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

STRAPS - Synthetic Training for Real Accurate Pose and Shape

Code repository for the paper:
Synthetic Training for Accurate 3D Human Pose and Shape Estimation in the Wild
Akash Sengupta, Ignas Budvytis, Roberto Cipolla
BMVC 2020
[paper][supplementary][SSP-3D dataset]

Update: Check out Hierarchical Probabilistic 3D Humans for a better synthetic training framework for human pose and shape estimation from images.




  • Linux or macOS
  • Python ≥ 3.6


We recommend using a virtual environment to install relevant dependencies:

python3 -m venv STRAPS
source STRAPS/bin/activate

After creating a virtual environment, first install torch and torchvision: pip install torch==1.4.0 torchvision==0.5.0

Then install (my fork of) detectron2 and its dependencies (cython and pycocotools):

pip install cython
pip install 'git+'
pip install 'git+'

The remaining dependencies can be installed by simply running: pip install -r requirements.txt. This will be sufficient for inference. If you wish run model training, you will require the PyTorch port of Neural Mesh Renderer: pip install neural_renderer_pytorch==1.1.3.

Additional files

You will need to download the SMPL model. The neutral model is required for training and running the demo code. If you want to evaluate the model on datasets with gendered SMPL labels (such as 3DPW and SSP-3D), the male and female models are available here. You will need to convert the SMPL model files to be compatible with python3 by removing any chumpy objects. To do so, please follow the instructions here.

Download the required additional files here. Place both the SMPL model and the additional files in the additional directory such that they have the following structure and filenames:

├── additional                                # Folder with additional files
│   ├── smpl
│       ├── SMPL_NEUTRAL.pkl                  # Gender-neutral SMPL model 
│   ├── cocoplus_regressor.npy                # Cocoplus joints regressor
│   ├── J_regressor_h36m.npy                  # Human3.6M joints regressor
│   ├── J_regressor_extra.npy                 # Extra joints regressor
│   ├── neutral_smpl_mean_params_6dpose.npz   # Mean gender-neutral SMPL parameters
│   ├── smpl_faces.npy                        # SMPL mesh faces
│   ├── cube_parts.npy
│   └── vertex_texture.npy                    
└── ...

Model checkpoints

Download pre-trained model checkpoints for our SMPL regressor, as well as for PointRend and DensePose (via detectron2) from here. Place these files in the checkpoints directory.

Training data

If you wish to train using the STRAPS framework, we provide pre-processed npz files with training and validation pose parameters here. Place these files in the data directory. Note that the model described in the STRAPS paper was trained on SMPL poses from Human3.6M (obtained using MoSh). We cannot distribute these due to license limitations. However, we have prepared npz files with SMPL poses from the AMASS dataset to replace H3.6M, which should provide similar performance.

Inference is used to run inference on a given folder of input images. For example, to run inference on the demo folder, do:

python --input ./demo --checkpoint checkpoints/straps_model_checkpoint.tar --silh_from pointrend

You can choose between using --silh_from pointrend and --silh_from densepose to obtain human silhouettes. PointRend provides more accurate silhouettes for easy body poses but DensePose is more robust to challenging body poses. Best results are achieved when the image is roughly cropped and centred around the subject person.

To test STRAPS against a wide range of body shapes, we recommend checking out the SSP-3D dataset.

If you are running inference on a remote machine, you might run into problems getting pyrender to run off-screen/headless rendering. If you have EGL installed, uncommenting the appropriate line in might work. If not, simply disable pyrender rendering during inference.


The training options within the file can be edited as desired. Since STRAPS does not require image data, the only training data you are required to download are the pre-processed npz files (above). Then, simply do python


If you find this code useful in your research, please cite the following publication:

               author = {Sengupta, Akash and Budvytis, Ignas and Cipolla, Roberto},
               title = {Synthetic Training for Accurate 3D Human Pose and Shape Estimation in the Wild},
               booktitle = {British Machine Vision Conference (BMVC)},
               month = {September},
               year = {2020}                         


Code was adapted from/influenced by the following repos - thanks to the authors!


Code repository for the paper: Synthetic Training for Accurate 3D Human Pose and Shape Estimation in the Wild (BMVC 2020)








No releases published


No packages published