning for Cross-Scene and Cross-Domain Correspondence Pruning” on OpenResearch

About this paper

Peiwen Xia, Tangfei Liao, Wei Zhu, Danhuai Zhao, Jianjun Ke, Kaihao Zhang, Tong Lu, Tao Wang

CorrMoE, developed by researchers at Nanjing University and collaborators, introduces a framework for robust correspondence pruning, integrating de-stylization learning and a Mixture of Experts (MoE) module to enhance generalization across diverse visual domains and scene structures. This approach yields superior performance in camera pose estimation and outlier rejection, with notable improvements in cross-domain generalization, achieving an 8.38 percentage point lead in average AUC@5° on the ZEB benchmark compared to prior methods.

Key insights

• Learning domain-invariant representations through style mixing on both implicit and explicit graph features is effective in mitigating domain shift for correspondence pruning.
• A Mixture of Experts (MoE) architecture enables adaptive processing of diverse scene structures by routing features to specialized experts, improving generalization across varying geometric patterns.
• The synergistic combination of de-stylization learning and the MoE mechanism enhances robustness and adaptability to both cross-domain shifts and scene variations in correspondence pruning.

Summary by alphaXiv.