Grocery Product Detection — NM i AI Competition
Ensemble model for detecting and classifying grocery products on store shelves.
Models
- yolo26x_1600.onnx — YOLO26x detector trained at 1600px resolution
- rfdetr_large_704.onnx — RF-DETR Large detector at 704px resolution
- efficientnet_b4_classifier.safetensors — Product classifier (356 classes, FP16)
- yolo26x_best.pt — YOLO26x PyTorch weights (accuracy-focused, 500 epochs)
- yolo26x_finetuned.pt — Fine-tuned on refined Roboflow dataset
Architecture
Two-stage ensemble:
- Detection: YOLO26x + RF-DETR with Weighted Boxes Fusion
- Classification verification: EfficientNet-B4 corrects low-confidence class predictions
Competition Score
- Best public leaderboard: 0.9021
- Scoring: 0.7 × detection_mAP@0.5 + 0.3 × classification_mAP@0.5
Training Data
- 248 shelf images from Norwegian grocery stores
- 22,731 COCO-format bounding box annotations
- 356 product categories
- Additional refined annotations via Roboflow