This is the PyTorch implementation of our ICCV 2021 paper Intrinsic-Extrinsic Preserved GANs for Unsupervised 3D Pose Transfer.
Haoyu Chen, Hao Tang, Henglin Shi, Wei Peng, Nicu Sebe, Guoying Zhao.

If you use our code or paper, please consider citing:

@inproceedings{chen2021IEPGAN,
  title={Intrinsic-Extrinsic Preserved GANs for Unsupervised 3D Pose Transfer},
  author={Chen, Haoyu and Tang, Hao and Henglin, Shi and Peng, Wei and Sebe, Nicu and Zhao, Guoying},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2021}
}

The current version will lead to a degenerated result, the training strategy should be the same as in the paper (not this version). Revised version will be uploaded soon. Thanks to Xiaokun's help!

Requirements:

• python3.6
• numpy
• pytorch==1.1.0 and above
• trimesh

Please download FAUST data from FAUST link, and DFAUST dataset from DFAUST link for training the model.

You need to convert the datasets into .obj format for better run the source code, converted version can be downloaded here: FAUST link

The script of generating the GIH of the data is provided now, see GIH_calculating/generate_geodesic_distance.py.

The generated GIH of the FAUST dataset will be around 13GB.

You can also directly download it from the link here: FAUST GIH link.

The usage of our code is easy, just run the code below.

bash run.sh

Your can change the hyperparameters according to your GPU settings and needs in the run.sh file:

python  train.py --batch 4 ./dataset_3d/FAUST/  --edge_loss_setting 'rec_rec' --experi_path 'full_adap_GIH_Lap_rec' --server 'local' --rec_epoch 1200 --geo_epoch 1600 --iter 2000 --geoloss 0.005 --n_crop 1 --ref_crop 2 --sampling_number 120 --limb_n 4 --limb_sampling 600  --sampling_pattern 'adaptive'

Part of our code is based on

3D transfer: NPT，

GAN framework: SWAP-Autoencoder (unofficial implementation) and

GIH implementation: LIMP. Many thanks!

MIT-2.0 License