#Boosting Cross-Domain Point Classification via Distilling Relational Priors from 2D Transformers

This repo is a PyTorch implementation for Boosting Cross-Domain Point Classification via Distilling Relational Priors from 2D Transformers Paper

The code has been tested with

• Python >= 3.7
• PyTorch == 1.8.0+cu111
• torch-scatter == 2.0.7
• torchsampler == 0.1.2
• torchvision == 0.9.0+cu111

Some dependent packages:

• PyTorchEMD

Please refer to issue #6 before installing.

cd PyTorchEMD
python setup.py install

Download the official PointDA-10 dataset and put the folder under [your_dataroot]/data/.
After download, the directory structure should be:

${ROOT}
|--PointDA_data
|  |--modelnet
|  |--scannet
|  |--shapenet

Download the MAE Pre-trained Vit Model and put the folder under pretrained/.

Training on both source and target

python main.py --src_dataset modelnet --trgt_dataset scannet --dataroot [your_dataroot] --batch_size 16
python main_spst.py --exp_name 'spst' --trgt_dataset scannet --dataroot [your_dataroot] --batch_size 16 --lr 5e-5

If you want to test with pre-trained model, download it from here and place it at experiments/

This repo benefits from PointCLIP_V2, MAE, GAST. Thanks for their wonderful works.