Object detection

How this question breaks down

1. 1DetectorsCore27 minTurn a shelf photo into a box and a label for every facing, price tag and gap: choose the input size so small tags survive, a one-stage anchor-free detector with a feature pyramid, a duplicate filter that does not merge neighbours, and an evaluation that can count past 100 objects.
2. 2DataCore21 minLabel ~196 boxes per shelf photo without spending two hours on each: a tight guideline with ignore regions, extreme clicking and model pre-labels, splits by store, agreement measured in IoU, and a queue that picks the next bays to label from the model's doubts and a random stream.
3. 3On deviceCore26 minKeep boxes moving at 30 frames a second on phones Aislemark does not control: a small distilled int8 model on a worker thread for live boxes, tiled passes for the captured photo, detect-then-track between frames, a thermal state machine that degrades gracefully, and models delivered per device class and checked against the server.

Work through the parts in order and you’ve covered the full design. Each part is one topic.