redactcam

Confidence cannot tell a plate from a sign

A plate detector fires hard on rectangular signage. A place-name sign scores about 0.6 — above any confidence threshold you can actually use without losing real plates. Raise the number and you start missing plates; lower it and you blur every sign on the road. There is no setting on that dial that separates the two, which is why the dial is the wrong control.

The axes that do separate them are geometric, and the figure above is them. A real plate is wide but bounded — around 2.4–3:1 seen at an angle, up to about 4.7:1 head-on — while signs box near-square and shorelines and bridge railings box at about 13:1. And a real plate is a tiny fraction of a 4K frame, never more than about 0.4%, while a stone wall the model misreads covers several percent. Either axis alone lets something through: the wall sits squarely inside the aspect window, and only the area ceiling drops it. Both together leave a window that keeps every real plate and rejects all three of the others.

The same problem turns up on faces, where the separable axis is time rather than shape. Foliage and shadows score 0.5–0.7, comfortably above a genuine small face at 0.19, so no confidence threshold works there either. But a real face re-detects across consecutive samples while a texture artefact fires once and stops. That is why the face confidence is set very low and paired with a rule that a track must accumulate several detections before it is emitted — lowering the confidence without that pairing gives you a video full of blurred hedges.

The blur is tracked, not stamped

Detection is expensive, so it runs on a sample of frames. The naive fix is to hold each detection’s box until the next one, or to union the two endpoints. Either way a plate caught in one sample has its box frozen for a fraction of a second while the car keeps moving: the blur covers empty road ahead of the plate, and the plate itself drives out from under it in plain view.

redactcam propagates each box with Lucas-Kanade optical flow on every frame, with a RANSAC scale term so an approaching object’s box grows with it, and snaps the box back to ground truth at each new detection. Detection stays sparse; propagation costs milliseconds. Occupants get a different treatment again — face detection through a moving windscreen latches onto reflections and roof lines at any confidence, so the cabin is covered geometrically instead: detect the whole vehicle, blur the top band of its box, which is where people sit at every viewing angle.

Verifying after the render is verifying too late

Once a multi-gigabyte file is encoded and shipped, discovering that the blur missed a driver costs a re-encode and a takedown. So the check runs before the encode, and it does not trust any stage above it. It re-runs detection from scratch on the source frames, corroborates each vehicle with a licence plate — so it chases cars rather than billboards — and confirms that the mask covers each corroborated car’s cabin. A cabin the mask leaves uncovered raises CoverageError instead of producing a file.

Checking the mask rather than the finished render is deliberate. A render usually carries other changes too — grading, overlays, a different codec — so differencing it against the source cannot isolate the blur. The mask is exactly what the blur will cover, which makes mask coverage equal to output coverage without the confound, and it exists before the expensive, irreversible step rather than after it.

What this page is and is not

The gate thresholds in the figure are read straight from the library’s tuned preset for forward-facing 4K road footage. The rectangles are the ranges that preset’s own comments record from real footage, drawn to show the decision the gate makes — they are per-class ranges rather than detections from any one video, and no output here is presented as a result. There is no blurred video on this page on purpose: the interesting claim is not that a detector found a face, it is what happens when one does not.

Requires Python 3.11 or newer and an ffmpeg/ffprobe binary on the path. The code is MIT, but the model weights it downloads by default are not, and several are copyleft — the README sets that out in full before you deploy anything. pip install redactcam. Source: github.com/Back-Road-Creative/redactcam, published on PyPI. The longer argument is the essay, Prove the Blur Landed.