Skip to content
/ BEP Public

This repository is devoted to the development of the facial emotion recognition (FER) system as a final bachelor project at the TU/e. Realised by Blazej Manczak. Supervisors: Dr. Laura Astola (Accenture) and Dr. Vlado Menkovski (TU/e)

Notifications You must be signed in to change notification settings

bmanczak/BEP

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

57 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Accurate, fast and automatic facial expression recognition

This repository is devoted to the development of the facial emotion recognition (FER) system under resource constraint as a final bachelor project at the TU/e. Realized by Blazej Manczak. Supervisors: Dr. Laura Astola and Dr. Vlado Menkovski. Please see the Report.pdf for a comprehensive overview of the methods used in this repository. A couple of things about this repo:

  • The complete workflow including the classification trained model is demonstrated in this Google Colab notebook
  • I am using two datasets: AffectNet and RAF-DB. These datases have 7 emotion categories: anger, disgust, fear, hapiness, sadness, suprise, neutral.
  • The repo contains a preprocessing pipeline responsible for CLAHE illumination normalization, facial landmark detection (using MTCNN), face alignment, preprocessing as done for VGG-Face, or SE-ResNet-50 models, and face extraction. The scrirpt can be found in BEP/ModelInference/withMtcnn.
  • The method described in the Report achieves close to the state-of-the-art results on the corresponding validation sets: 87.8% on the RAF-DB dataset and 64.06% on AffectNet.
  • The method is capable of real-time processing: 4.5-5 FPS on a 2016 baseline MacBookPro 13" (CPU only).
  • The model(s) can be successfully quantized with minimal accuracy loss, giving an opportunity of working in resource-constrained environments (~7 FPS on Google Dev Board).
  • Grad-Cam algorithm can be employed to create a coarse localization map highlighting important aspects of the image.

BlueJayTeam

Environment

We provide a conda environment called BEP which contains all packages you need to execute the scripts in this repo.

More information about the training and performance can be found on my Comet.ml experiment page.

About

This repository is devoted to the development of the facial emotion recognition (FER) system as a final bachelor project at the TU/e. Realised by Blazej Manczak. Supervisors: Dr. Laura Astola (Accenture) and Dr. Vlado Menkovski (TU/e)

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published