Before you train a model: a computer vision feasibility checklist
Most computer vision problems are decided by the camera, the lighting and the data — long before the model.
Computer vision projects often begin with a model choice. In practice, whether a system will work is usually determined by more basic questions. We run through these before committing to a build.
Can a person do it from the same image?
If a trained person can't reliably see the defect, read the label or count the items from the camera's image, a model probably can't either. Start by reviewing real frames from the real camera position.
What changes over a day — and a year?
Lighting, weather, seasonal products, camera cleaning and packaging redesigns all shift the input. Data collected in one week in one condition rarely represents production.
What accuracy does the process need?
Counting for a daily report tolerates different errors than a safety alert. Define what a false positive and a false negative cost, then decide where people review results.
Where does inference run?
Bandwidth, privacy and latency often favour running models on-site and sending only events. That affects hardware, model size and how updates are deployed.