Image data for AI is photos your model learns to see from, each one captured to a brief and labeled with what it shows. People from our 30,000+ network take the photos with our Image Collection tool, and our annotators label them in our own Image Annotation tool.
photo 0212 · captured
capturedboxes and classes
labeledwhat the model learns
Image data for AI is a set of photos, each paired with labels that say what's in it and where, that a computer vision model learns from. The photos teach what things look like, and the labels teach what they are.
A good image dataset is captured to a brief. It covers the conditions your model will face, like light, angle, distance and background, and the objects, places and people it has to recognize.
Labels come in levels: a class for the whole photo, a box around each object, or an outline down to the pixel. The level depends on what your model has to do with the result.
Whether boxes and masks agree from one annotator to the next is a quality question as much as a labeling one, and the guide on annotation quality walks through it.
One street photo, labeled four ways. Switch between them to see how much a single image can carry. The photo is an illustration, and you choose which layers you need.
Captured by a person from our network to your brief, on the kind of phone and in the kind of light your model will meet.
A tight box around every object your model has to find, each with its class, drawn to rules agreed for hidden and cut-off objects.
Every pixel assigned to a class, or each object outlined on its own, for when a box is too rough. See computer vision for the difference.
The conditions each photo was taken in, recorded with it, so you can check coverage and find where your model struggles.
New photos captured to your brief, and labels from a single class per photo down to an outline per pixel. Photos are captured in our own Image Collection tool and labeled in Image Annotation.
People from our network take new photos of products, shelves, streets and documents, in the light and on the phones your model will meet.
For new datasets and missing conditions
Our tool: Image Collection Free sample of image collection
One label for the whole photo, from a set you agree on, for models that sort or tag pictures.
For photo tagging, search and sorting
Our tool: Image Annotation Free sample of image classification
A box around each object with its class, for models that have to find things in a photo.
For object detection and driver assistance
Our tool: Image Annotation Free sample of bounding boxes
Every pixel given a class, or each object given its own outline, when the exact shape matters.
For scene understanding and robotics
Our tool: Image Annotation Free sample of segmentation
Labels that say which specific thing each object is, from product types on a shelf to signs on a road.
For product recognition and shelf checks
Our tool: Image Annotation Free sample of recognition labels
Photos spread across the regions, places and contributors your brief asks for, so no single setting dominates.
For models that work for every user
Typical jobs for image data. They show the kind of work we do, and none of them is a client name. Working with more than photos? There are pages for speech data, video data and text data as well. Example category: retail shelf and product photos.
Cars, people and signs boxed in street photos, so a detection model learns what's on the road.
Products recognized by type and size, so a model can check stock and spot gaps.
Every pixel of road, sky and building marked, for models that need the exact layout.
Whole photos labeled by class, so a model can sort and find pictures.
Night, rain or unusual angles captured on purpose, or generated when they're hard to find.
Held-back photos from every condition, to check how your model does in each.
Each page covers what that kind of photo project usually needs, and each one is built to your brief. Every category, including speech, video and text, is on the datasets hub.
Most image datasets that disappoint were under-specified, not too small. We pin these four down with you first.
What has to appear in the photos, and the label set each one gets.
The light, angles, distances and backgrounds your model will meet, planned into the capture brief.
Which devices take the photos, and how they're spread across regions and contributors.
A class for the whole photo, a box per object, or an outline to the pixel, chosen by what the model has to do.
One person on our side owns it from brief to delivery. Each step links to the service behind it.
You tell us the subjects, the conditions and what the images are for.
The NDA is signed before any data moves.
People from our network capture photos in Image Collection.
Annotators label them in Image Annotation.
A review runs before anything is delivered.
Held-back photos check your model condition by condition.
The vocabulary of a vision project, a sentence or two each.
Image projects tend to pass through a few of these. You can also compare all eight.
New photos captured by people in our network.
Detection, segmentation and recognition in depth.
Labeling across every data type.
Generated images for rare scenes, reviewed by people.
Scores broken down by class, light and angle.
What buyers ask us most about image data. The services FAQ covers how the services fit together.
A set of photos, each paired with labels that say what's in it and where, that a computer vision model learns from.
Both. People from our network capture new photos to your brief with our Image Collection tool, and our annotators label yours or ours in our Image Annotation tool.
Our latest was an image collection: 16,000 images from 2,400 participants.
It depends on the task, the number of classes and how varied the conditions are, so there's no standard number. We plan it with you, and the free sample shows quality before you commit to volume.
Image classes, bounding boxes, semantic and instance segmentation, and recognition labels. Our computer vision page goes into each one.
Yes. Conditions go into the capture brief: light, angle, distance, background and the phones people use, so the photos look like what your model will meet.
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's in the images and roughly how many, 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.