3D Environments Require 3D Visualisations

1Problem

We assess 3D spatial knowledge with a 2D medium.

Three examples of 2D sketch maps from prior work showing the difficulty of representing vertical structure on flat paper.

Three 2D sketch maps from Zhong & Kozhevnikov (2016). Each loses vertical information in a different way: floors drawn as separate plans cannot be aligned (A); annotations stand in for height (B); ad-hoc perspective distorts angles (C).

2Method

Two VR experiments. Each participant produced a 2D pen-and-paper and a 3D Gravity Sketch map of the same environment.

Experiment 1 environment: a three-storey building with the middle floor rotated 90 degrees, viewed in cutaway.

Exp 1 · Vertically layered building

n: 27
Levels: 3 (middle rotated 90°)
Landmarks: 6 (3 in, 3 out)
Locomotion: Joystick (free)

▶ Video demonstration of the experimental procedure

Experiment 2 environment: an urban scene of six buildings with landmarks placed at three different heights.

Exp 2 · Vertically volumetric urban scene

n: 37
Buildings: 6
Landmark heights: 3 levels
Locomotion: Drone fly-through

▶ Video demonstration of the experimental procedure

Each sketch was coded for occurrence and correctness of qualitative spatial relations along X, Y, Z axes separately. Bayesian Bernoulli mixed-effects models with random effects for participant and relation.

3Results

2D sketch maps act as a bottleneck: people store vertical spatial knowledge that the 2D medium hides but a 3D medium reveals.

H1 · supported

When drawing on paper, people show a bias toward the horizontal plane: they leave out vertical (up–down) relations far more often than horizontal ones.

Exp 1 · building

2.0× lower odds

OR 0.49, 95% CI [0.22, 1.10] · posterior 93%, moderate.

Exp 2 · urban

35× lower odds

OR 0.03, 95% CI [0.02, 0.05] · posterior ~100%, strong.

H2 · not supported

When a vertical relation is drawn in both formats, the 3D sketch is no more accurate than the 2D one. 3D doesn't make relations more correct — just more likely to appear.

Exp 1 · building

no difference

log-OR −0.32, 95% CI [−1.17, 0.52] · ~3× more likely under null.

Exp 2 · urban

no difference

log-OR −0.31, 95% CI [−0.76, 0.12] · ~7× more likely under null.

H3 · supported

When a relation is missing from someone's 2D sketch but appears on their 3D sketch, it's mostly correct — the spatial knowledge was there; the 2D medium was hiding it.

Exp 1 · building

91% correct

OR vs chance 10.5, 95% CI [6.5, 17.5] · posterior ~100%.

Exp 2 · urban

67% correct

OR vs chance 1.88, 95% CI [1.47, 2.41] · posterior ~100%.

4Sketch maps

Every participant’s 2D and 3D sketch, side by side. Click a card.