Occlusion errors immediately destroy spatial credibility
Aliases: broken occlusion · depth-order error · spatial credibility
What it is
A virtual lamp that should sit behind a real post floats in front of it, and the room flips from “there is a thing in here” to “a sticker on the glass.” Occlusion failure does not bruise polish. It bruises spatial credibility: the strongest cue people use to decide “it belongs in the world” reverses on the spot, and the other depth cues never get to catch up. The error need not be large. One reversed order is enough.
Credibility here is a snap verdict, not the kind of discomfort that accumulates with wear.
Why it happens
Occlusion is the strongest near-field depth cue. In Cutting and Vishton’s ranking it outranks disparity, perspective, and relative size inside arm’s reach. The visual system’s default is “whatever hides the other thing is closer.” A virtual object that violates that says “this picture is not of this room” — a category judgment, not a matter of degree. Stereo, shadows, and contact, even when they are right, cannot outvote a reversed order; they get reread as detail on a decal.
Time does not offer an adaptation window. People do not “get used to interpenetration after a while.” The first error already drops the object from scene membership, and later correct frames rarely buy it back.
Studying it
Run a depth-cue conflict: the same real–virtual pair, occlusion order flipped, stereo, shadow, and perspective held constant. Collect a forced choice — “object in the room” versus “layer stuck to the view” — plus spatial-realism subscales of presence inventories. Score brief flashes separately from sustained errors, to see how little violation it takes to reverse the category.
Independent variables: occlusion order, error duration, object distance. Dependent variables: the forced choice, spatial subscores of IPQ or a similar inventory, time to notice.
The forced choice outperforms mean satisfaction. Content can inflate satisfaction; it does not inflate the category judgment.
Where it stops holding
Stylized, openly layer-like marks — navigation chevrons, screen-space HUDs — are expected to sit on the view, so occlusion errors are barely read as spatial events. Far landscape and skybox objects already have weak occlusion; a reversed order there is less fatal than on near furniture. An error shorter than a saccade and parked in the periphery can be missed; the same error on the object being handled almost never is. “Unacceptable” measured on a static desktop prop does not transfer cleanly to walking, where the background keeps changing — motion may add depth information, or it may make the error louder; the direction is not settled.
Applying it
- Treat a wrong occlusion order as a defect that ejects the object from the scene, not as a polish item. If the broken occlusion can be turned off, turn it off; do not ship a flash.
- Spend correctness on near objects the wearer is looking at. Let distant decoration yield.
- When depth is unstable, shrink the virtual object and leave it clearly in mid-air rather than letting it fight a real piece of furniture for front and back.
- How to check: show someone new to the project ten seconds of the live scene and ask only “is this thing in the room?” Every “stuck on the glass” is a spatial claim that already failed.