To a person standing on the corner, a box truck rolling through the intersection is just a bigger vehicle — one more thing in the scene, a little slower, a little louder. To a camera watching that same corner, it is something else entirely. It is a wall. For the two or three seconds it takes to cross the frame, the truck is a slab of sheet metal moving between the lens and everything in the lane behind it, and everything in that lane stops reaching the camera. Three cars that were plainly there a moment ago are, as far as any pixel is concerned, gone. Not moved. Not parked. Erased.
The detector doesn’t fail when this happens, and that’s the part worth sitting with. A computer-vision model spots a car by the light that lands on the sensor, and no light is coming back from behind the truck, so the model does exactly the right thing and reports nothing there. It is being honest. The feed showed it a wall, and it drew a box around the wall and told the truth about the rest of the street: nothing here that I can see. The trouble is the last three words, which the map is about to drop.
The visible is a biased sample
If occlusion sprinkled its errors evenly across the day, you could shrug it off as noise. It doesn’t, and the shape is what makes it dangerous. Big occluders — trucks, buses, delivery vans — cluster in dense traffic, at the busy hours, on the busy corners. The empty residential feed at two in the afternoon has nothing to hide behind and counts honestly. The six-lane downtown junction at rush hour is wall after moving wall, each one blanking a slice of the scene for a beat. So a single camera undercounts precisely when the count is largest, and reports its cleanest numbers exactly where nobody needed them.
That should sound familiar, because it is the double-counting problem wearing its coat inside out. Overlapping cameras invent traffic where a corner is busy, plotting one truck as two. Occlusion erases traffic where a corner is busy, dropping three cars to zero. They look like opposite bugs, and they are — but both of them aim their error at the same target. A live map of a city is, in the end, a machine for answering where is it busy?, and here are two failures that both lie loudest in the busy places: one hands back the crowd exaggerated, the other hands it back thinned. Neither adds a little static everywhere. Both distort the map exactly where its whole purpose lives.
A detector’s silence has two meanings that look identical on the wire: there is nothing here, and I could not see here. Absence of a detection is not a detection of absence. The camera can’t tell those apart — but the map must, because it’s about to draw one of them as a fact.
Unknown is not empty
The tempting move, the one a demo never punishes you for, is to treat “no detection” as “no car.” One camera, one clear afternoon, and the equation holds well enough. Ship it against a real intersection at five o’clock and it turns the back of every truck into a little zone of confident zeros — a stretch of street the map swears is clear, at the exact instant it is anything but. The map isn’t wrong the way a typo is wrong. It’s wrong the way a false all-clear is wrong: most convincing at the worst possible moment.
So the first discipline is a refusal, and it’s the same shape as the refusals that run through the rest of the studio’s work. Thyseus does not get to pin a zero it didn’t earn. A patch of road that a truck is standing in front of is not empty; it is unknown, and those are different marks on the map. The honest rendering of a blocked view is a question, drawn as a question — a region the map admits it can’t currently speak for — not a cleared street drawn with the same confidence as a genuinely quiet one. Getting that distinction into the data model, so a gap can carry “I couldn’t see” instead of silently collapsing to “there’s nothing,” is most of the work. The rest is knowing what to do with the gap once you’re honest that you have one.
The wall moves, and that’s the tell
What saves the count is that the wall is in a hurry. A truck occludes for a beat and then it’s gone, and the two facts on either side of that beat are enormously informative. A car sat in that lane a half-second before the truck arrived. A car sat in that lane a half-second after it cleared. Cars do not teleport, and they do not wink out of the universe to make room for freight. The track that stitches boxes across time — the same machinery that keeps a single sedan from being counted as thirty as it crosses the frame — reaches straight across the occlusion and says the obvious thing out loud: it was there the whole time; you just couldn’t see it. Continuity turns the gap from a hole into a bridge.
And where time isn’t enough, the geometry is. The second camera whose overlapping angle caused the double-counting is the very same camera that sees around the truck. The angle that’s a liability when you’re trying not to count a thing twice becomes an asset the moment a thing has been hidden: what one lens has walled off, the other very often has in plain view. Two cameras that have to be reconciled into one story don’t just correct each other’s phantoms; they cover each other’s blind spots. When they disagree — one seeing three cars where the other sees a truck and nothing — the disagreement isn’t a problem to average away. It’s the answer. One of them is occluded, and the map can tell which.
None of this makes the blind spot vanish. A street that no camera and no recent track can vouch for stays unknown, and it should — that’s the honest edge of the coverage, and pretending otherwise would just be the confident-zero mistake in a bigger coat. The whole point is to hold the line between I can see this street is clear and I cannot currently see this street, and to spend cameras and time buying back as much of the second category as the network honestly can.
That’s the same failure the studio designs hardest against everywhere it builds: not the loud error you catch in the demo, but the quiet, confident one that only shows up in the busy room the tool has already walked out of. A map that undercounts and tells you it’s unsure is a tool you can work with. A map that undercounts and draws the gap as a cleared road is a tool that will lie to you at rush hour and look composed doing it. The truck was always going to be a wall. The only question a mapmaker gets to decide is whether the wall casts a zero or a question mark — and one of those is a great deal easier to live with when the street fills back in.