PoInt-Net

ICCV 2023, CV4Metaverse Workshop: Intrinsic Appearance Decomposition Using Point Cloud Representation (Official PyTorch Implementation)

News

[08, Sept. 2023] The code is released.

[16, Aug. 2023] Our work (short paper) is accepted by the CV4Metaverse workshop, ICCV 2023.

Requirements

• Python 3.7+
• Pytorch
• numpy
• scipy
• scikit-learn
• scikit-image
• MiDaS

Data Preparation

With depth:

• Run depth2normal.py to precompute the surface normal.
• Run img2pcd.py to preprocess the rgb-d image to point cloud.

Without depth:

• Go to MiDaS to precompute the depth （or any other depth estimation model）
• Run depth2normal.py to precompute the surface normal.
• Run img2pcd.py to preprocess the rgb-d image to point cloud.

Evaluation

• Run 'test.py' to decompose the albedo and shading. (We provide example data in folder: Data)

Pre-trained Model

We provide pre-trained models in the pre_trained_model folder.

Acknowledgement:

This work is funded by Bosch Center for Artificial Intelligence. If you find the work helps you, please consider citing the paper.

@inproceedings{xing2023intrinsic,
  title={Intrinsic Appearance Decomposition Using Point Cloud Representation},
  author={Xing, Xiaoyan and Groh, Konrad and Karaoglu, Sezer and Gevers, Theo},
  booktitle={ICCVW},
  year={2023}
}