. What is the core idea of a sliding-window detector, as built in this lesson?
A window classifier answers 'face or not, at this exact window,' and sliding it across every position turns that per-window classifier into a detector.
. In this lesson's box convention, how is each detected face represented?
Boxes here are stored as (cx, cy, w, h), with the center reported directly rather than a top-left corner.
. After thresholding the sliding-window score map, why does a single true face typically produce a *cluster* of detections rather than one?
Many nearby, overlapping windows all score above threshold near a real face, producing a cluster of near-duplicate detections.
. What does non-maximum suppression (NMS) do to resolve duplicate detections?
NMS greedily keeps the best-scoring box in each cluster and suppresses its highly-overlapping neighbors, collapsing duplicates down to one detection per object.