Reproduce “OV9D: Open-Vocabulary Category-Level 9D Object Pose and Size Estimation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Junhao Cai, Yisheng He, Weihao Yuan, Siyu Zhu, Zilong Dong, Liefeng Bo, Qifeng Chen

Researchers introduce OV9D, a framework for open-vocabulary category-level 9D object pose and size estimation for novel object categories, leveraging pre-trained visual-language foundation models. The system achieves mAPs of 26.8% for relative 5°5cm precision and 43.8% for relative 10°5cm on unseen categories, demonstrating consistent pose estimation and strong generalization to real-world data.

Key insights

  • Pre-trained visual-language foundation models can be effectively combined and adapted to provide powerful semantic and discriminative features for robust 3D object perception in open-vocabulary settings.
  • A large-scale synthetic dataset with careful symmetry annotations is crucial for training models that can learn consistent canonical object frames and generalize effectively to novel categories.
  • Evaluating open-vocabulary pose estimation requires new metrics like "Relative Rot&Trans precision" to assess the consistency of learned category-canonical frames, rather than solely relying on absolute ground-truth alignment.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.