This repository contains the code to our paper published at 3DV 2020.

This project uses data provided by Daniel Holden et al.'s paper A Deep Learning Framework for Character Motion Synthesis and Editing, which can be downloaded here. We preprocess the raw data and create training/validation splits with the scripts provided in data-preprocessing. For this to work, you will need some helper functions provided in Daniel Holden's code base in the motion module. The code base can be downloaded here.

This project is a bit older and hence the libraries it uses are too. It was tested with Python 3.5, CUDA 8 and TensorFlow 0.12. Tensorflow was installed via

(On Windows)
pip install --ignore-installed --upgrade https://storage.googleapis.com/tensorflow/windows/gpu/tensorflow_gpu-0.12.1-cp35-cp35m-win_amd64.whl

(On Linux)
pip install --ignore-installed --upgrade https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-0.12.1-cp35-cp35m-linux_x86_64.whl

A list of all dependencies is provided in requirements.txt. Lastly, make sure the folder tbase is on the python path. If you are using PyCharm, the simply mark the folders infilling and tbase as "sources root".

All necessary code for training and evaluation is provided under infilling. You need to adjust paths in two places:

• get_data_path() in flags_parser.py: return the path where the preprocessed data is located, currently set to ../data_preprocessed.
• get_checkpoints_path() in flags_parser.py: return the path where pre-trained models are stored, currently set to ../pretrained-models.

Under pretrained-models we provide the models described in the paper. The mapping is as follows:

To evaluate a pre-trained model, use the evaluate.py script. For example, to re-create the last row of Table 1 in the main paper use:

python evaluate.py --split validation --runs VGG/31f
python evaluate.py --split validation --runs VGG/31fp --perturbator column --perturbation_size [60]
python evaluate.py --split validation --runs VGG/31fp --perturbator column --perturbation_size [120]

The syntax of the --runs parameter is as follows. VGG/31 means that we want to evaluate the model under pretrained-models/VGG/run_031. f signals that foot contacts should be removed from the input data (we do not use them in our method) and p applies the perturbations. If the p is missing, no perturbations are applied, no matter how the remaining parameters are set.

As another example, to reproduce row 2 in Table 1 you can use:

python evaluate.py --split validation --runs HoldenCAE/3
python evaluate.py --split validation --runs HoldenCAE/3p --perturbator column --perturbation_size [60]
python evaluate.py --split validation --runs HoldenCAE/3p --perturbator column --perturbation_size [120]

To run the linear interpolation baseline, use --lerp.

The scripts starting with showcase_*.py were used to create the results shown in the video. Please refer to those scripts directly for more details. Very simple visualizations are available via visualize.py.

The main.py script lets you re-train models from scratch. Please refer to the file config.txt which is available for each pretrained model to see which configuration parameters were used. Models are implemented in models.py and some layer implementations can be found in ops.py.

You may use the code for non-commercial purposes as detailed in the license file. Please also respect the usage terms of the individual motion databases used in this work.

If you use this code, please cite

@inproceedings{MotionInfilling,
  author={Manuel {Kaufmann} and Emre {Aksan} and Jie {Song} and Fabrizio {Pece} and Remo {Ziegler} and Otmar {Hilliges}},
  booktitle={2020 International Conference on 3D Vision (3DV)}, 
  title={Convolutional Autoencoders for Human Motion Infilling}, 
  year={2020},
  pages={918-927},
  doi={10.1109/3DV50981.2020.00102}}

For questions or problems please create an issue or contact manuel.kaufmann@inf.ethz.ch.