Neural Body Fitting code repository

Setup:

• Requires Python 2.7
• git clone --recursive http://github.com/mohomran/neural_body_fitting
• create and activate a fresh virtualenv
• pip install tensorflow-gpu==1.6.0 (or tensorflow==1.6.0)
• inside the root folder run pip install -r requirements.txt
• navigate to external/up and run python setup.py develop (which will install the UP toolbox)
• download SMPL (at http://smpl.is.tue.mpg.de/downloads) and unzip to external/
• download the segmentation model and extract into models/
• download the fitting model and extract into experiments/states

Demo:

The following command will perform inference on 60 images from the UP dataset:

python run.py infer_segment_fit experiments/config/demo_up/ \
              --inp_fp demo/up/input/\
              --out_fp demo/up/output\
              --visualise render

The results can be viewed by opening the file demo/up/output/index.html in a browser. These were selected to demonstrate both success and failure cases. Most of the processing time (~80%) is taken up by the mesh renderer. Alternatively, you can use --visualise pose which is quicker and just plots the projected SMPL joints.

How to run for SwapNet

1. Make sure input images are 512x512. If they're not, scale them up. The model in this repository CANNOT do 128x128.

2. If the input directory has subfolders, flatten the directory. This can be done by replacing all "/" with "+" temporarily.

3. Specify output size in experiments/config/demo_up/options.py, input size and intermediate size.

4. Run python run.py infer_segment_fit experiments/config/demo_up/ --inp_fp path/to/input --out_fp path/to/output

5. If subfolders, replace all "+" back to "/"

Citation

If you find any parts of this code useful, please cite the following paper:

@inproceedings {omran2018nbf,
  title = {Neural Body Fitting: Unifying Deep Learning and Model-Based Human Pose and Shape Estimation},
  journal = {International Conference on 3D Vision (3DV)},
  year = {2018},
  author = {Omran, Mohamed and Lassner, Christoph and Pons-Moll, Gerard and Gehler, Peter V. and Schiele, Bernt}
  address = {Verona, Italy},
}

Acknowledgements

The repository is modelled after (and partially adopts code from) Christoph Lassner's Generating People project. The example data provided is from his Unite the People dataset.