This lesson builds a small synthetic task for rigid registration between a source point cloud and a transformed target point cloud.
- deterministic toy source shapes (cube/sphere surfaces)
- rigid pose targets with translation and Euler-angle supervision
- point-cloud registration wiring using
dlhub.pointcloud.registration.pointnetlk - compact MSE training and pose error tracking
python -m tracks.pointcloud.lesson_36_toy_pointcloud_registration.train \
--epochs 1 \
--max-train-batches 2 \
--max-eval-batches 1 \
--device cpu