Developer trains commercial-ready shelf-gap detector using 7 free datasets, zero cloud cost

A developer built a single-class empty-shelf detector by merging seven public Roboflow datasets totalling 11,667 images, training the YOLO11n model on Apple MPS hardware at no cloud GPU expense. Rather than identifying individual products, the system detects only empty shelf spaces, then cross-references a planogram to determine which SKU is missing — a design choice that simplifies training and generalises across store types. The model achieved a mAP50 score of 0.844 on a held-out test set and runs as a proof-of-concept on a Raspberry Pi 3. A strict licence filter — allowing only CC BY 4.0 datasets — was applied throughout, excluding several well-annotated sources that carried non-commercial or unclear terms. The developer argues that licence compatibility is more critical than dataset volume when building models intended for potential commercial deployment.