NM i AI 2026 — Grocery Shelf Detection
YOLO26l ensemble + MobileNetV3 classifier for detecting and classifying 356 grocery products on store shelves.
Competition score: 0.8656 (NM i AI 2026, NorgesGruppen task)
Files
| File | Description |
|---|
yolo26l_v1_best.pt | YOLO26l v1, mAP50=0.628, batch=2, 400 epochs |
yolo26l_v2_best.pt | YOLO26l v2, mAP50=0.636, batch=4, patience=75 |
mobilenetv3_classifier.pt | MobileNetV3-Small, 87.51% val accuracy |
Pipeline
YOLO26l v1 + v2 → Weighted Box Fusion → MobileNetV3 classifierTraining
- Detector: YOLO26l, imgsz=1280, copy_paste=0.3, mixup=0.15, cls=1.0
- Classifier: MobileNetV3-Small, 24,308 crops, stratified split, label smoothing=0.1
- Hardware: NVIDIA L4 24GB
Code
