Neural-Intrinsic-Embedding

This repository is a PyTorch implementation of Neural-Intrinsic-Embedding.

Requirements

To install requirements:

pip install -r requirements.txt

Installing PyTorch may require an ad hoc procedure, depending on your computer settings.

Data & Pretrained models

You can find the data and the pre-trained models in:

data
models

Evaluation

To evaluate the model FAUST\SCAPE, run:

python code/faust/test_faust_sample.py
or
python code/scape/test_scape_sample.py

And in matlab the script:

code/eval/FAUST_5k.m
or 
code/eval/SCAPE_5k.m

Training

First, you should compute the geodesic distance matrix for each shape in the dataset. You can use the code in this repository or this matlab code and put them in:

data/{DATASET_NAME}/geod
For example:
data/SCAPE_5k/geod/mesh000.npy  

if mesh000 has N points, then the distance matrix mesh000.npy has the shpae of [N,N].

To train the basis model, you may run:

python code/train_basis_sample.py --config config/train_scape_5k.yaml

Thenn, to train the descriptor model, you may run:

python code/train_desc_sample.py --config config/train_scape_5k.yaml 

License

If you use this code, please cite our paper.

@inproceedings{jiang2023neural,
  title={Neural Intrinsic Embedding for Non-rigid Point Cloud Matching},
  author={Jiang, Puhua and Sun, Mingze and Huang, Ruqi},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={21835--21845},
  year={2023}
}

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. For any commercial uses or derivatives, please contact us.