Progressive Point Cloud Deconvolution Generation Network

by Le Hui, Rui Xu, Jin Xie, Jianjun Qian, and Jian Yang, details are in paper.

Usage

1. requires:

   CUDA10.1
   Pytorch 1.7.1
   Python3.7
   

2. build ops:

   cd PDGN
   cd lib/pointops && python setup.py install && cd ../../
   
   cd evaluation/pytorch_structural_losses/
   make clean
   make
   

3. Dataset:

   We follow DPM and use its processed dataset. Please download shapenet.hdf5

4. Train:

   CUDA_VISIBLE_DEVICES=0 python main.py \
      --network PDGNet_v2 \
      --model_dir PDGNet_v2 \
      --batch_size 35 \
      --max_epoch 3000 \
      --snapshot 50 \
      --dataset shapenet15k \
      --choice chair \
      --phase train \
      --data_root dataset/shapenet.hdf5
   

5. Test (may take about 2 hours):

   CUDA_VISIBLE_DEVICES=0 python main.py \
      --network PDGNet_v2 \
      --batch_size 50 \
      --pretrain_model_G 600_chair_G.pth \
      --pretrain_model_D 600_chair_D.pth \
      --model_dir PDGNet_v2 \
      --choice chair \
      --phase test
   

Results

1. Results in Chair category (taken from paper DPM):

   Model	JSD ↓	MMD -CD ↓	MMD -EMD ↓	COV -CD ↑	COV -EMD ↑	1-NNA -CD ↓	1-NNA -EMD ↓
   PC-GAN (ICML 18)	6.649	13.436	3.104	46.23	22.14	69.67	100.00
   GCN-GAN (ICLR 18)	21.708	15.354	2.213	39.84	35.09	77.86	95.80
   TreeGAN (ICCV 19)	13.282	14.936	3.613	38.02	6.77	74.92	100.00
   PointFlow (ICCV 19)	12.474	13.631	1.856	41.86	43.38	66.13	68.40
   ShapeGF (ECCV 20)	5.996	13.175	1.785	48.53	46.71	56.17	62.69
   PDGN (ECCV 20)	6.764	12.852	2.082	53.48	39.33	60.71	75.53
   DPM (CVPR 21)	7.797	12.276	1.784	48.94	47.52	60.11	69.06

2. Pretrained model in Chair categroy:

   (1) Download and put in path: ./checkpoint/PDGNet_v2/PDGNet_v2

   (2) Run the test code.

3. We will provide more pretrained models for other categories soon.

Citation

If you find the code useful, please consider citing:

@inproceedings{hui2020pdgn,
  title={Progressive Point Cloud Deconvolution Generation Network},
  author={Hui, Le and Xu, Rui and Xie, Jin and Qian, Jianjun and Yang, Jian},
  booktitle={ECCV},
  year={2020}
}

Acknowledgement

Our Cuda code is from PointWeb.

Our data processing and evaluation code is from diffusion-point-cloud.