Model & API console

Detector settings, API and training

Detector settings

Model weights

Confidence threshold

0.45

NMS IoU

0.50

Max detections

600

Input size

Counting API

POST /api/v1/count

Multipart image in, JSON count out. The dashboard uses exactly this endpoint.

curl -X POST https://shelfcount.local/api/v1/count \
  -F "image=@aisle-04_vinegar_oils.jpg" -F "conf=0.45"

response

Pick a photo and send a request to see the JSON response.

Fine-tuning runs

Swap in a fine-tuned model without touching the dashboard.

shelf-density-v2in production412 images · 1 class09 Nov 2026mAP 0.912

shelf-density-v1180 images · 1 class04 Nov 2026mAP 0.884

baseline-coco-retail0 images · 1 class02 Nov 2026mAP 0.731

README

shelfcount/README.md

YOLOv8m pre-trained on SKU-110K (11,762 densely packed shelf images). It handles 100+ objects per image, which generic COCO models do not. Inference runs on CPU in ~0.3 s at 1280 px, on GPU in ~25 ms.