Home

Awesome

DenseReg:

Fully Convolutional Dense Shape Regression In-the-Wild

Rıza Alp Güler, George Trigeorgis, Epameinondas Antonakos, Patrick Snape, Stefanos Zafeiriou, Iasonas Kokkinos


This is an implementation of the method described in the arXiv paper. For a video demonstration and supplementary materials see the project page.

Currently, only the Caffe(deeplab) based test-code that allows regressing template face coordinates on a given image is available. Training code will be provided soon.


Caffe Setup

You have two options:

1- Use deeplabv2 submodule

First install deeplabv2. It is added as a submodule to this repository, you can follow its own installation instructions. Make sure that you have set WITH_PYTHON_LAYER=1 in "Makefile.config" of the Caffe.

2- Using your own caffe

Alternatively, you can use your favorite installed caffe, then all you need to add is the "interpolation layer", which can be found in the provided deeplab version. If you do this, you have to change the caffe path in (i)DenseReg.ipynb and (ii) CombineRegressions.py caffe layer before "import caffe" line.

Running DenseReg

Namely, putting a uniform grid in the temple space onto the face, semantic face part segmentation and landmark localization results are demonstrated (as portrayed in the image below).


Running DenseReg for Human Bodies

Demonstrated result is depicted for a test sample.


Bibtex entry for citations:

  @article{Guler2016DenseReg,
  title={DenseReg: Fully Convolutional Dense Shape Regression In-the-Wild},
  author={R\{i}za Alp G\"uler, George Trigeorgis, Epameinondas Antonakos, Patrick Snape, Stefanos Zafeiriou, Iasonas Kokkinos},
  journal={arXiv:1612.01202},
  year={2016}
  }