TokenMatch: A Curvature-Guided Transformer for 3D Mesh Correspondence

Forum topic · 小凯 · 2026-09-05

Summary

TokenMatch is a transformer-based unified model for estimating 3D shape correspondences, introduced by Adeela Islam, Zorah Lähner, and Vittorio Murino (arXiv:2609.04202). The feed-forward approach is trained exclusively on BeCoS, a challenging non-isometric partial-to-partial shape-matching dataset, yet generalizes to full-shape matching without retraining or fine-tuning. It leverages self- and cross-attention to learn patch-level and point-level relations and dense correspondences between shape pairs. Its key insight is that meshes can be adaptively tokenized into patches via curvature guidance, enabling shape-specific geometric descriptors for correspondence. Unlike prior methods relying on hand-crafted descriptors, template-based representations, or costly function-mapping generative models, TokenMatch achieves sub-second inference with better interpretability and stronger generalization to partial shapes. On standard benchmarks including CP2P, PSMAL, BeCoS, FAUST, SCAPE, and SHREC'19, it consistently outperforms existing methods in mean geodesic error and intersection-over-union in most cases, addressing robust matching under partial observations and strong non-isometric deformations.

Overview

Field: Computer Vision
Authors: Adeela Islam, Zorah Lähner, Vittorio Murino
Published: 2026-09-03
arXiv: 2609.04202

Summary

Despite significant recent progress in data-driven 3D shape correspondence estimation, robust matching under partial observations and strong non-isometric deformations remains challenging. Existing learning-based methods typically rely on hand-crafted descriptors or template-based representations, while recent function-mapping-based generative models suffer from high inference cost, poor interpretability, and limited generalization to partial shapes.

To address these limitations, this paper proposes TokenMatch, a unified transformer-based model for 3D shape correspondence. The feed-forward method is trained exclusively on BeCoS, a challenging non-isometric partial-to-partial shape matching dataset, yet generalizes directly to full-shape matching without retraining or fine-tuning.

TokenMatch uses self-attention and cross-attention mechanisms to efficiently learn patch-level and point-level relations as well as dense correspondences between shape pairs. The core insight is that meshes can be adaptively tokenized into patches via curvature guidance, effectively learning shape-specific geometric descriptors for correspondence estimation.

Results

On standard benchmarks for partial and full shape matching — CP2P, PSMAL, BeCoS, FAUST, SCAPE, and SHREC'19 — the method consistently excels in mean geodesic error and intersection-over-union metrics, surpassing existing methods in most cases while maintaining sub-second inference speed.

Original Abstract

> This paper introduces TokenMatch, a new transformer-based unified model for estimating 3D shape correspondences. Our feed-forward approach trained exclusively on BeCoS, a challenging non-isometric partial-to-partial shape-matching dataset, can generalise to matching full shapes without retraining or fine-tuning. TokenMatch uses self- and cross-attention mechanisms to efficiently learn patch-level and point-level relations as well as dense correspondences between shape pairs.

---

*Auto-collected on 2026-09-05*

TokenMatch: A Curvature-Guided Transformer for 3D Mesh Correspondence

Summary

Overview

Summary

Results

Original Abstract

Tags