Density must adapt to environment complexity
Aliases: view management · visual clutter · feature congestion
What it is
The same set of labels is crisp on a blank wall and a smear on a bench stacked with parts, where it also covers the parts’ contours. Adaptive overlay density takes the environment’s visual complexity as input: when the background is simple, labels may pack; when the background is already full of edges and texture, labels must thin, shrink, or appear only near the gaze. Density is not a product constant.
Adaptation runs both ways. It is not only subtraction when “it already looks full.” Empty scenes can have density added back.
Why it happens
Whether an overlay can be resolved is limited by feature congestion in the background. Rosenholtz’s clutter metrics use local variation in colour, orientation, and contrast to predict slower search; AR labels are another layer on that already congested picture. A blank wall has low congestion, so extra labels still separate; cabinets, cables, print raise local variation, and the same count fights both itself and the background for edges. View management (Bell, Feiner, Höllerer and after) moves and aggregates labels by projected position and overlap — density as a function of background geometry.
Complexity also changes with heading: a turn from a white wall to a window-side desk jumps congestion in one frame. Fixed density is then too sparse in one bin and too dense in the other. Adaptation that keeps up needs a computable complexity signal (edge density, feature congestion, depth layers) mapped onto a label cap, type size, or opacity.
Studying it
Prepare two rooms of different complexity (or two walls of one room), place the same candidate labels, and compare fixed density with density driven by a clutter metric. Dependent measures: label readability, whether real background contours can still be traced, time to find a named label. Evaluations of AR label layout by Tatzgern, Grasset and others run this “background changes, layout changes” route.
Independent variables: background feature congestion, the control law (fixed / congestion cap / gaze window), number of candidate labels. Dependent variables: recognition accuracy, preservation of background edges, search time.
A rendered fake background underestimates the congestion of real texture. Live rooms or photographs are what make complexity measurable.
Where it stops holding
In an environment that stays simple (one white wall in an empty gallery), adaptation barely fires and a fixed density is enough. If people have a density slider and use it, the algorithm yields — but a “follow the background” setting should still exist, because no one will twist a slider every time they turn. If the algorithm misreads complexity (an art wall of high contrast treated as “congested,” labels cleared), adaptation is worse than fixed. In a safety-critical mode the area cap still outranks adaptation — a simple background still cannot fill the central field.
Applying it
- Drive the current label cap from a signal you can compute live (edge density or feature congestion), not from one default count for the whole product.
- When complexity rises, aggregate first, then shrink type, then restrict to near the gaze. When it falls, put the withheld labels back, so the controller is not subtraction-only.
- How to check: open the same label set on a blank wall and on a parts-laden bench. The wall should be allowed more readable labels; the bench should still show part contours. The same density on both means there was no adaptation.