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

Forum topic · 小凯 · 2026-09-05

Summary

TokenMatch is a transformer-based unified model for 3D shape correspondence estimation, introduced by Adeela Islam, Zorah Lähner, and Vittorio Murino (arXiv:2609.04202). The method addresses persistent challenges in robust matching under partial observations and strong non-isometric deformations, where existing approaches rely on handcrafted descriptors, template-based representations, or costly function-map generative models with limited interpretability. TokenMatch is a feed-forward architecture 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. Its key insight is that meshes can be adaptively tokenized into patches using curvature guidance, enabling the model to learn shape-specific geometric descriptors for correspondence estimation. Self-attention and cross-attention mechanisms efficiently capture patch-level and point-level relations as well as dense correspondences between shape pairs. On standard benchmarks including CP2P, PSMAL, BeCoS, FAUST, SCAPE, and SHREC'19, TokenMatch consistently outperforms prior methods in mean geodesic error and intersection-over-union metrics in most cases, while maintaining sub-second inference speeds.

Overview

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

Abstract (translated)

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 handcrafted descriptors or template-based representations, while recent function-map-based generative models suffer from high inference costs, limited interpretability, and poor generalization to partial shapes. To address these limitations, this paper proposes TokenMatch, a transformer-based unified model for 3D shape correspondence estimation.

The feed-forward approach is trained exclusively on BeCoS, a challenging non-isometric partial-to-partial shape-matching dataset, and generalizes directly to full-shape matching without retraining or fine-tuning. TokenMatch leverages 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 achieves strong results on mean geodesic error and intersection-over-union metrics, outperforming existing approaches in most cases while keeping inference at sub-second speeds.

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*