Video data for AI is footage your model learns from over time: objects that move, actions that unfold, and speech that goes with the picture. People from our 30,000+ network film it in our Video Collection tool, and the same team labels, tracks and transcribes it.
clip 031 · captured
car 03 · tracked
events over time
walking
crossing
stopped
00:00 00:05 00:09
labeledVideo data for AI is footage paired with labels over time: boxes that follow each object from frame to frame, the start and end of actions, and what was said. A model learns from it how things move and what happens next.
Time is what makes video different from a stack of photos. An object has to keep the same identity across frames, and an action has a start, a middle and an end that someone has to mark.
A good video dataset is filmed to a brief, with the angles, settings and lengths your model will meet, and labeled to rules agreed before the first clip.
Per-frame labels multiply any inconsistency, so it is worth reading how annotation quality gets measured before you scope a labeling pass.
A nine-second street clip with its labels split out. Pick a layer. The clip is an illustration, and the layers you get are agreed up front.
Filmed by a person from our network to your brief for angle, setting and length, with our Video Collection tool.
The same object followed from frame to frame with one id, including the frames where it's partly hidden.
Actions and events marked with where they start and end, so a model learns what happens in a clip and in what order.
What people say in the footage, transcribed with timestamps that line up with the picture. See transcription.
New footage filmed to your brief, and labels that follow it through time: objects in every frame, tracks with one id, actions and events, and the speech in the clip.
Clips of actions, gestures, driving and everyday scenes, filmed by people from our network for the angle, setting and length you need.
For new datasets and missing scenes
Our tool: Video Collection Free sample of video collection
A box around each object in each frame, with its class, drawn to rules agreed for hidden and cut-off objects.
For object detection in video
Our tool: Image Annotation Free sample of per-frame boxes
The same object followed across frames with one id, so a model learns how things move.
For tracking, counting and movement
Our tool: Image Annotation Free sample of object tracking
The start and end of each action or event, labeled from a set you agree on.
For action and gesture recognition
Speech in the footage transcribed with timestamps that line up with the picture.
For captions, video search and assistants
Our tool: Transcription Free sample of video transcription
Footage spread across the times of day, weather and settings your model will meet, planned into the brief.
For models that work outside good daylight
A handful of common uses. They show the kind of work, and they aren't client names. If your footage comes with other data, see speech data, image data and text data. Example category: automotive driving scene video.
Cars, people and signs boxed frame by frame in driving footage, so a model learns what's on the road.
Actions marked with their start and end, so a model learns to tell walking from crossing from stopping.
Each person or object followed with one id, for models that count or track movement.
Gestures filmed by many people at many angles, so a model recognizes them from anyone.
Speech transcribed with timestamps, so what's said can be searched and shown on screen.
Held-back clips from every condition, to check how your model does in each.
Each page covers what that kind of video project usually needs, and each one is built to your brief. First-person footage from head mounts, action cameras and smart glasses has its own page: egocentric video data. Every category, including speech, image and text, is on the datasets hub.
Reshooting is the expensive fix. Deciding these four things first is the cheap one.
The actions, gestures and scenes your model has to learn, and the label set for each.
Where the camera sits, the settings and light, and how long each clip runs.
Every frame or key frames, and how each object keeps its id when it's hidden or leaves the frame.
Whether what's said gets transcribed, with speaker labels and timestamps that match the picture.
Your named contact runs it from the brief to the last clip. Each step links to the service behind it.
You tell us what happens in the clips, where, and what the data is for.
The NDA is signed before any data moves.
People from our network film clips in Video Collection.
Objects, tracks and events get labeled and checked.
Speech in the footage is written down with timestamps.
Held-back clips check your model condition by condition.
What people mean by the words that come up on a video project.
A video project usually needs more than one of these. You can also compare all eight.
New footage filmed by people in our network.
Detection, segmentation and tracking in depth.
Speech in the footage turned into reviewed text.
Labeling across every data type.
Held-back clips scored against your targets.
Straight answers about video data. The services FAQ covers how the services fit together.
Footage paired with labels over time, like boxes that follow each object from frame to frame, the start and end of actions, and what was said, that a model learns from.
Yes. People from our network film clips of actions, gestures, driving and everyday scenes to your brief for angle, setting and length, with our Video Collection tool.
Yes. People from our network film first-person footage on head- or chest-mounted phones, action cameras, smart glasses or devices you send us, and the same team labels hands, objects and task steps. See egocentric video data.
Objects in each frame, the same object followed across frames with one id, actions and events with their start and end, and speech, which gets transcribed. Our computer vision page covers the image labels.
It depends on the task. Some models need every frame, others only need key frames. We agree it with you before labeling starts, along with how an object keeps its id across frames.
Yes. Speech in video gets transcribed with timestamps that line up with the footage. See transcription.
It depends on the task and how varied the scenes are, so there's no standard number. We plan it with you, and the free sample shows quality before you commit to volume.
Yes. Send a short brief and we'll capture or label a small sample free before you commit.
Yes. The NDA is signed before any data moves, we use it only for the scope we agreed, and only the people working on your project can get to it. It's returned or deleted when you ask.
Pick what you need, say what happens in the clips and roughly how much footage, and send it. We'll capture 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.