Precision vs recall in object detection.

WHAT THIS TESTS The interviewer wants crisp definitions, awareness that they trade off, and judgment about which matters in a given application.

A GOOD ANSWER COVERS Precision is the fraction of the model's positive detections that are actually correct: true positives divided by true positives plus false positives. It answers, when the model says there is an object, how often is it right. Recall is the fraction of real objects the model successfully finds: true positives divided by true positives plus false negatives. It answers, of all the objects that exist, how many did we catch. In detection these depend on the confidence threshold and on the IoU criterion for counting a box as correct. Lowering the threshold catches more true objects, raising recall but adding false positives that lower precision; raising it does the opposite. You choose based on the cost of each error type. When a miss is dangerous, you favor recall; when a false alarm is expensive or erodes trust, you favor precision. Average precision and the precision-recall curve summarize the whole trade-off.

COMMON WRONG ANSWERS Swapping the definitions. Claiming you can maximize both freely with no trade-off. Forgetting the threshold and IoU dependence in detection specifically.

LIKELY FOLLOW-UPS F1 score, mAP across IoU thresholds, how to pick an operating point, and behavior on imbalanced data.

ONE CONCRETE EXAMPLE For a self-driving car's pedestrian detector you prioritize recall: missing a real pedestrian is catastrophic, so you accept some false alarms and brake conservatively. For a consumer photo app that auto-tags faces, you prioritize precision: a wrong tag is annoying and visible, so it is better to skip uncertain detections than to label incorrectly.