Submitted to: BOP Challenge 2024

Training data: None, our method does not require any task-specific training

Onboarding data: Model-based rendering, 40 rendered templates

Used 3D models: Default, CAD

Notes:

1) Testing data: RGB

2) Segmentation method: Default, CNOS

3) Our method does not introduce any additional or task-specific training

4) Pose refinement: Yes, the pose refinement network is from MegaPose with 5 hypotheses

5) Key idea: Our method leverages 3D-consistent features derived from 3D foundation models for training-free 6D pose estimation, rather than relying on 2D foundation models (e.g., DINO).