PointCloudSuperResolution.pytorch

This repository is implementation of AR-GCN(https://arxiv.org/abs/1908.02111) from "Point Cloud Super Resolution with Adversarial Residual Graph Networks" in Pytorch. You can find official Tensorflow implementation here.

The model implementations are in src/model

Note

The code is tested under Pytorch 1.9.0 and Python 3.8 on Ubuntu 18.04 LTS.

We used pytorch3d operation & loss functions for k-nn & cd_loss calculation.

Note that to enable non-symmetric chamfer distance calculation (only forward is used), we modified pytorch3d/loss/chamfer.py like below.

    cham_dist = cham_x + cham_y
    cham_normals = cham_norm_x + cham_norm_y if return_normals else None

    forward = cham_x
    backward = cham_y
    return forward, backward, cham_normals

Usage

Training

You can download training patches in HDF5 format in here. (Note that the link is not maintained by me. Please refere PU-Net official repository for the dataset.)

cd src
# ResGCN (80 epoch pre training. Train only generator)
python train.py config/train_config_res_gcn.yaml

# AR GCN (40 epoch GAN training. Must specify saved pre-trained weight from above)
python train.py config/train_config_ar_gcn.yaml

Prediction and Evaluation

You can download evaluation data from author's official repository.

cd src
python test.py # refer 'src/config/test_config.yaml' for test settings

Performance

	CD	F-Score
Res-GCN	0.0092	0.6349
AR-GCN	0.0090	0.6470

Prediction example

• Ground truth
• Input
• Prediction (Res-GCN)
• Prediction (AR-GCN)

Issues

• (fixed) ~~Training GAN takes too much time(~7s/it) compare to the author's implementation(~1s/it)~~
• (partially fixed) Performance is relatively low compared to the author's implementation. Especially during GAN training phase, discriminator does not trained well compare to the author's implementation.

Contact

[email protected]