IA Data Hackathon (French Ministry of Armed Forces)

A computer vision challenge focused on vehicle classification using Transfer Learning (ResNet-18) and model explainability.

PythonPyTorchResNet-18Grad-CAMTransfer LearningMatplotlib

Completed in April 2024 - took 2 days to be built

View on GitHub

About the Project

Participated in the 'Projet 6milarite' hackathon organized by the French Ministry of Armed Forces.

Our team of 5 was tasked with training a model to classify 196 distinct car models from the Stanford Cars Dataset (16k images). We handled significant class imbalance by implementing data augmentation and strategic train/test splitting (70/30).

We fine-tuned a pre-trained ResNet-18 model, achieving 86% accuracy with a loss of 0.03. To meet the 'explainability' requirement, we generated Grad-CAM heatmaps, specifically analyzing false positive predictions to visualize exactly where the model's attention failed.

Motivation

I wanted to push beyond simple accuracy metrics and understand the 'why' behind AI decisions. This project allowed me to apply deep learning theory to a messy, real-world dataset while learning visualization tools for model diagnostics.

Challenges Overcome

•Tackling fine-grained classification where inter-class visual differences are minimal and easily confused by the model
•Managing GPU memory constraints while processing high-resolution heatmaps for thousands of validation images
•Justifying the trade-off between model complexity and interpretability to the jury

What I Learned

•Implementing Transfer Learning by freezing initial layers of a ResNet-18 model pre-trained on ImageNet
•Diagnosing model performance using Confusion Matrices and F1-scores/Recall analysis per class
•Developing a pipeline to automatically overlay Grad-CAM heatmaps on misclassified images for debugging
•Optimizing data pipelines with PyTorch DataLoaders and custom transformations (CenterCrop, Normalize)

About the Project

Participated in the 'Projet 6milarite' hackathon organized by the French Ministry of Armed Forces.

Challenges Overcome

•Tackling fine-grained classification where inter-class visual differences are minimal and easily confused by the model

•Managing GPU memory constraints while processing high-resolution heatmaps for thousands of validation images

•Justifying the trade-off between model complexity and interpretability to the jury

What I Learned

•Implementing Transfer Learning by freezing initial layers of a ResNet-18 model pre-trained on ImageNet

•Diagnosing model performance using Confusion Matrices and F1-scores/Recall analysis per class

•Developing a pipeline to automatically overlay Grad-CAM heatmaps on misclassified images for debugging

•Optimizing data pipelines with PyTorch DataLoaders and custom transformations (CenterCrop, Normalize)