Design an active learning loop for detection

WHAT THIS TESTS: It probes whether you can build a continuous active-learning loop that spends a scarce annotation budget exactly where it most improves the detector, rather than labeling at random.

A GOOD ANSWER COVERS: Start by training a detector on a small seed labeled set. Run it across the large unlabeled pool to get predicted boxes and confidences. Define an acquisition function: uncertainty signals include low maximum class confidence, many boxes hovering near the score threshold, or high disagreement under test-time augmentation, an ensemble, or Monte-Carlo dropout. Combine uncertainty with a diversity or representativeness term, often via feature-space clustering or core-set selection, so the chosen batch covers genuinely different scenes rather than near-identical frames. Send the top batch to human annotators through a labeling tool, ingest the new annotations, retrain or fine-tune the detector, and repeat. Track a held-out validation curve to decide when added labels stop yielding gains.

COMMON WRONG ANSWERS: Selecting only the single most uncertain images, which returns redundant similar shots from the same clip and wastes budget. Ignoring class imbalance, so rare object categories never get queried. Retraining from scratch every cycle when fine-tuning would suffice. Forgetting a stopping criterion, so the loop runs forever.

LIKELY FOLLOW-UPS: How do you measure uncertainty for a detector that emits many boxes per image. Why add a diversity term on top of uncertainty. How do you keep annotator label noise from compounding over rounds. How often should you actually retrain versus accumulate labels.

ONE CONCRETE EXAMPLE: With a hundred thousand unlabeled driving frames and budget for two thousand labels, you label five hundred seed frames, train, then each round pick the five hundred frames scoring highest on uncertainty times diversity, such as dusk scenes with low-confidence pedestrian boxes from different roads, label them, and retrain. This typically reaches the target mAP with far fewer labels than uniform random sampling would require for the same accuracy.