InfoCD

The official repository of the paper "InfoCD: A Contrastive Chamfer Distance Loss for Point Cloud Completion" published at NeurIPS 2023

Illustration of comparison among CD, MI, and InfoCD with different numbers of samples.

UPDATE

Our code and model weights will be released soon!!!

UPDATE

We update SeedFormer + InfoCD in Oct 13th

SeedFormer + InfoCD

Installation

The code has been tested on one configuration:

• python == 3.6.8
• PyTorch == 1.8.1
• CUDA == 10.2
• numpy
• open3d

pip install -r requirements.txt

Compile the C++ extension modules:

sh install.sh

Datasets

The details of used datasets can be found in DATASET.md

Pretrained Models are attached

Usage

Training on PCN dataset

First, you should specify your dataset directories in train_pcn.py:

__C.DATASETS.SHAPENET.PARTIAL_POINTS_PATH        = '<*PATH-TO-YOUR-DATASET*>/PCN/%s/partial/%s/%s/%02d.pcd'
__C.DATASETS.SHAPENET.COMPLETE_POINTS_PATH       = '<*PATH-TO-YOUR-DATASET*>/PCN/%s/complete/%s/%s.pcd'

To train SeedFormer + HyperCD on PCN dataset, simply run:

python3 train_pcn.py

Testing on PCN dataset

To test a pretrained model, run:

python3 train_pcn.py --test

Or you can give the model directory name to test one particular model:

python3 train_pcn.py --test --pretrained train_pcn_Log_2022_XX_XX_XX_XX_XX

Save generated complete point clouds as well as gt and partial clouds in testing:

python3 train_pcn.py --test --output 1

Using ShapeNet-55/34

To use ShapeNet55 dataset, change the data directoriy in train_shapenet55.py:

__C.DATASETS.SHAPENET55.COMPLETE_POINTS_PATH     = '<*PATH-TO-YOUR-DATASET*>/ShapeNet55/shapenet_pc/%s'

Then, run:

python3 train_shapenet55.py

In order to switch to ShapeNet34, you can change the data file in train_shapenet55.py:

__C.DATASETS.SHAPENET55.CATEGORY_FILE_PATH       = './datasets/ShapeNet55-34/ShapeNet-34/'

The testing process is very similar to that on PCN:

python3 train_shapenet55.py --test

Acknowledgement

Code is borrowed from SeedFormer, InfoCD loss can be found in loss_utils.py, This loss can be easily implement to other networks such as PointAttN and CP-Net.

Publication

Please cite our papers if you use our idea or code:

@inproceedings{
lin2023infocd,
title={InfoCD: A Contrastive Chamfer Distance Loss for Point Cloud Completion},
author={Fangzhou Lin and Yue Yun and Ziming Zhang and Songlin Hou and Kazunori D Yamada and Vijaya Kolachalama and Venkatesh Saligrama},
booktitle={Advances in Neural Information Processing Systems},
editor={},
year={2023},
url={}
}