Skip to content

ralpguler/DenseReg

master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 

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

  • You have to download the caffemodel by running the script: get_densereg_model.sh.
  • Then, you can use the ipyton notebook DenseReg.ipynb, which very basically demonstrates how example results for the Lena's face are obtained.

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

  • You have to download the caffemodel by running the script: get_densereg_model.sh.
  • Then, you can use the ipyton notebook DenseRegHumanBody.ipynb, which demonstrates dense-correspondences for human bodies on sample images. Note that this network is not trained to be invariant to changes in scale.

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}
  }

About

Code repository for DenseReg.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages