LEARN. PRACTICE. GROW.Concept and mechanism
Classification assigns image categories; detection locates instances with boxes. Data preparation must match the required output and separate training from evaluation examples. Precision measures how many positive predictions are correct; recall measures how many actual positives were found. Global accuracy can look excellent when the critical class is rare. Threshold selection involves false-positive cost, false-negative cost, and review capacity. Publishing an iteration and training an iteration are distinct Custom Vision steps: the prediction target must point to the approved version. In video, transcripts and time-based insights locate segments but do not establish identity or authorization.
Guided application
In fictional attachment triage, first define what must be located and who reviews uncertain results. Evaluate each class using representative examples, including degraded images and rare cases. If 80 of 100 alerts are correct and there are 200 actual positives, precision is 80% and recall is 40%; half the alerts does not mean half the positives were found. Retain the published version and acceptance set to compare releases. Include lifecycle in planning: Custom Vision and Image Analysis 4.0 have announced retirement on 2028-09-25. This requires assessing migration for new long-running projects without generalizing retirement to all vision services.
A model with 98% accuracy can miss many critical documents if those documents are a small fraction of the total.
Common pitfalls
Accuracy as risk; training as publishing; threshold without review-capacity assessment; one product retirement date applied to the whole family.
Related topics: Services, identity, and operations · Generation, grounding, and evaluation · Agents, tools, and control
Relate per-class metrics, published version, and support horizon to the release decision.
Reference: Custom Vision model quality and representative data · AI-102 historical skills measured 2025-12-23; exam retired 2026-06-30