Computer vision data means labels: object detection, segmentation and recognition for images and video, across automotive, retail, healthcare and other industries. We label in our own Image Annotation tool, and when you need new footage, our Image Collection and Video Collection tools capture it.
Computer vision data is images and video with labels that tell a model what's in each frame and where: a box around each object, an outline of its exact shape, or a name for what it is. A vision model learns to see from those examples.
The label type follows the job. Detection finds objects, segmentation marks their exact pixels, and recognition says which specific thing each one is. Video adds time, so an object has to keep the same label from one frame to the next.
Most of the effort goes into the hard frames, like objects partly hidden behind others or cut off at the edge of the image. The rules for those get written down before labeling starts. Our image data and video data pages walk through one labeled sample of each.
Box and mask quality decide what a detector can learn. Our guide to what makes annotation quality good or bad explains how agreement checks and review catch the errors.
Object detection puts bounding boxes around the things your model needs to find. Semantic segmentation assigns every pixel to a class. Instance segmentation draws a separate outline for each object. Recognition labels follow whatever your industry needs.
A bounding box around every object your model needs to find, each with its class, drawn to your rules for partly hidden things.
For driver assistance, autonomy and safety cameras
Our tool: Image Annotation Free sample of object detection
Every pixel assigned to a class, like road, pedestrian or sky, for when a box is too rough.
For drivable area, scene understanding and robotics
Our tool: Image Annotation Free sample of semantic segmentation
A separate outline for each individual object, even when several of the same kind overlap.
For counting, tracking and crowded scenes
Our tool: Image Annotation Free sample of instance segmentation
Labels that say which specific thing each object is, following whatever your industry needs, from product types on a shelf to signs on a road.
For product recognition, search and shelf checks
Our tool: Image Annotation Free sample of recognition labels
Video labeled frame by frame, with an agreed rule for how an object is treated across frames, so the same car keeps the same label.
For object tracking and video analytics
Our tool: Image Annotation Free sample of video labeling and tracking
When you need footage you don't have, our Image Collection and Video Collection tools capture it to your spec, and the same team labels it.
For rare conditions and new markets
Our tools: Image Collection and Video Collection Free sample of image and video collection
Common jobs for vision data, and the kind of label behind each one. They're examples of the work, not a client list.
Cars, people, cyclists and signs boxed frame by frame, so a detection model learns what's on the road.
Every pixel of road, curb and sidewalk marked, for models that have to judge exactly where it's safe to drive.
Each product recognized by type and size, so a model can check stock and find gaps on a shelf.
Each person outlined separately even when they overlap, so a model can count a queue or a crowd.
The same object followed across frames with one id, for models that track movement over time.
Recognition and evaluation tasks in clinical and health-tech settings, labeled to rules your team sets.
Bounding boxes are faster to produce and enough for many detection tasks. Segmentation is more precise and takes more effort per image. The right choice depends on what the model has to do with the result.
A model that only needs to know a pedestrian is somewhere in the frame can work from boxes. A model that has to judge exactly where the drivable road ends needs pixel-level labels. Not sure? Tell us the task and we'll say which fits.
Most labeling disagreements come from the hard frames. Decide the rule for each of these before labeling starts, and in video, agree how an object is treated across frames.
Objects behind other objects. Agree whether the box covers only the visible part or the full object.
Objects only partly inside the image. Decide whether they get labeled, and how they're marked.
Far-away people and signs a few pixels wide. Set a minimum size below which things aren't labeled.
Night, glare, rain and motion blur. Agree what gets labeled when the object is hard to make out.
A named person on our side owns the project from the first examples to delivery. This is how it usually goes.
You share the task, some example images or clips and the classes you need.
The NDA is signed before any of your data moves.
We settle boxes or pixels and the rules for the awkward frames.
We label a small sample free, so you can check it before you commit.
Annotators work in Image Annotation. New footage comes through Image Collection and Video Collection.
A review runs before delivery, with one named person for your questions.
The words that come up when you plan an image or video labeling project, and what each one means.
One team can capture the footage, label it and check your model against it. You can also compare all eight services. Example category: retail shelf and product photos, labeled for detection.
New images and video captured to your spec.
Text and audio labeling alongside the images.
Accuracy checked for each class and condition.
Generated images for rare scenes, reviewed by people.
Short answers about this service. The services FAQ covers the other services and how they fit together.
Images and video with labels that tell a model what's in each frame and where: a box around each object, an outline around its exact shape, or a name for what it is.
Semantic segmentation assigns every pixel to a class, so every car gets labeled as a car. Instance segmentation also separates the individual objects, so car one and car two each get their own outline.
Boxes are faster to produce and enough for many detection tasks. A model that has to judge exactly where the drivable road ends needs pixel-level labels. Not sure? Tell us the task and we'll say which fits.
We write the rule down with you before labeling starts: objects partly hidden behind others, cut off at the edge of the image, very small, or in poor light. In video, we also agree how an object is treated across frames.
Yes. We use our Image Collection and Video Collection tools to capture them, then label in our Image Annotation tool. Our data collection page has the details.
Automotive, retail and healthcare, among others.
The task, a few example images or clips, the classes you need and a rough volume. If you have labeling rules already, send them too.
Yes. Send a short brief and we'll 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 the labels you need, say what's in the images and roughly how many there are, and send it. We'll 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.