FERAtt: Facial Expression Recognition with Attention Net

This repository is under construction ...

Paper | arXiv

Pedro D. Marrero Fernandez¹, Fidel A. Guerrero-Peña¹, Tsang Ing Ren¹, Alexandre Cunha²

• 1 Centro de Informatica (CIn), Universidade Federal de Pernambuco (UFPE), Brazil
• 2 Center for Advanced Methods in Biological Image Analysis (CAMBIA) California Institute of Technology, USA

Introduction

Pytorch implementation for FERAtt neural net. Facial Expression Recognition with Attention Net (FERAtt), is based on the dual-branch architecture and consists of four major modules: (i) an attention module $$G_{att}$$ to extract the attention feature map, (ii) a feature extraction module $G_{ft}$ to obtain essential features from the input image $I$, (iii) a reconstruction module $G_{rec}$ to estimate a good attention image $I_{att}$, and (iv) a representation module $G_{rep}$ that is responsible for the representation and classification of the facial expression image.

Prerequisites

• Linux or macOS
• Python 3
• NVIDIA GPU + CUDA cuDNN
• PyTorch 1.5

Installation

$git clone https://github.com/pedrodiamel/pytorchvision.git
$cd pytorchvision
$python setup.py install
$pip install -r installation.txt

Docker:

docker build -f "Dockerfile" -t feratt:latest .
./run_docker.sh

Visualize result with Visdom

We now support Visdom for real-time loss visualization during training!

To use Visdom in the browser:

# First install Python server and client
pip install visdom
# Start the server (probably in a screen or tmux)
python -m visdom.server -env_path runs/visdom/
# http://localhost:8097/

How use

Step 1: Train

./train_bu3dfe.sh
./train_ck.sh

Citation

If you find this useful for your research, please cite the following paper.

@InProceedings{Fernandez_2019_CVPR_Workshops,
author = {Marrero Fernandez, Pedro D. and Guerrero Pena, Fidel A. and Ing Ren, Tsang and Cunha, Alexandre},
title = {FERAtt: Facial Expression Recognition With Attention Net},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2019}
}

Acknowledgments

Gratefully acknowledge financial support from the Brazilian government agency FACEPE.