Public datasets are too clean
Most detection benchmarks are built from photographs. They were framed by a person, exposed well, and compressed once. A warehouse camera sees none of that. It looks down from four metres through a wide lens, under sodium lights, with rolling shutter skew on anything that moves and noise that climbs every time the dock door closes.
A model trained only on photographs learns the photographer, not the scene. It does well on the leaderboard and then loses a third of its recall on the first night shift.
Modelling the module
We characterise every sensor module we support. For each one we measure read noise, the response curve, lens distortion, rolling shutter timing and how the image processor sharpens and denoises. Those measurements become a sensor model that is applied to training images before the network ever sees them.
- Noise is added in the raw domain, before demosaicing, the way the sensor produces it.
- Lens distortion and vignetting match the real optics, not a generic fisheye.
- Motion blur and shutter skew are rendered from real object speeds on site.
The result is a training set that looks the way the deployed camera looks. The network spends its capacity on the scene instead of on undoing the sensor.
Calibrate once, ship everywhere
Customers do not train anything. When a camera is added in Parallax Studio, a two-minute calibration capture identifies the module and selects the matching variant of the model. If the module is new to us, the capture is enough to fit a sensor model, and a tuned variant follows without new labels.
The best augmentation is the one the sensor would have applied anyway.
What changed in the field
Across 14 customer sites, sensor-matched Depth-S recovered most of the gap between daytime and night recall. Night recall on people rose from 0.71 to 0.89 at the same false positive rate, with no change to model size or latency. On the COCO validation set the two variants score within half a point of each other, which is the point. The leaderboard did not move. The loading dock did.