Neural Kinematic Networks for Unsupervised Motion Retargetting

This is the code for the CVPR 2018 paper Neural Kinematic Networks for Unsupervised Motion Retargetting by Ruben Villegas, Jimei Yang, Duygu Ceylan and Honglak Lee.

Please follow the instructions below to run the code:

Requirements

Our method works with works with

• Linux
• NVIDIA Titan X GPU
• Tensorflow version 1.3.0

Installing Dependencies (Anaconda installation is recommended)

• pip install --ignore-installed --upgrade https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.3.0-cp27-none-linux_x86_64.whl

Download and install blender version 2.79

Download and install from:

• https://www.blender.org/download/

Downloading Data

Train data:
Firstly, create an account in the Mixamo website. Next, refer to Apendix D in our paper, and download the corresponding fbx animation files for each character folder in ./datasets/train/. Once the fbx files have been downloaded, run the following blender script to convert them into BVH files:

blender -b -P ./datasets/fbx2bvh.py

Finally, preprocess the bvh files into npy files by running the following command:

python ./datasets/preprocess.py

Test data (already preprocessed):

./datasets/download_test.sh

Training

NKN Autoencoder:

CUDA_VISIBLE_DEVICES=GPU_ID python ./src/train_online_retargeting_mixamo.py --gpu=GPU_ID --min_steps=60 --max_steps=60 --gru_units=512 --num_layer=2 --learning_rate=0.0001 --keep_prob=0.9 --alpha=100 --gamma=10.0 --omega=0.01 --euler_ord=yzx --optim=adam

NKN with Cycle Consistency:

CUDA_VISIBLE_DEVICES=GPU_ID python ./src/train_online_retargeting_cycle_mixamo.py --gpu=GPU_ID --min_steps=60 --max_steps=60 --gru_units=512 --num_layer=2 --learning_rate=0.0001 --keep_prob=0.9 --alpha=100 --gamma=10.0 --omega=0.01 --euler_ord=yzx --optim=adam

NKN with Adversarial Cycle Consistency:

CUDA_VISIBLE_DEVICES=GPU_ID python ./src/train_online_retargeting_cycle_adv_mixamo.py --gpu=GPU_ID --min_steps=60 --max_steps=60 --gru_units=512 --num_layer=2 --learning_rate=0.0001 --keep_prob=0.9 --beta=0.001 --alpha=100 --gamma=10.0 --omega=0.01 --euler_ord=yzx --optim=adam --norm_type=batch_norm --d_arch=2 --margin=0.3 --d_rand

Inference from above training (BVH files will be saved in ./results/blender_files)

NKN Autoencoder:

CUDA_VISIBLE_DEVICES=GPU_ID python src/test_online_retargeting_mixamo.py --gpu=GPU_ID --prefix=Online_Retargeting_Mixamo_gru_units=512_optim=adam_learning_rate=0.0001_num_layer=2_alpha=100.0_euler_ord=yzx_omega=0.01_keep_prob=0.9_gamma=10.0

NKN with Cycle Consistency:

CUDA_VISIBLE_DEVICES=GPU_ID python src/test_online_retargeting_mixamo.py --gpu=GPU_ID --prefix=Online_Retargeting_Mixamo_Cycle_gru_units=512_optim=adam_learning_rate=0.0001_num_layer=2_alpha=100.0_euler_ord=yzx_omega=0.01_keep_prob=0.9_gamma=10.0

NKN with Adversarial Cycle Consistency:

CUDA_VISIBLE_DEVICES=GPU_ID python src/test_online_retargeting_mixamo.py --gpu=GPU_ID --prefix=Online_Retargeting_Mixamo_Cycle_Adv_beta=0.001_gru_units=512_optim=adam_d_arch=2_learning_rate=0.0001_omega=0.01_norm_type=batch_norm_d_rand=True_num_layer=2_alpha=100.0_euler_ord=yzx_margin=0.3_keep_prob=0.9_gamma=10.0

Inference with your trained models:
Simply change the --prefix input

Evaluation

Evaluate outputs

python ./results/evaluate_mixamo.py

Evaluate on your trained models
Please change the paths in ./results/evaluate_mixamo.py

Results location
Results will be located in ./results/quantitative/result_tables_mixamo_online.txt

Generating videos

Download blender files with character skins

./results/download.sh

Generate videos

blender -b -P ./results/make_videos.py

Citation

If you find this useful, please cite our work as follows:

@InProceedings{Villegas_2018_CVPR,
  author = {Villegas, Ruben and Yang, Jimei and Ceylan, Duygu and Lee, Honglak},
  title = {Neural Kinematic Networks for Unsupervised Motion Retargetting},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2018}
}

Please contact "[email protected]" if you have any questions.