Embodied AI Glossary中文

Field of View

视场角FoVCommon

The angular range a camera or sensor can see at once, usually given as horizontal, vertical, and diagonal figures.

Field of view is the angular range a camera or sensor covers at any given instant, usually reported separately as horizontal, vertical, and diagonal, in degrees. For an ordinary pinhole camera, it's set by the sensor size s and the focal length f: FoV = 2·arctan(s/2f) — a shorter focal length gives a wider field. Lidar and depth-camera spec sheets likewise list horizontal and vertical FoV separately. In embodied AI, field of view determines how much a camera can actually see: a wrist camera sits close to the object, so too narrow a FoV loses the surrounding context, which is why UMI added a 155° fisheye lens to its wrist-mounted GoPro. Conversely, a wider FoV means more distortion at the edges and fewer pixels devoted to each degree of view. Swapping a camera or lens changes the distribution of the images a policy sees, so an already-trained vision policy often needs extra data to adapt.

ExampleA camera with a 36 mm-wide sensor and an 18 mm focal length has a horizontal field of view of 2·arctan(36/36) = 90°; switching to a 36 mm focal length shrinks that to about 53°.

Also called
FoV, Angle of View
Related
Pinhole Camera Model · Camera Intrinsics · Fisheye Camera · Lens Distortion · Wrist Camera · Depth Camera
Sources
Field of view - Wikipedia
Universal Manipulation Interface (arXiv:2402.10329)

See it in the full glossary →