Improve YOLO26 Crate Detection

Job ID: 40529820

Budget: €50 – €100 EUR

My current YOLO 26 model struggles with Eurobox and bread-crate detection, hovering below 50 % accuracy. With only ~100 training images (each holding 30–40 crates), I need to push performance past 94 % without relying on power-hungry cloud instances.

I’m open to every practical angle—tighter algorithmic tuning, smart preprocessing and creative data augmentation—so long as the final solution can run locally on a mid-range GPU or even CPU if possible. Feel free to experiment with lighter YOLO variants, pruning, quantisation, mosaic augmentation, rotation/flip tricks, colour tweaks or any other ideas you trust; I care about the end result and the ability to reproduce it on my hardware.

Acceptance criteria
• Provide the updated model weights, training script and a concise README.
• Demonstrate ≥94 % mAP (or equivalent class-level accuracy) on a hold-out set I’ll supply.
• Keep inference times reasonable for real-time (≤50 ms per image on RTX 3060 or similar).

If you can get me there efficiently, let’s talk—quick wins, clear metrics and clean code are what I’m after.