/3dmm_cnn

Regressing Robust and Discriminative 3D Morphable Models with a very Deep Neural Network

Primary LanguagePython

Python code for estimating 3DMM parameters using our very deep neural network

This page contains end-to-end demo code that estimates the 3D facial shape and texture directly from an unconstrained 2D face image. For a given input image, it produces a standard ply file of the face shape and texture. It accompanies the deep network described in our paper [1].

This release is part of an on-going face recognition and modeling project. Please, see this page for updates and more data.

A new version of this code which includes also pose and expression fitting, is being prepared and will be released soon.

Teaser

Features

  • End-to-End code to be used for 3D shape and texture estimation directly from image intensities
  • Designed and tested on face images in unconstrained conditions, including the challenging LFW, YTF and IJB-A benchmarks
  • The 3D face shape and texture parameters extracted using our network were shown for the first time to be descriminative and robust, providing near state of the art face recognition performance with 3DMM representations on these benchmarks
  • No expensive, iterative optimization, inner loops to regress the shape. 3DMM fitting is therefore extremely fast

Dependencies

Library requirements

The code has been tested on Linux only. On Linux you can rely on the default version of python, installing all the packages needed from the package manager or on Anaconda Python and install required packages through conda. A bit more effort is required to install caffé.

Data requirements

Before running the code, please, make sure to have all the required data in the following specific folder:

  • Download our CNN
  • Download the Basel Face Model
  • Download dlib face prediction model and move the .dat file into the dlib_model folder.
  • Move the CNN model (3 files: 3dmm_cnn_resnet_101.caffemodel,deploy_network.prototxt,mean.binaryproto) into the CNN folder
  • Copy the Basel Face Model (01_MorphableModel.mat) in the same folder of demo.py file.
  • Run the script python trimBaselFace.py. This should output a file BaselFaceModel_mod.mat and remove the original one automatically.

Usage

Input

The demo script can be used from the command line with the following syntax:

$ Usage: python demo.py <inputList> <outputDir> <needCrop> <useLM>

where the parameters are the following:

  • <inputList> is a text file containing the paths to each of the input images, one in each line.
  • <outputDir> is the path to the output directory, where ply files are stored.
  • <needCrop> tells the demo if the images need cropping (1) or not (0). Default 1. If your input image size is equal (square) and has a CASIA-like [2] bounding box, you can set <needCrop> as 0. Otherwise, you have to set it as 1.
  • <useLM> is an option to refine the bounding box using detected landmarks (1) or not (0). Default 1.

Example for <inputList>:

data/1.jpg
data/2.jpg
....

Output

The demo code should produce an output similar to this:

user@system:~/Desktop/3dmm_release$ python demo.py input.txt out/
> Prepare image data/1.jpg:
>     Number of faces detected: 1
> Prepare image data/2.jpg:
>     Number of faces detected: 1
> CNN Model loaded to regress 3D Shape and Texture!
> Loaded the Basel Face Model to write the 3D output!
> Processing image:  tmp_ims/2.png   2.png   1/2
> Writing 3D file in:  out//2.ply
> Processing image:  tmp_ims/1.png   1.png   2/2
> Writing 3D file in:  out//1.ply

The final 3D shape and texture can be displayed using standard off-the-shelf 3D (ply file) visualization software such as MeshLab. Using MeshLab, the output may be displayed as follows:

user@system:~/Desktop/3dmm_release$ meshlab out/1.ply

user@system:~/Desktop/3dmm_release$ meshlab out/2.ply

which should produce something similar to:

Teaser

Citation

If you find this work useful, please cite our paper [1] with the following bibtex:

@inproceedings{tran2017regressing,
  title={Regressing Robust and Discriminative 3D Morphable Models with a very Deep Neural Network},
  author={Tran, Anh Tuan and Hassner, Tal and Masi, Iacopo and Medioni, Gerard},
  booktitle={Computer Vision and Pattern Recognition (CVPR)},
  year={2017}
}

Troubleshooting

Problem: Old Caffe Engine

"F0210 10:49:17.604714 24046 net.cpp:797] Check failed:
target_blobs.size() == source_layer.blobs_size() (5 vs. 3) Incompatible
number of blobs for layer bn_conv1"

Solution: install caffe 1.0.0-rc3 or above.

For more info on caffe verson please see https://github.com/BVLC/caffe/releases

To check your caffe version from python:

In [3]: import caffe
In [4]: caffe.__version__
Out[4]: '1.0.0-rc3'

References

[1] A. Tran, T. Hassner, I. Masi, G. Medioni, "Regressing Robust and Discriminative 3D Morphable Models with a very Deep Neural Network", arxiv pre-print 2016

[2] Dong Yi, Zhen Lei, Shengcai Liao and Stan Z. Li, "Learning Face Representation from Scratch". arXiv preprint arXiv:1411.7923. 2014.

Changelog

  • Jan 2017, First Release

Disclaimer

The SOFTWARE PACKAGE provided in this page is provided "as is", without any guarantee made as to its suitability or fitness for any particular use. It may contain bugs, so use of this tool is at your own risk. We take no responsibility for any damage of any sort that may unintentionally be caused through its use.

Contacts

If you have any questions, drop an email to anhttran@usc.edu , hassner@isi.edu and iacopoma@usc.edu or leave a message below with GitHub (log-in is needed).