Hybrid Dual-Attention Functional Map Learning for Robust Shape Matching

siyu li, Minqi Li, Kaibing Zhang, Xiangjian He, Liu Runqing

Keywords: 3D shape matching，Deep functional maps，Attention mechanism，Optimal transport

Abstract: Although deep functional map methods have significantly advanced the field of 3D shape matching, many existing approaches still rely on conventional network architectures for feature enhancement and use only the Laplace–Beltrami operator (LBO) to construct eigenbases for functional map computation. This often results in performance degradation when the learned features are insufficiently distinctive. To overcome these limitations, we propose an efficient unsupervised framework for deformable shape matching. Our method incorporates a feature extraction module with a dual-layer attention mechanism, a differentiable functional map solver, and an optimal transport (OT) post-processing step to produce accurate point-to-point correspondences. The attention mechanism learns discriminative and structurally invariant descriptors, significantly improving robustness under complex geometric deformations. Additionally, we introduce a hybrid matching strategy that integrates both Laplacian and elastic modal representations, optimized via Sinkhorn iterations to yield a transport matrix. This facilitates robust and accurate correspondence recovery. Extensive experiments across diverse challenging scenarios demonstrate that our approach outperforms state-of-the-art methods in matching accuracy. Our code is publicly available.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 15411