TokenMatch: 3D Mesh Correspondence Transformer with Curvature-Guided Tokenisation

What it claims

TokenMatch is a feed-forward transformer for 3D shape correspondence that is trained exclusively on BeCoS (a non-isometric partial-to-partial matching dataset) yet generalises to full-shape matching with no retraining or fine-tuning. It claims to run at sub-second inference and to outperform existing methods on most partial and full matching benchmarks in mean geodesic error and intersection-over-union.

How it works

The stated core insight is that meshes are adaptively tokenised into patches using surface curvature as the guide, letting the model learn shape-specific geometric descriptors rather than relying on hand-crafted descriptors or template-based representations. Self- and cross-attention then learn patch-level and point-level relations plus dense correspondences between a shape pair. The paper positions this against two prior lines: descriptor/template approaches, and recent generative models over functional maps, which it says suffer from high inference cost, limited interpretability, and poor generalisation to partial shapes.

Evidence

Evaluated on CP2P, PSMAL, BeCoS, FAUST, SCAPE, and SHREC'19, spanning both partial and full matching. Metrics reported are mean geodesic error and IoU, with the claim of "consistently high performance, in most cases outperforming existing methods." Note this is abstract-only text: no per-benchmark numbers, no named baseline scores, and no ablation isolating the curvature-guided tokenisation are given on the page, so the head-to-head margins cannot be checked here. The one concrete quantitative claim is sub-second inference speed.

Limits

• The page is an abstract only (v1, submitted 3 Sep 2026); no tables, code link, weights, or licence are present. Treat all comparative claims as unverified until the PDF is read.
• Trained on a single dataset (BeCoS); the cross-dataset generalisation is the central claim but rests entirely on the unshown benchmark results.
• "In most cases outperforming" implies it does not win everywhere — which benchmarks it loses on is not stated.
• No model size, parameter count, training compute, or hardware for the sub-second figure is given.

If you match non-isometric or partial shapes and currently pay the inference cost of functional-map generative methods, this is worth pulling the PDF for the actual error/IoU tables and any released code.