Pointcept/Utonia

[ICML'26] Official repository of Utonia: Toward One Encoder for All Point Clouds

What it solves

Utonia provides a cross-domain pre-trained encoder for 3D point clouds, aiming to create a single, universal encoder that can be applied to various downstream 3D tasks across different domains (such as indoor, outdoor, and object-level data) without needing separate specialized models.

How it works

It is based on the Point Transformer V3 (PTv3) architecture and is modified from the Sonata and Concerto frameworks. The model acts as an encoder-only system that takes 3D point cloud data (coordinates, colors, and normals) and produces hierarchical encoded features. These features can then be mapped back to the original point cloud scale for use in tasks like semantic segmentation or visualization.

Who it’s for

Researchers and developers working on 3D computer vision, specifically those needing high-quality pre-trained representations for point cloud analysis, semantic segmentation, and 3D scene understanding.

Highlights

  • Cross-Domain Capability: Designed to work across diverse 3D datasets (indoor, outdoor, and objects).
  • Inference Demos: Includes ready-to-use scripts for PCA visualization, similarity heatmaps, and semantic segmentation.
  • Flexible Integration: Offers both a standalone mode for quick testing and a package mode for integration into existing codebases.
  • Video Support: Can be used with VGGT to process point clouds lifted from videos.

Related

  • Project
    Pointcept/Concerto

    A joint 2D-3D self-supervised pre-trained Point Transformer V3 model for extracting high-quality spatial representations from 3D point clouds.

  • Project
    Pointcept/Pointcept

    A powerful and flexible codebase for point cloud perception research, integrating various 3D backbones and self-supervised pre-training frameworks.

  • Project
    NVlabs/PointWorld

    PointWorld is a large pre-trained 3D world model that predicts full-scene 3D point flows from partial RGB-D captures and robot actions to improve in-the-wild robotic manipulation.

  • Project
    yuxumin/PoinTr

    PoinTr is a transformer-based model for point cloud completion that reconstructs full 3D objects from partial point clouds using a geometry-aware encoder-decoder architecture.

  • Project
    isl-org/Open3D-ML

    An extension of Open3D that provides tools, pretrained models, and pipelines for 3D machine learning tasks like semantic segmentation and object detection.