Gardenia is a face alignment project using Convolutional Neural Network. The project illstrate the training and testing process on 300W dataset.
- Download and install [caffe from xuzhenqi] (https://github.com/xuzhenqi/caffe/tree/eye_detection). We use both the c++ and python interfaces.
- Matlab
We need first to prepare the data, suppose the original data root path are data/. Run
scripts/prepare.sh
To crop the testset and augment the trainset, and we will get the data/augment/train.txt, data/train_mean.blob and data/crop/test.txt.
After preparing the data, we can run
scripts/train.sh
to train our model. The prototxt are set done on proto folder.
After training a model, suppose the model path to be $MODEL, run
scripts/inference.sh # You may need to set some options in inference.sh to predict labels
scripts/evalution.sh # report mean err on 300W dataset and CED curve.
The inference.sh will use trainded model to predict shapes, the pretrainded can be downloaded from [here] (http://pan.baidu.com/s/1cBOph8). Feel free to put an issue if you encounter problems.
Tested on Ubuntu 16.04, Nvidia TITAN X, without cudnn, without image load and data processing.
| proto | num of layers | average forward speed | average backward speed |
|---|---|---|---|
| v3 | 11 | 17.5707 ms / 1 sample | 38.4078 ms / 1 sample |
| v4 | 16 | 20.783 ms / 1 sample | 41.3371 ms / 1 sample |
| v5 | 26 | 22.3556 ms / 1 sample | 54.5015 ms / 1 sample |
| v6 | 36 | 25.0671 ms / 1 sample | 54.6937 ms / 1 sample |
| v7 | 41 | 26.0211 ms / 1 sample | 55.7797 ms / 1 sample |
@inproceedings{xu2016learning,
title={Learning Facial Point Response for Alignment by Purely Convolutional Network},
author={Xu, Zhenqi and Deng, Weihong and Hu, Jiani},
booktitle={Asian Conference on Computer Vision},
pages={248--263},
year={2016},
organization={Springer}
}
You can get a pdf copy from https://pan.baidu.com/s/1qXKo85Q