How do you evaluate the performance of a computer vision model?

Instruction: Describe the methods and metrics used for evaluating model performance.

Context: This question assesses the candidate's ability to critically analyze the effectiveness of computer vision models using appropriate evaluation metrics.

Updated

Example Answer

I would choose metrics from the task and the failure cost. For classification, inspect per-class precision and recall rather than accuracy alone. For detection, evaluate matched boxes and classes with a defined IoU threshold and average precision convention; IoU by itself cannot count missed objects or false alarms. For deployment, measure the complete preprocessing-to-output pipeline, including tail latency, memory and important data slices. A model can improve average quality while becoming worse on small objects or missing its deadline.

Make it your own

Illustrative review table: overall AP, small-object AP, night-time recall, false alarms per image, p95 pipeline latency and peak memory. State dataset and matching rules.

Why this works

This prevents one convenient number from hiding operational failures. Follow-up: what would you prioritize if AP improves but night-time recall drops?

Sources: TorchVision detection models, scikit-learn model evaluation.

Related Questions