Vision task definition and acceptance bar
Most vision projects that stall were never clear on what the model must output or what happens next with that output. Classification, detection, segmentation and counting need different labels, different metrics and different budgets, and the cost of a missed defect is rarely the cost of a false alarm.
First tasks
- Ask the person who acts on the output what they do with each prediction
- Write the cost of a miss and the cost of a false alarm side by side
- Agree one numeric acceptance bar for the most important class
- Record the latency and hardware limits of the target location
- Revisit the acceptance bar with the action owner @recurring(yearly)