Facial Emotion Recognition with PyTorch

Project for Software Development Practical: Computer Vision and Deep Learning @ LMU Munich in SS2025. It demonstrates (real-time) facial emotion recognition using a custom CNN architecture with ResNet-style blocks and Squeeze-and-Excitation (SE) channel attention.

📄 Read the report here

Overview

This project implements a facial emotion recognition model that can:

• Classify 6 emotions: Angry, Disgust, Fear, Happy, Sad, Surprise
• Provide real-time webcam inference
• Generate Grad-CAM visualizations for Explainable AI
• Process video files with emotion predictions and heatmap overlays

Project Structure

├── model.py                        # Core CNN architecture
├── live_demo.py                    # Real-time webcam demo
├── inference_video.py              # Video processing with Grad-CAM
├── gradcam.py                      # Grad-CAM implementation
├── csv_gen.py                      # .csv generation file
├── final_model_ferplus.ipynb       # Main training notebook
├── best-weights.pt                 # Model weights
├── README.md                       # This file
├── experiments_ferplus/                        
│   ├── ferplus_moderate_aug.ipynb  # FER+ experiment moderate augmentation
└── experiments_fer2013/            # Various Augmentation strategy 
    ├── final_model_fer2013.ipynb   # Final model FER-2013
    ├── no_aug.ipynb
    ├── horizontalflip.ipynb
    ├── verticalflip.ipynb
    └── moderate_aug.ipynb

Installation

1. Install dependencies:

pip install -r requirements.txt

Usage

Real-time Webcam Demo

Run the live emotion detection demo:

python3 live_demo.py

Press 'q' to quit the demo.

Video Processing with Grad-CAM

Process a video file with emotion predictions and attention heatmaps:

python3 inference_video.py --input path/to/video.mp4 --output output_with_emotions.mp4

.csv file generation

Iterates through images in a folder and outputs the corresponding classification scores in a csv file

python3 csv_gen.py path/to/imagefolder --ouptput predictions.csv

Training

The main training pipeline is in final_model_ferplus.ipynb.

Experimental Variations

The experiments_fer2013/ directory contains experiments with different augmentation strategies:

1. no_aug.ipynb: Baseline without augmentation (~60% accuracy)
2. horizontalflip.ipynb: Basic horizontal flipping (~62% accuracy)
3. verticalflip.ipynb: Vertical flipping experiments (~57% accuracy)
4. moderate_aug.ipynb: Moderate augmentation (~70% accuracy)

The experiments_ferplus/ directory contains experiment on the FER+ dataset with moderate augmentation strategy.