Computer vision annotation is the labeling of images and video so computer vision models can learn to see, identify, locate, and segment objects. It spans the full range of visual labels, classes, boxes, masks, and keypoints, across both still images and video.
Computer vision annotation is data annotation scoped to visual data: the images and video that computer vision models learn from. It covers every visual label type, image-level classes, bounding boxes, polygons and segmentation masks, keypoints, and, in video, object tracks and events, and it is the supervised foundation of the field, the ground truth behind detection, segmentation, classification, and tracking.
The term sits between the broad "data annotation," which spans all modalities including text and audio, and the narrower "image annotation," which covers only still images: computer vision annotation includes both images and video.
Key takeaways
It is annotation for computer vision, labeling images and video for visual tasks.
It spans classes, boxes, masks, keypoints, and, in video, tracks and events.
It is broader than image annotation, which is images only, and narrower than data annotation, which is all modalities.
What computer vision annotation provides
Common computer vision annotation types.
Common computer vision annotation types.
Type
What it labels
Classification
A class for the whole image or frame
Bounding boxes
Object location for detection
Polygons and masks
Object shape for segmentation
Keypoints
Landmarks for pose and structure
Tracks and events
Identity and timing across video
How it works
You define a schema, label the visual data against it, increasingly with model assistance, review for quality, and export for training. FiftyOne is built for visual data: it stores images, video, and their labels together, visualizes them, and lets you curate what to label and compare labels against model predictions.
Why it matters
Computer vision annotation is where nearly every vision model gets its ground truth, and its quality and coverage set the model's ceiling. The highest-leverage decision is coverage, not volume. Vision models fail on what they rarely saw labeled, the unusual viewpoint, the odd lighting, the rare class, so a dataset's blind spots, not its size, predict where the model breaks. Curating for coverage of the hard, rare cases beats labeling more of the common ones, which is why visual datasets are increasingly built by finding gaps rather than scraping more easy data.
Frequently asked questions
What is computer vision annotation?
Labeling images and video so computer vision models can learn to detect, segment, classify, and track.
How is it different from image annotation?
Image annotation covers still images. Computer vision annotation includes both images and video.
How is it different from data annotation?
Data annotation spans all modalities. Computer vision annotation is the visual subset.