After launch
Something is now in the way. Here is what it does.
Vegetation, new plant or stored material sits between the camera and the asset, and part of the view is quietly gone.
From the floor
A tree grew. Pallets were stacked in a new place. A temporary structure went up and stayed. The camera still returns a good image of whatever it can see, and nothing reports that a portion of the scene has been unavailable for a month.
Which failure is it
Covariate shift
The visible input changed. Worth separating from other covariate cases because the fix is often physical rather than statistical: no amount of retraining recovers an asset that is genuinely no longer in shot.
Four modes, and they need different responses: covariate, concept, prior, and a pipeline event wearing drift’s coat.
What moves first
Accuracy is the last thing to tell you.
Detections fall for the assets in one region of the frame while the rest of the frame is unaffected. Spatial rather than global, which is what distinguishes it from lighting or weather.
Measuring accuracy directly needs fresh labels, so teams lean on proxies between audits. By the time a quarterly audit shows a drop, the system has been wrong for a quarter. The cheap signals are confidence per class, detection rate per camera against its own history, and how often a reviewer overrides the model.
What to do
Check the frame before you retrain. If the asset is occluded, that is a maintenance ticket, not a labelling job. Retraining around a blocked view teaches the model to stop expecting the thing.
What it costs to ignore
The most misdiagnosed condition here. Teams retrain repeatedly against a view problem and conclude the model cannot do the task.
Where this shows up
What it gets first