Egocentric video data is first-person footage from a camera on the head, chest or a pair of glasses, showing hands and tasks the way the wearer sees them. People from our 30,000+ network film it to your brief on mounted phones, action cameras, smart glasses or your own devices, and the same team labels the hands, objects and steps.
Egocentric video data is footage recorded by a camera worn by the person doing the activity, usually on the head, chest or a pair of glasses. It shows the world from the wearer's point of view: their hands, the objects they touch and each step of the task, in the order they do it.
The word comes from "ego", the self. Footage from a separate camera watching the person is called exocentric, or third-person, video. Some projects record both at once, so a model can learn to link the two views.
First-person video is in demand because it's how robots and wearable AI see. Teams building robot manipulation, smart-glasses assistants and AR task guidance need footage of real people doing real tasks, with their hands clearly in frame.
Public datasets such as Ego4D and EPIC-KITCHENS turned first-person video into a research field. They were built for research, though, and their licenses set terms on commercial use. A product team usually needs footage of its own tasks, in its own settings, on its own devices, filmed to a brief. Our guide to planning an egocentric collection walks through that brief.
The device decides how wide the view is, how steady the picture is and what else gets recorded with it. We match it to what your model will see once it ships.
A contributor's own phone in a head strap or chest harness, recording through our Video Collection tool. It's the easiest setup to spread across many people, homes and regions.
For large, varied collections
Our tool: Video Collection Free sample of phone-mounted footage
Wide-angle cameras worn on the head or chest. The wide view keeps both hands and the work surface in frame during close, hands-busy tasks.
For manipulation and close-up tasks
Camera glasses that film from eye level, close to where the wearer is looking, with both hands free. The footage matches what a glasses-based assistant will see.
For wearable assistants and AR
If your model is tied to a specific camera, headset or sensor set, send us the hardware and contributors film with it to your protocol.
For matching a production device exactly
New footage filmed to your brief, and labels that follow the wearer's hands and task through time. The computer vision page covers the image labels in depth.
A box or an outline around each hand, with left and right kept apart, drawn to rules agreed for hands that leave the frame or hide behind an object.
For hand detection and tracking
Our tool: Image Annotation Free sample of hand labels
Boxes on the objects being reached for, held or used, with their class, and the frames where a hand touches them and lets go.
For hand-object interaction and grasping
Our tool: Image Annotation Free sample of object labels
The same hand or object followed across the clip with one id, including while the camera turns with the wearer's head.
For tracking and movement
Our tool: Image Annotation Free sample of tracking
The start and end of each action or step, like open the drawer or tighten the bolt, from a label set you agree on. Long tasks get split into the steps a model has to recognize.
For action and procedure recognition
If contributors talk through what they're doing as they film, that speech gets transcribed with timestamps, so each step comes with a description in the wearer's own words.
For video-language models and assistants
Our tool: Transcription Free sample of transcribed speech
Footage spread across the people, rooms, regions and light your model will meet, planned into the brief before anyone films.
For models that work outside the lab
A handful of common uses. They show the kind of work, and they aren't client names. For footage from a fixed or handheld camera, see video data.
People doing household and workplace tasks, with hands and steps labeled, for robot models that learn how a job gets done by watching it done.
Close-up footage of hands picking up, turning and using objects, so a model learns grips and contact from many different hands.
Eye-level footage of everyday life, for assistants that answer questions about what the wearer is looking at.
Assembly, picking and repair filmed step by step, for AR guidance that knows which step a worker is on and what comes next.
Aisles, shelves and products filmed the way a shopper sees them, for models that recognize products in hand and on the shelf.
Held-back clips from every device, place and condition, to check how your model does in each before it ships.
Each page covers what that kind of first-person project usually needs, and each one is built to your brief. Every other category is on the datasets hub.
Reshooting first-person footage means sending people back to redo the task. Settling these four things first is the cheap fix. The egocentric collection guide has a brief template that covers them all.
A fixed list of steps gives clean, comparable clips. Letting people do the task their own way gives the variety a model meets in real homes and workplaces. Many briefs mix the two.
Head, chest and glasses each frame the hands differently. Match the device to your product, and agree the resolution, frame rate and any motion data before collection.
The mix of contributors, how much footage any one person can give, and the homes, sites and regions involved, so one kitchen doesn't end up teaching the whole model.
Where filming is allowed, how people who haven't agreed to be filmed are kept out, and what gets removed before delivery. Each contributor's consent is recorded before collection starts.
Your named contact runs it from the brief to the last clip. Each step links to the service behind it.
You tell us the tasks, the settings, the device and what the data is for.
The NDA is signed before any data moves.
A few contributors film a small sample free, so you can check framing and quality.
People from our network film the tasks first-person, on the agreed device.
Hands, objects and steps get labeled, and speech gets transcribed.
Footage and labels are reviewed before they're delivered to you.
What people mean by the words that come up on a first-person video project. The full glossary covers every data type.
A first-person video project usually needs more than one of these. You can also compare all eight.
New first-person footage filmed by people in our network.
Hands, objects and tracks labeled frame by frame.
What the wearer says, turned into timed text.
Held-back first-person clips scored against your targets.
Fixed, handheld and driving footage, labeled the same way.
Straight answers about first-person video. The services FAQ covers how the services fit together.
First-person footage recorded by a camera worn by the person doing the activity, usually on the head, chest or a pair of glasses, labeled so a model can learn from it. It shows the wearer's hands, the objects they use and each step of the task.
Egocentric video is filmed from the wearer's point of view. Exocentric, or third-person, video is filmed by a separate camera watching the person. Some projects record both at the same time so a model can link the two views.
Yes. We've delivered first-person video projects for clients before, as part of 150+ projects since 2021. We don't publish client names.
Phones on head or chest mounts, action cameras worn on the head or chest, smart glasses, or devices you send us. We pick the setup with you, based on what your model will see once it ships.
Yes. If your model depends on a specific device or sensor set, send us the hardware and contributors film with it to your protocol.
They're a good place to start research, but they were built for research and their licenses set terms on commercial use. They also cover the tasks, places and devices their creators chose, not yours. Read the license first, and film the gaps to your own brief.
Hands and the objects they use in each frame, the same hand or object followed with one id, actions and task steps with their start and end, and what the wearer says, transcribed with timestamps.
It's settled in the brief before anyone films: where filming is allowed, how people who haven't agreed to be filmed are kept out of the footage, and what gets removed before delivery. Each contributor's consent is recorded before collection starts, and the NDA is signed before any data moves.
Yes. We take calls in US Eastern mornings, which is evening in India. We're ISO 27001 certified, our SOC 2 report is available under NDA, and we sign GDPR data processing agreements and HIPAA business associate agreements on request.
Yes. Send a short brief and we'll film or label a small sample free before you commit.
Pick what you need, say which tasks get filmed and roughly how much footage, and send it. We'll film or label a small sample free before you commit.
This opens your email app with the brief filled in. Nothing is sent until you press send. You can also use the contact page.