Neural-Template

Neural Template: Topology-aware Reconstruction and Disentangled Generation of 3D Meshes (CVPR 2022)

Environments

You can create and activate a conda environment for this project using the following commands:

conda env create -f environment.yml
conda activate NeuralTemplate

Dataset

For the dataset, we use the dataset provided by IM-NET for training and evaluation. We provide another zip file (link) which should be unzipped in data before any training or inference.

Training

There are altogether three commands for three training phases (Continuous training, Discrete training, Image encoder training).

For the continuous phase, you can train the model by using the following command:

python train/implicit_trainer.py --resume_path ./configs/config_continuous.py

For the discrete phase, you should first specify the network_resume_path variable in the configs/config_discrete.py as the model path in the previous phase. Then, you can run the following command:

python train/implicit_trainer.py --resume_path ./configs/config_discrete.py

Lastly, for the Image encoder training, you should specify the auto_encoder_config_path and auto_encoder_resume_path in config_image.py similarly, and run the following command to start training:

python train/image_trainer.py --resume_path ./configs/config_image.py

Testing

We provide the pre-trained models used in the paper for reproducing the results. You can unzip the file (link) in the pretrain folder.

For the evaluation of the autoencoder, you can use this command:

python utils/mesh_visualization.py

For the evaluation of the single view reconstruction task, you can use this command:

python utils/image_mesh_visualization.py

If you need test your own trained model, you change the testing_folder variable in these two python files.

Note:

you need to set PYTHONPATH=<project root directory> before running any above commands.

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{hui2022template,
    title = {Neural Template: Topology-aware Reconstruction and Disentangled Generation of 3D Meshes},
    author = {Ka-Hei Hui* and Ruihui Li* and Jingyu Hu and Chi-Wing Fu(* joint first authors)},
    booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2022}
}