A panel intersecting a real object breaks depth judgment
Aliases: UI clipping through reality · virtual-real intersection · interpenetration
What it is
A floating settings panel cuts into a real table: the top half still in air, the bottom half sunk into the slab, the table edge coming out through the middle of the panel. That is not a slight misalignment. It is interpenetration. The visual system is told both “the panel is in front of the table” and “the table is in front of the panel,” and depth judgment collapses on the spot — people stop asking where the button is and start asking whether the table is fake.
Why it happens
Depth is a vote among cues. Stereo disparity parks the panel at, say, a metre and a half; occlusion, following the geometric intersection, draws the table edge in front of or behind panel pixels; motion parallax then walks the intersection line across the tabletop as the head moves. The three cues no longer support one 3D explanation, so percepts jump between “the panel is a ghost” and “the table has been sliced.”
Interpenetration is worse than “a bit off” because it draws an impossible contour: the real object has no slot for a virtual rectangle to enter, and vision has no tag that says “these are two coordinate systems.” Once that contour appears, the panel’s depth credit is vetoed, and neighbouring widgets that have not yet intersected are pulled into doubt.
Studying it
In a furnished room, place a world-locked panel so that it does or does not intersect a table edge, a monitor, a door frame. Ask people to judge “panel in front or behind the furniture,” and which object looks wrong. A pointing probe helps: ray-point the intersection line and see whether they land at table depth or panel depth.
Independent variables: intersection or not, intersection depth (graze / deep cut), texture complexity of the real object. Dependent variables: front/behind accuracy, localisation error of the intersection line, count of spoken presence breaks.
An empty lab has nothing to cut through, so interpenetration never happens and the conclusion inflates. Video see-through and optical see-through look different at an intersection; do not treat them as one stimulus.
Where it stops holding
In a fully virtual scene the same geometry is two virtual meshes intersecting, which breaks internal consistency rather than a real–virtual depth vote — similar mechanism, different social cost. At distances where stereo disparity is already weak, the intersection line is less sharp and may be read as a soft overprint. A file that is meant to sit in a desktop, with the intersection dressed by contact shadow, will be explained as a story rather than as an error; that is managed interpenetration, not harmless interpenetration. In the dark, when real contours are invisible, the break is postponed until a hand meets the real table.
Applying it
- Before placing a panel, test whether its bounding box cuts a table, a screen, a door frame. If it does, translate or shrink it; do not rely on the wearer to cock their head and “look past.”
- Treat an intersection line in-session as a depth incident: move the panel, do not hide the line with an outline. An outline hides the look, not the vote.
- Walk the real room. Do not only place interfaces in an empty scene.
- How to check: first-person video. Any frame where a real edge crosses through panel pixels is one interpenetration. If a walkthrough contains such a frame, the placement has not passed.