Exocentric Video

Video shot from a camera outside the performer's own body, looking at a person or robot from a bystander's vantage point.

Exocentric video is the counterpart to first-person (egocentric) video, where the camera is worn on the performer's head or chest. In exocentric video, the camera sits outside the performer's body — a tripod-mounted camera, a security camera, or handheld footage shot by someone nearby. Most instructional videos and sports broadcasts on the internet fall into this category, and there's far more exocentric footage available online than egocentric footage. It shows full-body posture and the overall relationship between a person and their environment clearly, but hand detail is often blocked by the body or by objects, and the viewpoint doesn't match what a robot's own onboard camera sees. In embodied AI, it's commonly used to estimate human body motion, which is then retargeted for a humanoid robot to imitate, or mined for task procedures; the third-person cameras fixed beside a table during robot data collection also produce this kind of footage. Datasets like Ego-Exo4D record both viewpoints at once specifically to study the correspondence between them.

ExampleGVHMR is used to estimate human body motion from a monocular exocentric video, and GMR then retargets that motion into joint trajectories for a humanoid robot.

Also called

Third-Person Video, Exo

Related

Egocentric Video · Third-Person Camera · Ego-Exo4D · Internet Video Data · Human Pose Estimation · Motion Retargeting

Sources

Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives (arXiv)
Ego-Exo4D 官网 (Chinese)