What I learned merging seven public datasets to train a single-class shelf detector

Description and use

Trained a YOLO11n empty‑shelf detector on 11,667 CC BY 4.0 images from seven public datasets, using Apple MPS for training and deploying on a Raspberry Pi 3 with zero cloud GPU cost.

Problem

Build a commercial‑ready empty‑shelf detector with minimal licensing risk and zero cloud GPU cost.

Outcome

Merged 11,667 CC BY 4.0 images, trained YOLO11n on Apple MPS, achieved mAP50 0.844 on a held‑out test set, and deployed the model on a Raspberry Pi 3 with zero cloud GPU spend.

Workflow

• 1. Identify use‑case: empty‑space detection only.
• 2. Filter datasets by licence (strict CC BY 4.0).
• 3. Select datasets with empty‑space or derivable annotations and ≥200 images.
• 4. Merge datasets, remap annotations to single class, preserve original splits, redistribute to 80/10/10 split.
• 5. Deduplicate and check cross‑dataset contamination using perceptual hashes.
• 6. Fine‑tune YOLO11n on Apple MPS (mps device, imgsz 640, batch 16, epochs 60).
• 7. Evaluate on held‑out test set (277 images, 1,255 instances).
• 8. Export model to NCNN for Pi 3 inference.
• 9. Deploy on Pi 3 and run proof‑of‑concept scans.
• 10. Document results and limitations.
• 11. Plan for future commercial deployment and licence compliance.
• 12. Publish operational metrics when available.

Evidence

"I merged seven public Roboflow datasets into 11,667 images — all CC BY 4.0 — trained YOLO11n on Apple MPS, and reached mAP50 0.844 on a held‑out test set of 277 images and 1,255 instances."

Source

"The licence filter was strict: CC BY 4.0 only."

Source

"Training used Ultralytics on my Mac with Apple's MPS backend."

Source