A project that tries to yield good classification results with little data available using a pipeline of different models to improve.

Objective: For 65 classes with 50 images, try to make the best possible predictor, using MoE like model structure, data augmentation, background removal and other tools

This project is a continuation on an attempt three years ago to make a Pokemon-like game, based on sweet water fish from West Europe. The final metrics are the

• Highest Probability Prediction
• Top 5 Probability Prediction

The results of the different models used and their impact on the predictive power is explained below. Current implementation includes:

• SAM Model background removal
• Tree-like RESNET-50 structure
• Data Augmentation for increased data

Highest Score:

Current Project is maintained by Ian Ronk, available on [email protected]