MogMeter

Inspiration

The concept of "mogging" (a term from bodybuilding/fitness culture referring to dominating or overshadowing others through physical presence) has become a popular topic in fitness communities. We saw an opportunity to create a fun, technological approach to quantify this subjective concept using modern AI techniques.

What it does

MogMeter is an AI-powered application that analyzes images to determine a "mog score" - a measure of how much someone is "mogging" in a given photo. The system uses a fine-tuned ConvNeXT model to process images and output a confidence score between 0 and 1, effectively creating a "Mog Meter" that can quantify presence and dominance in photos.

How we built it

Used PyTorch Lightning for creating and training a deep learning model based on the ConvNeXT architecture
Implemented transfer learning by fine-tuning the pretrained ConvNeXT model
Created a custom dataset pipeline with proper image normalization and augmentation
Used a sophisticated data organization system to handle training and validation splits
Implemented early stopping and model checkpointing for optimal training
Built with modern best practices like learning rate scheduling and dropout for regularization
Used Flask for the web application interface

Challenges we ran into

Managing image preprocessing across various formats while maintaining aspect ratios
Implementing proper data augmentation to prevent overfitting while maintaining relevant features
Balancing model complexity with training efficiency by selectively freezing and unfreezing ConvNeXT layers
Creating an efficient data pipeline that could handle various image formats and sizes -Lack of objective data points

Accomplishments that we're proud of

Successfully implemented a complex deep learning pipeline using modern best practices
Created a robust image normalization system that handles various input formats
Built an efficient training system with proper validation and early stopping
Developed a clean, modular codebase with proper separation of concerns

What we learned

Advanced PyTorch Lightning techniques for model training and validation
Transfer learning best practices with large models like ConvNeXT
Efficient image preprocessing and augmentation techniques
Proper model training practices including learning rate scheduling and early stopping