whztt07 / emotionchallenge Goto Github PK

If you use these models or code in your research, please cite:

@inproceedings{guo2017multi,
  title={Multi-modality Network with Visual and Geometrical Information for Micro Emotion Recognition},
  author={Guo, Jianzhu and Zhou, Shuai and Wu, Jinlin and Wan, Jun and Zhu, Xiangyu and Lei, Zhen and Li, Stan Z},
  booktitle={Automatic Face \& Gesture Recognition (FG 2017), 2017 12th IEEE International Conference on},
  pages={814--819},
  year={2017},
  organization={IEEE}
}

@article{guo2018dominant,
  title={Dominant and Complementary Emotion Recognition from Still Images of Faces},
  author={Guo, Jianzhu and Lei, Zhen and Wan, Jun and Avots, Egils and Hajarolasvadi, Noushin and Knyazev, Boris and Kuharenko, Artem and Jacques, Julio CS and Bar{\'o}, Xavier and Demirel, Hasan and others},
  journal={IEEE Access},
  year={2018},
  publisher={IEEE}
}

First, you should generate the crop and aligned data on test Chanllenge dataset. Just change to crop_align dir

python landmark.py

The crop and aligned of test data(final evaluation phase data) of 224x224 will place in $ROOT/data/face_224

Then change to cnn dir, just type

python extract.py

It will load data preprocessed and caffe model to generate labels named predictions.txt and predictions.zip for test data. All details were considered.

In this repo, some directory path may be confused, just be careful, contact me if any questions occured.

The trained caffe model is just a experiement model, it may not has the best perfomance in this challenge.

Just upload predictions.zip to submit window then.

We use Dlib to do face and landmark detection, and use landmark to do face cropping and alignment, then we use Caffe to with landmark and cropping image to train a cnn model to do the face expression recognition task.

First, run landmark.py to get all the origin image landmark, then build the crop_align binary, and run crop_align.py to get all the 224x224 size image.

Build crop_align

cd crop_align
mkdir build
cd build
cmake ..
make

All the preprocessed data except the images are in data dir.

Change to cnn dir, run prepare_data.py to prepare training, validation and test data. Then run train_val.sh to start training.

Just run extract.py to generate the result, the input is the test image and its landmark offset info.

We use the landmark offset and image info to do this task. In detail, the landmark offset is calculated by substraction of 224x224 image landmark and each id's mean landmark, and we concact this feature to modified alexnet's last output feature. We change softmax loss to hinge loss to get a little better result.

More detail is in the fact_sheet.tex.