Custom Vision Classifier Evaluation: Precision and Recall
Short answer
CE Precision and recall are the two performance metrics Custom Vision exposes on the Performance tab after classifier training.
Question
After training is complete, you need to evaluate the classifier.
Which two metrics are available for review? Each correct answer presents a complete solution.
NOTE: Each correct selection is worth one point.
Key Takeaways
The trap is F-score: it is derived from precision and recall and sounds reviewable, but Custom Vision's Performance tab does not present it, while AUC and weighted accuracy belong to other evaluation tooling or do not exist.
Azure Custom Vision reports precision and recall per iteration and per tag when you evaluate a trained classifier, making C and E the two correct metrics. F-score, AUC and weighted accuracy are not surfaced by the service's performance report.
Common mistake: Choosing F-score (A) because it combines precision and recall and is a common classifier metric; Custom Vision does not display F-score as a reviewable performance metric.
Expert Analysis
What the question is really testing
After you train a Custom Vision multiclass or multilabel classifier, the Performance tab of the project reports two core measures per iteration and per tag: precision and recall. Precision answers "of the images the model labelled as this tag, how many were right?" and recall answers "of all the images that truly carry this tag, how many did the model find?" These are the two metrics the service exposes in its performance grid and threshold sliders, so C (precision) and E (recall) are the correct pair.
The attractive trap: F-score (A)
F-score looks like a very credible answer because it is simply the harmonic mean of precision and recall, and many Azure ML-flavoured evaluations report an F1 value. It also superficially satisfies the phrasing "which metrics are available for review" — but Custom Vision's own performance report does not list F-score as one of the reviewable metrics. It is a derived figure you could compute yourself, not a metric the service presents. That single distinction is what disqualifies it, even though the underlying numbers (precision and recall) are exactly those the service does show.
Why AUC (B) and weighted accuracy (D) fail
Area under the ROC curve is a threshold-independent evaluation normally surfaced in Azure Machine Learning experiments, not in the Custom Vision performance view — what Custom Vision surfaces is per-tag precision/recall at a probability threshold. "Weighted accuracy" is a distractor with no basis in Custom Vision at all; the service does not report an accuracy figure, weighted or otherwise, in its performance metrics.
Bottom line
The reviewable Custom Vision classifier metrics are precision and recall: C and E.
Official Reference
https://learn.microsoft.com/en-us/azure/ai-services/custom-vision-service/evaluation
This link opens the official vendor documentation in a new tab.
Exam Strategy
When a question names a specific Azure service's reporting UI, anchor on what that service actually surfaces rather than on generic ML theory — Custom Vision shows precision and recall, not derived or external metrics. Treat multi-select questions as two separate verifications: for each candidate, ask "does this exact product expose this exact metric?" Reject options that belong to a sibling service (Azure ML AUC) or that you would have to compute yourself (F-score).
Frequently Asked Questions
Does Azure Custom Vision show an F1 score for trained classifiers?
How do precision and recall differ when reviewing a Custom Vision classifier?
Why is area under the curve not the answer for Custom Vision evaluation?
Community Votes
50% of anonymous learners picked answer C. Votes are pick records left by other test-takers — they are not the verified answer.
Discussion 2 contributions
Compiled from public community threads and reader submissions, lightly edited for clarity. Handles and vote tallies are shown as recorded, and comments reflect the views of individual contributors.
From public community threads 2
-
AIBoss 👍 2Precision and recall are two different measurements of the effectiveness of a classifier: Precision indicates the fraction of identified classifications that were correct. For example, if the model identified 100 images as dogs, and 99 of them were actually of dogs, then the precision would be 99%. Recall indicates the fraction of actual classifications that were correctly identified. For example, if there were actually 100 images of apples, and the model identified 80 as apples, the recall would be 80%.
-
8ac3742 👍 1precision and recall Imagine a basket containing 10 apples and 10 oranges. You build a classifier to find all the apples. The model inspects the basket and picks out 5 fruits, claiming they are all apples. Upon inspection of the 5 fruits it picked: 4 are actually apples (True Positives) 1 is an orange (False Positive) 6 apples were left behind in the basket (False Negatives) precision = 4/5 recall = 4/10
Reader comments & corrections
No reader corrections yet — spotted an error or have a note? Share it below.
Related Analysis
Practice All AI-103 Questions
Access 88 questions with complete answers and detailed explanations.
View Full AI-103 Practice Test →