One-stage vs two-stage detectors

WHAT THIS TESTS The interviewer checks that you understand the pipeline difference and can articulate the speed-versus-accuracy trade-off rather than just naming models.

A GOOD ANSWER COVERS Two-stage (Faster R-CNN): stage one is a Region Proposal Network that suggests candidate object regions; stage two pools features from those regions and classifies and refines each box. This sparse, focused second stage yields high accuracy, particularly for small and overlapping objects, but the two passes make it slower. One-stage (YOLO, SSD): a single network densely predicts class scores and box offsets across the whole image grid in one forward pass, eliminating the proposal step. This is much faster and better for real-time use, but the dense predictions create heavy background-foreground class imbalance that historically hurt accuracy; focal loss, introduced with RetinaNet, downweights easy negatives and closes much of the gap. Modern one-stage models are both fast and accurate.

COMMON WRONG ANSWERS Saying two-stage is always more accurate today, ignoring how modern one-stage detectors with focal loss compete strongly. Or claiming one-stage skips anchors or NMS entirely. Or framing the difference as model size rather than the proposal mechanism.

LIKELY FOLLOW-UPS What does the RPN do and how are proposals scored. Why does one-stage detection cause class imbalance and how does focal loss fix it. Where does NMS fit in each. When would you pick each for a product.

ONE CONCRETE EXAMPLE For a real-time phone camera filter you choose YOLO for its single-pass speed at acceptable accuracy, since the user needs immediate feedback and a few missed detections are tolerable. For an offline medical imaging task where missing a small lesion is costly and latency does not matter, you choose Faster R-CNN, whose region proposal stage focuses computation on candidate regions and gives higher precision on small, overlapping objects. The decision hinges on the speed-versus-accuracy trade-off and the cost of errors, not on a blanket claim that one family is universally superior.