SoftPoolNet: Shape Descriptor for Point Cloud Completion and Classification

The implementation of our paper accepted in ECCV (EUROPEAN CONFERENCE ON COMPUTER VISION, 16th, 2020)

Authors: Yida Wang, David Tan, Nassir Navab and Federico Tombari If you find this work useful in yourr research, please cite:

@article{DBLP:journals/corr/abs-2008-07358,
  author    = {Yida Wang and
               David Joseph Tan and
               Nassir Navab and
               Federico Tombari},
  title     = {SoftPoolNet: Shape Descriptor for Point Cloud Completion and Classification},
  journal   = {CoRR},
  volume    = {abs/2008.07358},
  year      = {2020},
  url       = {https://arxiv.org/abs/2008.07358},
  archivePrefix = {arXiv},
  eprint    = {2008.07358},
  timestamp = {Fri, 21 Aug 2020 15:05:50 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2008-07358.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Thanks for @slothfulxtx who has corrected my PyTorch implementation of SoftPool operator!

SoftPoolNet

Object Completion

Soft-Pool Operation

Train

SoftPool operators are provided in Pytorch frameworks, both using CUDA 10.2, we recommend to use the Pytorch version 1.2.0.

As we have some comparison experiments on GRNet and MSN, we suggest that you compile python libs in chamfer_pkg, emd, expansion_penalty and extensions. You can go towards each folder which includes the mentioend libs by cd, then

python setup.py install --user

Suppose that GPU 0 is supposed to get used for training

CUDA_VISIBLE_DEVICES=0 python3 train.py --batch 16 --n_regions 8 --num_points 2048 --dataset shapenet --savepath ijcv_shapenet_softpool --methods softpool

In case there are pretrained models

CUDA_VISIBLE_DEVICES=0 python3 train.py --batch 16 --n_regions 8 --num_points 2048 --dataset shapenet --savepath ijcv_shapenet_softpool --model log/ijcv_shapenet_softpool/network.pth --methods softpool

Benchmarks

Currently you can train and validate related works which are posted in Complete3D benchmark using the same infrastructure

CUDA_VISIBLE_DEVICES=0 python3 val.py --n_regions 1 --num_points 2048 --model log/ijcv_shapenet_pcn/network.pth --dataset shapenet --methods pcn # PCN
CUDA_VISIBLE_DEVICES=1 python3 val.py --n_regions 1 --num_points 2048 --model log/ijcv_shapenet_pointcnn/network.pth --dataset shapenet --methods pointcnn # PointCNN
CUDA_VISIBLE_DEVICES=0 python3 val.py --n_regions 1 --num_points 2048 --model log/ijcv_shapenet_folding/network.pth --dataset shapenet --methods folding # FoldingNet
CUDA_VISIBLE_DEVICES=0 python3 val.py --n_regions 1 --num_points 2048 --model log/ijcv_shapenet_grnet/network.pth --dataset shapenet --methods grnet # GRNet
CUDA_VISIBLE_DEVICES=0 python3 val.py --n_regions 8 --num_points 2048 --model log/ijcv_shapenet_softpool/network.pth --dataset shapenet --methods softpool # SoftPoolNet
CUDA_VISIBLE_DEVICES=0 python3 val.py --n_regions 1 --num_points 2048 --model log/ijcv_shapenet_msn/network.pth --dataset shapenet --methods msn # MSN
CUDA_VISIBLE_DEVICES=0 python3 val.py --n_regions 1 --num_points 2048 --model log/ijcv_shapenet_pointgcn/network.pth --dataset shapenet --methods pointgcn

Listed approaches (until ECCV 2020) are reported in complete3d dataset where you can reproduce our results with scripts in 'benchmark' folder.

wangyida/softpool