<!doctype html>

UniPose9D: Universal Category Agnostic Object Pose Estimation

UniPose9D: Universal Category Agnostic Object Pose Estimation

Yang You, Yi Du, Cole Harrison, Leonidas Guibas
arXiv Paper Code Dataset 🤗 Demo
UniPose9D pose estimates across cluttered tabletops, occlusions, robotic manipulation, and everyday photos

Given an object mask and either color with depth or color with predicted depth, UniPose9D estimates rotation, translation, and a metric 3D bounding box. It does not need category labels, CAD models, shape priors, or reference images. Yellow cuboids show the predicted metric boxes, while the colored axes show orientation.

Abstract

UniPose9D is a universal, category agnostic model for 9D object pose estimation. From one masked color and depth observation, or one color image with predicted metric depth, it estimates rotation, translation, and metric bounding box size. At inference time, it does not need a category name, template, shape prior, CAD model, or reference image.

Rather than regress NOCS coordinates independently at every pixel, UniPose9D samples pairs of points from the observed object cloud and combines geometric features with DINOv2 visual features. It predicts NOCS coordinates for each pair, creating many more correspondences for robust pose estimation with RANSAC and Kabsch Umeyama. Metric scale comes from the relationship between distances in the predicted NOCS space and the observed point cloud. An iterative Kabsch Umeyama procedure then refines the inlier threshold. Flow matching represents multiple valid poses for symmetric objects without requiring an explicit list of symmetries.

A single model is trained on a mixture of public pose datasets. It performs competitively on standard benchmarks and also transfers to unseen object categories and everyday scenes.

Method

UniPose9D pipeline: feature extraction, point pair NOCS prediction, and iterative Kabsch Umeyama pose recovery
Pipeline. (A) Visual and geometric features are extracted for each point in the masked observation. (B) Point pairs are sampled and a flow matching MLP head predicts NOCS coordinates and metric box scale for each pair. (C) An iterative Kabsch Umeyama solver refines the inlier set and recovers the complete 9D pose.

Interactive Demo

Try UniPose9D in your browser. Upload an image, mark the object with point prompts, then estimate its 9D pose. If the embedded app does not load, open it directly on 🤗 Hugging Face.

Results in the Wild

Pose estimates in robotic, artistic, and natural scenes
In the wild. UniPose9D transfers from robotic manipulation to artwork and natural imagery. From a single color image with predicted depth, it recovers a metric 3D box and object orientation.

BibTeX

@misc{you2026unipose9duniversalcategoryagnosticobject,
  title         = {UniPose9D: Universal Category-Agnostic Object Pose Estimation},
  author        = {Yang You and Yi Du and Cole Harrison and Leonidas Guibas},
  year          = {2026},
  eprint        = {2607.09985},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2607.09985},
}