Reproduce “Category-Level 6D Object Pose Estimation via Cascaded Relation and Recurrent Reconstruction Networks” on OpenResearch

About this paper

Jiaze Wang, Kai Chen, Qi Dou

A framework combining cascaded relation networks and recurrent reconstruction networks improves category-level 6D object pose estimation for unseen instances by enhancing feature representation from multi-modal inputs and iteratively refining 3D model reconstruction in canonical space. The method achieves improvements of 4.9% for 3D@75% and 17.7% for 5°2cm on the CAMERA25 dataset, and 2.7% for 3D@75% and 8.5% for 5°2cm on the REAL275 dataset over previous state-of-the-art.

Key insights

• Explicitly modeling relations across multi-modal inputs (RGB, point cloud, category prior), particularly using Transformer-based structures, significantly enhances feature representation for category-level 6D pose estimation.
• Iterative refinement of 3D NOCS model reconstruction and correspondence estimation addresses the limitations of single-step dense regression, leading to more geometrically precise models and improved pose accuracy.
• The synergistic combination of a cascaded relation network for enhanced feature learning and a recurrent network for iterative refinement yields a robust approach for generalizing 6D pose estimation to unseen object instances within known categories.

Summary by alphaXiv.