What is Computer Vision?

The task types, which matter more than the label

'Computer vision' covers several genuinely different problems, and choosing the wrong framing is a common early error. Classification assigns a label to a whole image — is this product defective. Detection locates objects with bounding boxes and counts them. Segmentation classifies every pixel, which is what you need for measuring area or precise boundaries. Keypoint estimation finds specific landmarks, as in pose tracking. Tracking maintains object identity across video frames. OCR extracts text.

The framing determines labelling cost, which dominates budget. A classification label takes a second; a segmentation mask can take many minutes per image. Choosing segmentation when detection would answer the business question can multiply annotation cost by two orders of magnitude.

Where it delivers

Industrial inspection is the most consistently successful application. Fixed camera, controlled lighting, repetitive judgement, and a clear economic value per catch. Conditions are constrained, which is exactly what vision models want.

Document processing — extracting fields from invoices, forms, and identity documents — has become reliable and is often the fastest route to measurable savings.

Medical imaging shows strong results in narrow, well-defined tasks, with regulatory approval rather than model accuracy as the usual constraint.

Retail and logistics use it for shelf monitoring, damage assessment, and package dimensioning. Agriculture uses it for yield estimation and disease detection.

Almost all of this now starts from a pretrained backbone — see transfer learning — so a production-quality model from a few thousand labelled images on a single GPU is a realistic expectation rather than an ambitious one.

What actually goes wrong

Failures in deployed vision systems are overwhelmingly environmental rather than algorithmic. Lighting changes between shifts and seasons. A camera is nudged and the framing shifts. A supplier changes packaging. A lens accumulates dust. The model was trained on one visual world and is now operating in another — data drift with unusually literal causes.

The practical defences are unglamorous and effective. Control the physical environment where you can; fixed mounting and consistent lighting are worth more than a better architecture. Collect training data across the full range of real conditions, including night shifts and weather. Use augmentation that reflects genuine variation — brightness, blur, rotation, occlusion. Monitor input image statistics in production so a dirty lens raises an alert rather than silently degrading accuracy.

Two further cautions. Models exploit spurious correlations readily: if every defective part in your dataset was photographed on a different bench, the model may learn the bench. Explicitly test on data collected under conditions absent from training. And be careful with class imbalance — defect rates are often well under 1%, so accuracy is meaningless and you should be reading precision and recall from a confusion matrix instead.