Point2SSM++

Implementation of Point2SSM++: Self-Supervised Learning of Anatomical Shape Models from Point Clouds. If using this code, please cite the paper.

Run training by calling 'train.py' with a specificed config file, for example:

python train.py -c cfgs/point2ssm++.yaml

This will write the model, logged info, and a copy of the config file to a folder in experiments/, such as experiments/spleen_all/point2ssm++_cd_l2_dgcnn/.

To run inference, call consist_test.py with the config file and dataset, for example:

python consist_test.py -c experiments/spleen_all/point2ssm++_cd_l2_dgcnn/point2ssm++.yaml -d spleen

This will write the predicted correspondence points to the experiment directory, for example experiments/spleen_all/point2ssm++_cd_l2_dgcnn/spleen/test/output/.

See cfgs/point2ssm++_4d.yaml for an example with 4D/spatiotemporal data, and cfgs/point2ssm++_classifier.yaml for multi-anatomy data.

Acknowledgements

This code utilizes the following Pytorch 3rd-party libraries and models:

• Chamfer Distance
• PointNet2
• DGCNN
• PointAttN (SFA Block)
• PSTNet2