Awesome
Deep3DMM
Official repository for the CVPR 2021 paper Learning Feature Aggregation for Deep 3D Morphable Models.
Requirements
This code is tested on Python 3.7 and Pytorch versoin 1.4 with CUDA 10.0. Requirments can be install by running
pip install -r requirements.txt
Install mesh processing libraries from MPI-IS/mesh.
Train
To start the training, follow these steps
-
Update default config file, default.cfg as needed, especially data_dir path.
-
Run the training of Deep3DMM by
python main.py -m ComaAtt
Note that the 'sliced' dataset split is used by default.
Evaluation
Run the evaluation by
python main.py -m ComaAtt --eval
Note that the checkpoint with best validation accuracy is evaluated by default.
Acknowledgement
This implementation is built upon the Pytorch implementation of COMA (Link). We also build our Deep3DMM with spiral convolution based on the implementation of Neural3DMM. Many thanks to the authors for releasing the source code.
License
This code is free for non-commerical purposes only. For commercial usage, please contact the authors for more information.
Cite
Please consider citing our work if you find it useful:
Zhixiang Chen and Tae-Kyun Kim, "Learning Feature Aggregation for Deep 3D Morphable Models", CVPR, 2021.