Fractal Hunter — Robust Image Classification Under Extreme Distortion
About the project
Up the Gain — Electric & Distortion-Ready
Blorbos (toon penguins) hide by transforming into distorted, non-intuitive representations. You are given a dataset derived from a small set of source images, transformed through unknown but structure-preserving processes.
📦 Dataset & full documentation: https://github.com/andrewrgarcia/fractal-hunter-dataset
🔬 Kaggle (for experimentation): https://www.kaggle.com/datasets/drandrewgarcia/fractal-hunter-extreme-distortion
Task
Build a model that determines whether an image originated from Blorbo.
Constraints
- Transformation is unknown
- No explicit inversion tricks
- Must generalize to unseen images and transformations
Goals
- Robust classification under severe distortion
- Clear approach + failure analysis
Evaluation
Held-out test set uses unseen transformations
Stretch
- Learn invariant representations
- Recover structure without knowing the transform
Stack: Python · PyTorch / sklearn
About Prince Fractalius ↗
Sworn to the Crown of Recursion
Operator, Fractal Kingdom Enforcement Guild (FKEG) Jurisdiction: unstable geometries and unauthorized Blorbo proliferation 💀🐧
Status: Still deployed. Compensation: Still pending.
Per Inhuman Resources, I require form CJ-9021 from the Burning Empire. The Empire requires form X549 from FKEG. FKEG will not issue X549 without CJ-9021. The Empire will not issue CJ-9021 without X549.
The loop persists. And still, I remain.
Working. Waiting. Caught in a bureaucracy that does not end.
My life for the Fractal King! ⚔️
— Prince Fractalius
*Prince Fractalius is Andrew Garcia's avatar, Founder of WorkDog.