Reproduce “Learning Canonical Shape Space for Category-Level 6D Object Pose and Size Estimation” on OpenResearch

About this paper

Dengsheng Chen, Jun Li, Zheng Wang, Kai Xu

A method introduces a correspondence-free, end-to-end deep learning approach for category-level 6D object pose and size estimation from a single RGB-D image by learning a Canonical Shape Space (CASS) via a VAE and a view-factorized RGB-D embedding. This approach achieves 23.5% on the 5°5cm metric on the NOCS-REAL275 dataset, outperforming previous state-of-the-art methods like NOCS (16.1%).

Key insights

• Implicitly learning a Canonical Shape Space (CASS) using a VAE can effectively model intra-class shape variations and serve as a unified representation for category-level objects.
• A view-factorized RGB-D embedding, which separates shape from pose information, can be achieved by training RGB-D and 3D encoders to share a common latent space via mixed training batches and shared gradients.
• Estimating 6D pose by contrasting pose-independent shape features with pose-dependent visual features provides a robust, correspondence-free approach that generalizes well to unseen object instances.

Summary by alphaXiv.