Reproduce “Generalizable Single-view Object Pose Estimation by Two-side Generating and Matching” 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
Yujing Sun, Caiyi Sun, Yuan Liu, Yuexin Ma, Siu Ming Yiu
A generalizable method for single-view object pose estimation is introduced, leveraging generative diffusion models and a novel two-side matching technique to accurately estimate 3D poses for unseen object categories, even with large viewpoint differences, without requiring 3D models or multi-view training data.
Key insights
- Large viewpoint changes, challenging for direct generation by diffusion models, can be effectively bridged by generating and matching multiple intermediate viewpoints from both reference and query images.
- A two-side matching approach, comparing generated views from both the reference and query, provides a robust mechanism to evaluate and optimize the relative 3D pose.
- A score function derived from the noise prediction of a diffusion model can serve as an effective proxy for image similarity, guiding the pose optimization process without explicit 3D models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.