In this repo, we include the 1st Place code to FG2020 Compound Emotion challenge;
We use Dlib to do face and landmark detection, and use landmark to do face cropping and alignment, then we use Pytorch version 1.1 to with landmark and cropping image to train cnn model to do the face expression recognition task.
First, you should generate the crop and aligned data on Chanllenge dataset. Change to crop_align dir
Build crop_align
mkdir build
cd build
cmake ..
make
cd ..
Dataset path need to be change in landmark.py and crop_align.py
python landmark.py ### gen train_landmark.json, val_landmark.json, test_landmark.json
python crop_align.py ### crop face images
python landmark_224.py ### gen landmark_224.json
python prepare_data.py ### gen val_ld.txt, train_ld_shuffle.txt
The crop and aligned data of 224x224 will place in $ROOT/faces_224/
Then change to $ROOT dir, just type
python pred_csv.py
python gen_submit.py
It will load data and model to generate labels named predictions.txt and predictions.zip for test data.
The trained model is just a experiement model, it may not has the best perfomance in this challenge.
Just upload predictions.zip to submit window then.
Run sh train_kd.sh to start training,(10-folder cross-validation).