A good object detector doesn’t hand you a yes or a no. Point one at a traffic camera and for every box it draws it also prints a number — car 0.62, car 0.94, person 0.31 — a confidence, its own estimate of how sure it is. That number feels like a gift. It looks like the detector doing the hard part for you, sorting the certain from the shaky so you don’t have to. All that’s left, surely, is to draw a line: keep everything above some score, drop the rest. Pick the line and you’re done.

Thyseus is a live map built from public cameras, and that line — the confidence threshold — is the single most consequential number in the whole system. It’s also the one number that cannot be set correctly, and learning why is most of what it means to build a map that doesn’t lie.

One knob, two ways to be wrong

The threshold is a knob, and a knob only moves one way at a time. Set it low — keep everything scored above, say, 0.30 — and you are generous: almost every real car makes the cut. You also keep a crowd of things that were never there. A wet reflection the detector briefly reads as a hood. A shadow under an overpass. A billboard car. A parked truck’s wing mirror catching the sun. These are ghosts: confident-looking boxes over nothing, and at a low threshold they pour onto the map as pins that a user will drive toward and find empty.

So you raise the knob. Set it to 0.88 and the ghosts starve — almost nothing spurious scores that high. But the same move that kills the ghosts also kills cars. The genuine car far down the road, half-occluded by a bus, seen at a bad angle in the rain, scored 0.71 — gone. It was real. It was on the road. And your clean, ghost-free map now quietly omits it, which is its own kind of lie: not a thing that isn’t there, but a thing that is and goes unshown.

Two failure modes, and the knob trades one for the other. Down: ghosts on the map — false things shown. Up: cars missed — true things hidden. Every turn of the dial buys down one cost by paying up the other.

The overlap is the whole problem

You might hope this is just a tuning exercise — that somewhere in the middle sits a magic setting where the ghosts have all fallen away and the real cars are all still standing. There isn’t one, and the reason is a fact about the scores themselves, not about your patience.

Confidence is not truth. It’s the detector’s feeling about a patch of pixels, and that feeling is only loosely coupled to whether the patch is a car. Real cars, seen badly — distant, dim, clipped by the frame, streaked by rain — score low. Ghosts, seen under just the right accident of light, score high. So the two populations don’t sit in tidy, separated piles. Their tails overlap. There is a stretch of the scale — call it 0.45 to 0.70 — where a faint-but-real car and a bold-but-fake ghost earn the exact same number. Inside that band, the score tells you nothing. A cut placed anywhere in it leaves ghosts above and cars below, every time.

A threshold can only separate what the scores already separate. Where a real car and a ghost earn the same confidence, no line runs between them — because on the evidence the line is given, there is nothing there to divide.

That overlap sets a floor. In the figure above, whatever you do with the dial, the best it reaches is a tie: one ghost still shown, two cars still missed. You can move that three around — trade the ghost for another dropped car, trade a miss back for another ghost — but you cannot get below it by turning the knob, because the knob was never the thing standing between you and a clean answer. The overlap was. The dial doesn’t remove the mistake. It only decides its shape.

So stop trying to win it with the dial

Here’s the turn, and it’s the part that makes Thyseus a map and not a demo. Once you accept that no single threshold is honest, the question changes. It stops being what’s the right line? and becomes why are we asking one camera’s one glance to be right in the first place?

A threshold is a decision made with a single observation: this box, this score, keep-or-drop. But a ghost and a car differ in a way a single frame can’t see — they differ across observations. A reflection that reads as a hood is a one-off: it’s gone the next frame, it doesn’t move like a car, and no other tab watching a nearby camera ever agrees that something is there. A real car persists, travels a plausible path, and — where coverage overlaps — gets seen by more than one feed at once. The signal that actually separates the two isn’t the confidence score. It’s agreement over time and across cameras.

That’s exactly the discriminator a threshold throws away. So Thyseus sets the per-tab dial generously — deliberately toward the ghost-heavy end, keeping the marginal detections a strict map would discard — and then refuses to trust any of them on their own. A pin earns its place on the map only when independent observations agree: the same spot, from a couple of unrelated tabs, holding across a few seconds. Ghosts don’t survive that test — they have no corroboration to offer. Faint real cars often do, because being faint in one frame doesn’t stop you from being solid in the next, or obvious to the camera down the block. The overlap that no line could split gets split by consensus instead.

It isn’t free. Corroboration takes a beat, so a genuinely fleeting car — one camera, one second, never seen again — can still slip through, and we’d rather lose that than invent a ghost. Honest to the meter means saying so. But the trade is a far better one than the dial alone could ever offer, because it’s made with the evidence a threshold never had.

The number was a tell, not a verdict

The confidence score isn’t useless — it’s a genuinely good hint, and throwing it away entirely would be its own mistake. But it is a tell, not a verdict, and the whole error of the naive approach is treating one camera’s hint as the final word. The detector’s number answers how sure am I, from this one look? The map has to answer something harder — is it really there? — and that question was never going to be settled by one glance, however confidently it was scored.

So we stopped hunting for the honest threshold, because there isn’t one. The dial only moves the mistake around; it takes more than one pair of eyes to make it disappear.