Localized Gaussians as Self-Attention Weights for Point Clouds Correspondence

How to run

Training

To train the model run:

python train.py --run_name MODEL_NAME [--gaussian_heads H] [--sigma SIGMA_1 ... SIGMA_H] [--learn_sigma] [--noise NOISE_LEVEL] [--noise_p P]

where:

• MODEL_NAME is the name of the model you want to train. The model will be saved in the models folder with the .pt extension.
• The --gaussian_heads flag specifies the number of Gaussian heads to use in the model. The --sigma flag specifies the standard deviation of the Gaussian heads.
• If --learn_sigma is used, the standard deviation will be learned during training. If both --sigma and --learn_sigma are used, the standard deviation will be learned and initialized with the values provided.
• The --noise flag specifies the noise level to add to the input point clouds. The noise is added with probability --noise_p.

Testing

To test the trained model on FAUST1K, simply run:

python test.py --run_name MODEL_NAME

where MODEL_NAME is the name of the model you want to test. The model should be in the models folder and is to be used without the .pt extension.

To test using a different dataset, you can use the --dataset flag:

python test.py --run_name MODEL_NAME --dataset DATASET_NAME

where DATASET_NAME is the name of the dataset you want to test on, without extension. The dataset should be in the dataset folder.

Acknowledgements

Training and testing code are based on the implementation of raganato: [https://github.com/raganato/SGP23_AttPos4ShapeMatching]

The code uses a modified version of the X-Tranformers library by Lucidrains: [https://github.com/lucidrains/x-transformers]

Datasets, Utils and Testing from the transmatching are based on the code by GiovanniTRA: [https://github.com/GiovanniTRA/transmatching]