Vision systems in production.
WHAT THE PIPELINE INVOLVES
Image capture
Preprocessing: resizing, normalisation, orientation
Inference
Postprocessing: thresholds, suppression of overlapping detections
Action or review
WHAT ORIENTATION PROBLEMS CAUSE
Images rotated by metadata rather than pixels, silently degrading accuracy.
WHAT TO DO
Apply orientation explicitly during preprocessing.
WHAT AFFECTS REAL ACCURACY MOST
Capture conditions: lighting, distance, angle, motion, focus.
WHAT THAT MEANS
Improving capture frequently beats improving the model.
WHAT TO PROVIDE USERS
Guidance at capture: framing, lighting, steadiness.
WHY
It raises accuracy more cheaply than any modelling.
WHAT TO RUN ON DEVICE WHERE POSSIBLE
Inference, avoiding upload of images entirely.
WHY THAT MATTERS HERE
Images are large, connections are metered, and upload frequently fails.
WHAT THAT REQUIRES
A model small enough, and an update mechanism.
WHAT TO MONITOR
Confidence distribution Rejection and review rates Accuracy on a sampled, manually checked set
WHAT TO ALWAYS PROVIDE
A manual route when recognition fails.
