Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Self-supervised Learning of Motion Capture

This is code for the paper: Hsiao-Yu Fish Tung, Hsiao-Wei Tung, Ersin Yumer, Katerina Fragkiadaki, Self-supervised Learning of Motion Capture, NIPS2017 (Spotlight)

Check the project page for more results.


  • Environment setup and Dataset
  • Data preprocessing
  • Pretrained model and small tfrecords
  • Training
  • Citation
  • License

1. Environment setup and Dataset

  • python We use python2.7.13 from Anaconda and Tensorflow 1.1

  • SMPL model: We need rest body template from SMPL model.

You can download it from here.

  • SURREAL Dataset: If you plan to pretrain or test on surreal dataset.

Please download surreal from here

  • H36M Dataset: If you plan to test on real video with some groundtruth (to evaluate).

Please download H3.6M Dataset from here

2. Data preprocessing

  • Parse Surreal Dataset into binary files

In order to speed up the read write for tfrecords, we parse surreal dataset into binary files. Open file


and change the data path and output path.

  • Build up tfrecords

change the data path to the path you built in the previous step in


and run it. You can specify how many examples you want to have in each tfrecords by changing value for num_samples. If "is_test" is False, we use sequences generated from actor 1, 5, 6, 7, 8 as training samples. If "is_test" is True, we use only sequence "" from actor 9 as validation. You can change this split by modifying the "get_file_list" function in

3. Pretrained model and small tfrecords

You can downdload a pretrained model using supervision from here surreal_quo0.tfrecords is a small training data and surreal2_100_test_quo1.tfrecords

Note: To make this code pack, I calculate 2d flow directly from 3d groundtruth during testing. But you should replace this with your own predicted flow and keypoints.

4. Train model

open up, there is one commend for pretraining using supervision, and one commend for finetuning with testing data. Commend out the line that you need


If you use this code, please cite:

@incollection{NIPS2017_7108, title = {Self-supervised Learning of Motion Capture}, author = {Tung, Hsiao-Yu and Tung, Hsiao-Wei and Yumer, Ersin and Fragkiadaki, Katerina}, booktitle = {Advances in Neural Information Processing Systems 30}, editor = {I. Guyon and U. V. Luxburg and S. Bengio and H. Wallach and R. Fergus and S. Vishwanathan and R. Garnett}, pages = {5236--5246}, year = {2017}, publisher = {Curran Associates, Inc.}, url = {} }


Code for Paper: Self-supervised Learning of Motion Capture






No releases published


No packages published