Generic Foundation-model-driven 6D Unseen Detection

Training data: None

Onboarding:
Following FoundPose to generate templates (~800 rendered images) and descriptors using the provided CAD models for coarse pose estimation.

2D detection: Using the default model-based CNOS (FastSAM) for unseen 2D detection.

Authors: Temporary Anonymity

For coarse 6D pose detection, we extend FoundPose to support correct and unified perspective cropping for pinhole/fisheye query images.

For fine 6D pose detection, we directly employ GoTrack to refine the coarse results.