AI-900 Practice Question: Describe features of computer vision workloads on Azure

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  The percentage of training images that contained this type of object

  Why it's wrong here

  The frequency of an object class in the training set is a dataset statistic, often called the prior or class prevalence, which tells you how often that label appears among all training images. It does not change per image at inference time; a specific detection's confidence score depends on the input image and the learned model parameters. For instance, a rare class can still have high confidence on a clear example, and a common class can have low confidence on an ambiguous one. The confidence score is conditional on the current input, not on historical class distribution.

• ✓

  The model's certainty about a detection, used to set thresholds balancing false positives vs misses

  Why this is correct

  A confidence score is the model's probabilistic estimate, per detection, that the bounding box actually contains the target object. Setting a threshold on this score creates an explicit precision/recall tradeoff: a higher threshold keeps only high-certainty detections (fewer false positives but more missed true objects), while a lower threshold catches more objects at the cost of more false positives. This score is computed at inference time for each individual detection, not as an aggregate training metric.

• ✗

  The accuracy of the model measured on the test dataset during training

  Why it's wrong here

  Test accuracy is an aggregate metric computed across the entire labeled test dataset after training, reporting the overall fraction of correct predictions. It describes model performance in bulk, not the model's confidence in any single object detection. A confidence score is a per-inference output: each detected box gets its own score, which can vary even when two detections are both ultimately correct or incorrect. Test accuracy summarizes how often the model is right overall; the confidence score tells you how sure the model is about a specific prediction.

• ✗

  A quality rating assigned by human reviewers to confirm the detection is correct

  Why it's wrong here

  A human reviewer's quality rating is an external verification step, such as re-labeling or approving detections in a review UI. In object detection services, the confidence score is generated automatically by the model's inference engine as a float value (e.g., 0 to 1) for every predicted box; it requires no human judgment and is returned with the detection results. Human feedback may be used to retrain a model, but it is not what the score represents—the score exists regardless of any human review.

Go deeper

Related to this question