Point clouds are everywhere that measurement happens: astronomical surveys, medical imaging, laser scans of buildings. And every one of those fields shares a task — interactive selection, picking out a meaningful subset of points for closer analysis.
The data is inherently three-dimensional. The traditional workflow is a 2D mouse operating on a 2D projection, which inherits every limitation of a flat screen: occlusion, ambiguous depth, and the need to orbit the camera repeatedly just to establish where things are relative to each other.
A paper by Boaz Keren-Gil, James Gain, Lucia Marchetti, Patrick Marais and Alexander Sivitilli asks whether VR’s six-degree-of-freedom hand interaction produces measurable gains — and measures five things rather than the usual one.
The design
A counterbalanced within-subjects experiment, N = 34. Each participant selected 90 target galaxies from a catalogue of 16,816.
On the desktop: mouse-driven selection with keyboard camera control. In VR: a hand-held spherical brush that adds or removes particles as it passes through them.
The measures were task accuracy, completion time, and three standard instruments — the System Usability Scale (SUS), the NASA Task Load Index (NASA-TLX) and the Flow State Scale (FSS).
The results, including the one that didn’t go VR’s way
Accuracy was effectively identical: 91.6% in VR against 93.1% on the desktop, p = 0.858. No significant difference, and the point estimate marginally favours the mouse.
This is the result that makes the paper credible. The standard VR-beats-desktop study reports a win on every axis, which usually means the desktop condition was not implemented with much care. Here the mouse holds its own on the thing a mouse is good at — precision on a fixed projection.
Everything else favoured VR, and not marginally:
| Measure | VR | Desktop | Effect |
|---|---|---|---|
| Completion time | 23% faster | — | p = 0.001, d = 0.93 |
| Usability (SUS) | 84.6 | 54.0 | d = −1.78 |
| Workload (NASA-TLX) | 20.4 | 34.7 | d = 0.90 |
| Flow (FSS) | 83.2 | 70.2 | d = −1.23 |
The SUS gap is the one to sit with. 84.6 against 54.0 is not a preference, it is a different category of experience: a score in the mid-80s is conventionally read as excellent, while 54 falls below the commonly cited average of around 68. Participants did the same quality of work either way, and found one method genuinely pleasant and the other a slog.
The NASA-TLX figures say the same thing in the opposite direction: 20.4 against 34.7 in workload, with a large effect size. The accuracy was equal because people paid for it on the desktop.
Why the generalisation is reasonable
The authors are careful about scope but explicit about the implication: the underlying problem — making a 3D selection through a 2D interface — is not specific to astronomy. They expect the pattern to carry to point-cloud selection in other fields.
That is a defensible claim precisely because the task was abstract. Nothing about selecting 90 galaxies from 16,816 depends on them being galaxies; the difficulty is depth disambiguation in a dense cloud, which is identical in a LiDAR scan of a façade or a segmented CT volume.
The useful reading for interface designers
The temptation with a result like this is to conclude that VR wins and migrate the tool. The more useful conclusion is narrower.
Six degrees of freedom did not make people better at this task. It made the same performance cheaper — in time, in cognitive load, and in how willing someone is to keep doing it. For a working scientist who performs selection dozens of times a day, that is arguably the more valuable axis, and it is the one a pure accuracy benchmark would have missed entirely.