🚀Excited to release PartField—a feedforward model that learns part-based feature fields for 3D shapes! It enables lightning-fast⚡️, robust, open-world hierarchical 3D part seg and unlocks cross-shape applications like co-seg and correspondence! 🔗 1/n

3D part seg remains an open challenge in computer vision. Recent open-world methods distill 2D priors via per-shape optimization but suffer from lengthy runtimes and noisy results. PartField is a feedforward model that delivers lightning-fast speed and robust performance. 2/n

Instead of relying on text prompts, PartField converts any 3D shape into a feature field that captures the general concept of 3D parts. It then decomposes the shape into hierarchical, multi-granularity parts by applying a clustering algorithm to the field. 3/n

We train PartField at scale with contrastive learning on both 2D data (distilled masks) and 3D supervision (when available). It showcases strong open-world capabilities across diverse categories, 3D modalities (meshes, Gaussians), and shape styles (artistic, Gen AI, CAD). 4/n

Instead of recognizing a single part or producing a fixed-granularity clustering, PartField implicitly learns a hierarchy of multi-scale parts and outputs a part tree. Users can interactively choose branches to decompose further based on their desired granularity. 5/n

Another interesting point is that, while we do not explicitly incorporate any cross-shape supervision, consistency surprisingly emerges in the learned feature space across different shapes. The figure visualizes similarities across the field relative to a selected location. 6/n

This emergent consistency enables various cross-shape applications, such as shape co-segmentation and correspondence, and demonstrates that PartField learns open-world, general-purpose, hierarchical, and consistent 3D feature fields. 7/n

Check out our project page, released code, and checkpoints! 🔗 Many thanks to our amazing collaborators: @mikacuy, @DonglaiXiang , @haosu_twitr , @FidlerSanja , @nmwsharp and @JunGao33210520! 8/n

The first community-owned crypto media outlet has partnered with DAIC, a leading Web3 infrastructure & non-custodial staking provider, to pioneer a new community validator model. Stake with us.

Have u tried with ai genwrated mesh ? 😂