Egocentric vs Third-Person Video Data

Egocentric (first-person) video is recorded from a camera worn by the person doing the task — on the head, chest, or wrist — so the camera shares the actor’s own viewpoint and moves with their body. Third-person video is recorded from a camera placed away from the actor, observing the scene from the outside, the way a bystander or a fixed security camera would.

The distinction is about where the camera is relative to the actor, not about resolution, frame rate or any other production quality. A phone in a chest mount and a glasses-style camera are both egocentric because they move with the wearer; a camera bolted to a wall or a tripod across the room is third-person because it is not.

Egocentric video puts the actor’s hands, tools and the objects they are manipulating close to the camera, entering frame the way the actor actually reaches for them — including occlusion during grasp and motion blur during reach, which are part of the training distribution a manipulation policy has to handle, not noise to remove.

It also carries gaze and head-motion information a fixed external camera cannot: where the actor was looking and how their head moved relative to the task. On this site’s multi-camera tier, that head view is paired with synchronized wrist cameras, depth and IMU on one hardware-triggered clock, so hand position, contact and the head view are all aligned to the same timeline rather than recorded in isolation. See /glossary/egocentric-video.

A third-person or fixed external camera shows the actor’s whole body and the surrounding scene in a single frame — posture, position in the room, and spatial relationships between the actor and other people or objects that a worn camera, by definition, cannot show because it is attached to the person it would need to frame.

This site describes third-person video here only as a general category for comparison; it is not a capture mode Firsthand offers. If an existing page on this site states otherwise, that statement should be treated as authoritative over this general description.

A robot with a head- or chest-mounted camera, or an embodied-AI system reasoning over a first-person observation stream, sees the world the way an egocentric capture records it — hands and objects entering frame from the body’s own position, not from an external vantage point. Training on footage that matches that observation distribution is the reason egocentric capture is the default viewpoint for manipulation and embodied-AI training data on this site. See /custom-collection/egocentric-video-data-collection.

A third-person view only matches what a robot sees if the robot itself is the subject being observed externally — for example, reviewing a teleoperated robot’s behavior from a fixed camera in the workspace, rather than training the robot’s own onboard perception.

On this site’s egocentric capture, labels such as 3D hand pose, depth-aligned segmentation and action segments are annotated against the same hardware clock and calibration the video was recorded on, because the head and wrist cameras, depth and IMU all share one PTP clock domain. That synchronized geometry is what makes metric 3D hand pose and contact events possible rather than estimated after the fact.

Generic third-person video, as a category, does not carry that same worn-camera proximity to the hands, so label types that depend on close, synchronized hand-object geometry are harder to produce from it; whole-body pose or scene-level annotation is a better match for what a third-person frame actually contains. See /glossary/hand-pose-estimation.

In Firsthand’s egocentric capture, every contributor signs a documented release as part of the engagement, and faces, plates and screens appearing in frame — including bystanders who are not the task performer — are detected, tracked and irreversibly blurred before anything leaves QA, with a human reviewer confirming the result. See /anonymization and /provenance.

A fixed or external third-person camera, as a general category, is more likely to capture multiple bystanders across a wider field of view than a camera worn by one task performer and pointed at their own task, which is one reason consent and anonymization requirements can be harder to satisfy for third-person capture in a shared space. This site does not offer third-person capture, so it cannot speak to how any particular third-person vendor handles that chain.

A program training a manipulation or embodied-AI policy needs egocentric data as the core training signal, because that is the viewpoint the deployed system will actually occupy. Third-person footage, where a team sources it separately, can be a useful secondary signal for context that a worn camera structurally cannot provide — whole-body posture, scene layout, or reviewing a teleoperated robot’s own behavior from outside it.

Where the gap is egocentric footage of a specific skill, environment or population that does not exist yet, that is what a custom collection engagement is built to close. See /custom-collection/pretraining-data-collection and /guides/robot-pretraining-data.

A custom egocentric collection engagement starts with a written skill spec — the task, objects, environment, camera tier and acceptance test a batch has to clear — runs a small Spec Pilot batch to prove the spec and QA bar hold on real footage, then scales capture across phone, head-rig or multi-camera tiers as needed, with delivery under a buyer-owned license. See /custom-collection/egocentric-video-data-collection for the tiers and /custom-collection/pricing for cost drivers.