Home

Awesome

Point-Cloud-Color-Constancy

CVPR 2022:Point Cloud Color Constancy

update: Release training code (24/08/2022)

[pdf] [data]

poster

Data

Introduction

We provide the extended illumination labels of NYU-v2, DIODE, and ETH3D as well as the point cloud, the raw format image(for ETH3D), and the linearization sRGB image (for NYU-2 and DIODE).

Each dataset consists of following parts:

For the full depth information and images on the three open-source datasets, please refer to their website.

NYU-v2:https://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html

DIODE:https://diode-dataset.org/

ETH3D: https://www.eth3d.net/datasets#high-res-multi-view

Code

We provide an example of the data processing, which include the image aligned, point cloud building, and point cloud visualization (based on open3D). We also provide the train & test code of the PCCC network.

Environment & Packages

For creating a new environment on sever

conda env -f create environment.yaml

For adding the necessary packages

pytorch==1.2.0
torchvision==0.4.0
open3D #(for point cloud visualization)
openCV

Data processing

If you use our depthAWB data for training, you can skip this phase. If you use your own data, you can refer to PcdGeneration.py to create your own point cloud data and visualize it.

Network

For training

python train_main.py --datasets NAME OF DATASET --foldn FOLD NUMBER --sizes INPUT SIZE OF POINT --batch_size BATCH SIZE --nepoch EPOCH --gpu_ids GPU ID

The ./pointnet/DataLoader.py can be changed if your use your own data.

Citation

If our work helps you, please cite us:

@InProceedings{Xing_2022_PCCC,
    author    = {Xing, Xiaoyan and Qian, Yanlin and Feng, Sibo and Dong, Yuhan and Matas, Ji\v{r}{\'\i}},
    title     = {Point Cloud Color Constancy},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {19750-19759}
}

Acknowledgement

This code of PCCC network is developed on the bias of PointNet.Pytorch and PointNet2.Pytorch. We thank the authors for their contribution.