The short answer
Markerless multi-view motion capture measures a person’s 3D motion by triangulating keypoints seen by two or more ordinary cameras, with nothing worn on the body. Motion retargeting maps that motion onto a robot’s joints. Held-out camera error checks the result against a camera left out of the solve. Egocentric video is filmed from a camera the person wears.
- Markerless multi-view motion capture
- Measuring a person’s 3D motion by triangulating body and hand keypoints seen by two or more ordinary cameras at once, with no markers, suits or sensors on the body; synchronisation and camera positions can be recovered from the footage itself.
- Motion retargeting
- Mapping a person’s captured motion onto a robot with different proportions and joints, usually by inverse kinematics on matched body points, so a robot policy can learn to reproduce it.
- Held-out camera error
- An accuracy check for multi-camera capture: one camera is left out of the solve and the reconstruction is compared with what that camera saw, in pixels and in millimetres.
- Egocentric (first-person) video
- Video from a camera worn by the person, usually on the head; it sees the hands well and the wearer’s own body poorly. Exocentric video is filmed from outside the person.
- Bullet time
- A shot in which time slows or stops while the camera moves round the subject. Made with a camera array or by rebuilding the moment in 3D.
- Time-slice
- Bullet time made by firing an arc of cameras at the same instant and playing their frames in order.
- Free-viewpoint video (FVV)
- Video of a real scene that can be watched from camera positions that did not film it.
- Volumetric video
- A recording of a person or scene in 3D over time, usually from a stage of dozens to hundreds of cameras, playable from any angle.
- Volumetric capture stage
- An indoor studio ringed with synced cameras, often around 100, built for volumetric video.
- 4D replay
- A replay of a moment rebuilt in 3D over time, so the viewer can move the camera and scrub through it.
- 3D Gaussian splatting (3DGS)
- A way to represent a scene as millions of small coloured blobs that can be rendered from any view in real time.
- 4D Gaussian splatting (4DGS)
- Gaussian splatting extended over time, so the blobs move and change to represent a dynamic scene.
- Neural radiance field (NeRF)
- A neural network that maps a 3D position and view direction to colour and density, rendered by ray marching.
- Novel view synthesis
- Rendering an image of a scene from a viewpoint no camera occupied.
- Multi-view reconstruction
- Recovering 3D shape and motion from several cameras that saw the same thing.
- Markerless motion capture
- Measuring how a body moves from ordinary video, without suits, markers or sensors.
- Parametric body model
- A 3D human body controlled by a few shape and pose numbers, fitted to images to recover a person’s body and pose.
- 2D keypoints
- Joint positions (shoulders, knees, ankles) detected in each image by a pose model.
- Triangulation
- Finding a point’s 3D position from where it appears in two or more calibrated cameras.
- Camera calibration
- Working out each camera’s lens (intrinsics) and position and direction (extrinsics).
- Intrinsics
- A camera’s internal parameters: focal length, image centre and lens distortion.
- Extrinsics
- A camera’s position and orientation in the scene.
- Structure from motion (SfM)
- Solving camera positions and a sparse 3D point cloud from overlapping images.
- Bundle adjustment
- Refining cameras and 3D points together so that every point lands where it was seen in every image.
- Reprojection error
- The distance in pixels between where a 3D point projects into an image and where it was actually seen.
- Held-out view
- A camera left out of a reconstruction and used to test it: rebuild its view from the others and compare.
- PSNR
- Peak signal-to-noise ratio, in decibels: how close a rendered image is to the real one. Higher is closer.
- Genlock
- A shared signal that makes cameras capture frames at exactly the same moment.
- Timecode
- A clock stamp written into each frame so footage from several cameras can be lined up.
- Motion sync
- Lining up cameras with no shared clock by matching the subject’s movement as each camera saw it.
- Rolling shutter
- A sensor that reads the image line by line, so fast motion is skewed within one frame.
- Electronic image stabilisation (EIS)
- Software that crops and warps each frame to remove shake, which changes the camera’s geometry from frame to frame.
- Occlusion
- When something blocks a camera’s view of part of the subject.
- Virtual camera
- A camera path defined after the shoot and rendered from the reconstruction.
- Frozen orbit
- Bullet time in which time stops completely and the camera circles the subject.
- Multi-angle replay
- The same moment shown from several real cameras, synced to one timeline.
- FAA Part 107
- The US rules for commercial small-drone flight. Flying several drones at once needs a waiver, so each drone usually has its own pilot.
- Skeletal tracking
- Tracking each player’s joints over time from fixed venue cameras, as Hawk-Eye SkeleTRACK does with 29 points.