pvnieo / DPFM

Pytorch code for "DPFM: Deep Partial Functional Maps" - 3DV 2021 (Oral)

DPFM

Code for "DPFM: Deep Partial Functional Maps", published at 3DV 2021 (recognized with a Best Paper Award!).

Installation

This implementation runs on python >= 3.7, use pip to install dependencies:

pip3 install -r requirements.txt

Download data & preprocessing

The data should be downloaded and placed in the data folder. Each data folder should have two subfolders: shapes which contains the 3D meshes, and maps which contains the point-to-point ground truth maps that are used for training.

├── dpfm
│   ├── data
│   │   ├── my_dataset
│   │   |   ├── shapes
│   │   │   │   ├── shape1.off
│   │   |   │   ├── shape2.off
│   │   │   │   ├── ...
│   │   │   ├── maps
│   │   │   │   ├── gt_p2p1.map
│   │   │   │   ├── gt_p2p2.map
│   │   │   │   ├── ...
│   ├── diffusion_net
│   │   ├── ...
│   ├── eval_shrec_partial.py
│   ├── model.py
│   ├── ...

The data will be automatically processed when the training script is executed.

The datasets used in our paper are provided the dataset repository.

Usage

To train DPFM model on the shrec 16 partial dataset, use the training script:

python3 train_shrec_partial.py --config shrec16_cuts
# OR
python3 train_shrec_partial.py --config shrec16_holes

To evaluate a trained model, use:

python3 eval_shrec_partial.py --config shrec16_cuts --model_path path/to/saved/model --predictions_name path/to/save/perdiction/file

We provide two pre-trained models on the shrec 16 partial dataset which are available in data/saved_models.

DPFM on full non-rigid dataset

We also evaluated the performance of DPFM on a full non-rigid dataset. Specifically, we experimented with FAUST-remeshed (FR) and SCAPE-remeshed (SR) used in many previous works, such as DiffusionNet.

The results we found are provided in the table below. The syntax X on Y means that the model was trained on X and tested on Y.

Setting	FR on FR	SR on SR	FR on SR	SR on FR
GeomFMaps + DiffusionNet	2.7	3.0	3.3	3.0
DPFM	2.1	2.3	2.7	2.5

We provide two pre-trained models on the FAUST-remeshed and SCAPE-remeshed datasets which are available in data/saved_models.

Citation

@inproceedings{attaiki2021dpfm,
  doi = {10.1109/3dv53792.2021.00040},
  url = {https://doi.org/10.1109/3dv53792.2021.00040},
  year = {2021},
  month = dec,
  publisher = {{IEEE}},
  author = {Souhaib Attaiki and Gautam Pai and Maks Ovsjanikov},
  title = {{DPFM}: Deep Partial Functional Maps},
  booktitle = {2021 International Conference on 3D Vision (3DV)}
}