This AI can match incomplete 3D shapes and still work on complete ones without extra training.

Computers struggle to match points between two three-dimensional shapes when parts are missing or surfaces stretch and bend. To solve this, researchers built TokenMatch, a system based on a transformer model. Instead of looking at an entire object at once, it breaks the shape's surface mesh into smaller patches. It decides where to make these patches by looking at curvature, which is how sharply the surface bends. Next, it uses attention tools to compare patches and points between two objects to learn their connections. The team trained TokenMatch only on broken, bent shapes. Without extra training, it learned to accurately match both complete objects and incomplete pieces. It beats existing methods in most tests, achieves high performance, and runs fast, finishing its calculations in under a second.

Why it matters

For people working with 3D models, like animators or engineers, this makes it faster and more reliable to match shapes even when they're incomplete or stretched out. It means less time fixing matches and more time creating or analyzing.

Questions answered