第六讲
Iterative Closet Point (ICP)
It takes two point clouds as input and computes an R & T matrix that transforms one point cloud to align with the other as closely as possible.
- Make data centered
- Compute chamfer distance between two point clouds (for every , search the closest to it) and get correspondences from to .
- Find using constrained orthogonal Procrustes
- Update
Category-Level 6D Object Pose Estimation
Normalized Object Coordinate Space
- rotation normalization: align object orientation
- translation normalization: zero-center the objects
- scale normalization: uniformly normalize the scales
Category-Level 6D Pose
transformation from NOCS to camera space
Pose Estimation
RGB Image → NOCS point cloud
Depth Image → Depth point cloud
可以得到NOCS point cloud和Depth point cloud的dense correspondence。