This is implementation of PCN——Point Completion Network in pytorch. PCN is an autoencoder for point cloud completion. As for the details of the paper, please refer to arXiv.

• Ubuntu 18.04 LTS
• Python 3.7.9
• PyTorch 1.7.0
• CUDA 10.1.243

Compile for cd and emd:

cd extensions/chamfer_distance
python setup.py install
cd ../earth_movers_distance
python setup.py install

Hint: Don't compile on Windows platform.

As for other modules, please install by:

pip install -r requirements.txt

Please reference render and sample to create your own dataset. Also, we decompressed all .lmdb data from PCN data into .ply data which has smaller volume 8.1G and upload it into Google Drive. Here is the shared link: Google Drive.

In order to train the model, please use script:

python train.py --exp_name PCN_16384 --lr 0.0001 --epochs 400 --batch_size 32 --coarse_loss cd --num_workers 8

If you want to use emd to calculate the distances between coarse point clouds, please use script:

python train.py --exp_name PCN_16384 --lr 0.0001 --epochs 400 --batch_size 32 --coarse_loss emd --num_workers 8

In order to test the model, please use follow script:

python test.py --exp_name PCN_16384 --ckpt_path <path of pretrained model> --batch_size 32 --num_workers 8

Because of the computation cost for calculating emd for 16384 points, I split out the emd's evaluation. The parameter --emd is used for testing emd. The parameter --novel is for novel testing data contains unseen categories while training. The parameter --save is used for saving the prediction into .ply file and visualize the result into .png image.

The pretrained model is in checkpoint/.

I trained the model on Nvidia GPU 1080Ti with L1 Chamfer Distance for 400 epochs with initial learning rate 0.0001 and decay by 0.7 every 50 epochs. The batch size is 32. Best model is the minimum L1 cd one in validation data.

The threshold for F-Score is 0.01.

• PCN: Point Completion Network
• PCN's official Tensorflow implementation