GeoGuessr Modelling
We trained a deep neural network that predicts the country of origin for a given image. The project tackles the challenging problem of geographical image classification with systematic data engineering and performance optimization.
Overview
The system uses transfer learning with a ResNet152 backbone combined with a custom classification head. Key achievements include:
Data Engineering
- Curated a balanced dataset of ~100,000 images
- Filtered outlier countries and balanced distributions to match real-world Google Maps data
- Transformed highly skewed initial distributions into uniform geographical coverage
Performance Optimization
- Reduced epoch training time from 30 minutes to 5 minutes (83% speedup)
- Created a tensor-based dataset with pre-extracted ResNet features
The result is an accurate country predictor that efficiently processes large-scale geographical image data.