Three degrees of freedom track rotation only
Aliases: 3DoF · orientation-only tracking · rotational DoF · yaw-pitch-roll tracking
What it is
Sit in a swivel chair with a cardboard viewer: turn the head and the panorama turns; lean toward a vase and it does not grow. Three degrees of freedom (3DoF) track head rotation only — yaw, pitch, and roll. Translation — fore-aft, left-right, up-down — never enters the pose. The viewpoint is pinned to an invisible sphere centre. The wearer can look around and cannot walk up.
This is not a crippled six-degree system. It is a complete pose model: it declares that the world responds to orientation alone. 360 video, seated viewing, and early phone VR all run on it.
Why it happens
Head pose has six independent components. The three rotational ones can be integrated from an inertial measurement unit with no external reference: gyros give angular velocity, accelerometers help estimate gravity and so pin pitch and roll. The three translational ones have no equally cheap sensor path — double-integrating acceleration drifts within seconds — so a 3DoF device chooses not to measure them. At render time the camera has an orientation matrix and no position update; the head moved ten centimetres in the room, the virtual camera stayed at the origin.
The perceptual cost is missing motion parallax. In a real room, leaning toward a table makes near structure shear faster than far structure, and the table edge grows into the field of view. Under 3DoF that optic flow never arrives, however far the wearer leans. Rotation is still correct, so “turn to look” works and “lean to look” does not. Seated, with the head roughly in place, the missing segment is rarely demanded, and the model is enough.
Studying it
Use a controlled lean task: a target an arm’s length in front of the viewpoint, head turns allowed, standing forbidden. Compare rotation-only pose updates with pose updates that also include translation, on judgments of relative distance and occlusion.
Independent variables: whether pose updates include translation, lean amplitude, target distance. Dependent variables: distance-estimation error, whether a partly occluded object can be named, discomfort during the lean, whether a post-trial sketch of relative layout is correct.
Field observation is simpler: in seated 360 content, count how many people instinctively lean toward a point of interest, and whether they immediately sit back — sitting back usually means a translation channel was expected and not delivered.
Where it stops holding
A swivel chair plus 3DoF is enough for looking: lectures, panoramic documentary, seated cinema. Reaching, looking under a table, sidestepping an obstacle are not resolution problems; they are missing position. In a car or cockpit the body is belted and translation is already small, so 3DoF matches the task. Children and tall adults seated at the same origin do not get different eye heights, because position is not tracked — that is the model, not a bug. Strap a 3DoF headset to a walking person and rotation remains correct; the first step produces a different error, which is no longer “rotation only” but a mismatch between channel and locomotion.
Applying it
- Design 3DoF content for sitting, looking around, and selecting. Do not design “walk up to that table for a closer look.” Put points of interest on a sphere centred on the head; acquire them by pointing with orientation, not by translating closer.
- Pin the model with one action line at the start: “Turn your head to look; there is no need to lean.” Cheaper than explaining after a failed lean.
- Size interactive hotspots in angle, not in world centimetres. Put buttons on the viewing direction; do not put them where a forward lean is required to reach.
- How to check: have someone sit through the whole piece, then ask them to point at “the one thing you wanted to lean toward.” Every object named is content demanding a translation channel the format did not promise.