Face detection and bounding boxes, explained

Face detection answers "where are the faces?" — not "whose face is this?" (that is face recognition, a different and more sensitive task). A detector takes a frame and returns a list of bounding boxes, one per face, each with a confidence score. For anonymization that is exactly the right primitive: you need to know where every face is so you can cover it, and you specifically do not want to identify anyone.

A bounding box is the smallest axis-aligned rectangle that encloses the object — here, a face. It is stored as pixel coordinates, most often (x, y) for the top-left corner plus width and height, or as two corners (x1, y1, x2, y2). Some formats normalise the numbers to 0–1 so they survive resizing. The box is what a downstream stage acts on: to blur a face, you blur the pixels inside its box.

You keep detections above a chosen confidence and discard the rest. A low threshold maximises recall — you catch nearly every face, at the cost of occasionally boxing something that is not a face. A high threshold maximises precision — almost every box is a real face, but a few faint or distant faces slip through. For anonymization, recall matters more than precision: an extra blurred patch is harmless, a missed face is a leaked identity, so the threshold is set low and a human catches the false positives.

This is why anonymization pipelines tune for recall and keep a reviewer in the loop, rather than chasing a single "accuracy" number. The right threshold is the one that meets your acceptable miss rate on a labelled sample of your own footage.

A box marks a face; it does not hide it. After detection comes tracking (linking boxes across video frames so coverage does not flicker), redaction (blurring the pixels inside each box, irreversibly), and review (a human confirming nothing was missed). Detection is necessary but not sufficient — the guarantee that a video is anonymized comes from the whole pipeline, not the detector alone.