Boxes, points, polygons and poses over a picture, in its own pixel space: predicted solid, ground truth dashed, the previous version dotted, each on a halo; class words, a line-style legend, selection by keyboard, raster overlays, loading and failed states.
- Predicted
- Ground truth
Annotated image evaluation
Install
npx shadcn add @syncrobotic/annotated-image Also installs: lucide-react, @syncrobotic/cn, @syncrobotic/skeleton
Overview
A picture with boxes, points, polygons or poses drawn over it: what a model predicted, what a person labelled, the previous version.
When to use
- Reviewing a model's detections against ground truth, frame by frame.
- Comparing two versions of a model or of a dataset's labels on the same picture.
- Showing what a camera saw when an event fired, with the object that fired it.
Use something else
- Many pictures at once, without marks: Image grid.
- Drawing or editing boxes: a labelling tool; this one shows and selects, it does not edit.
More examples
States and real uses. Each example is also a Storybook story and a visual regression test.
- Predicted
- Previous version
Annotated image previous version
- Predicted
- Ground truth
Annotated image keypoints
- Predicted
- Ground truth
- Class
- forklift
- Source
- Predicted
- Score
- 0.88
Annotated image selectable
- Predicted
- Ground truth
- Class
- forklift
- Source
- Predicted
- Score
- 0.88
Annotated image labels selected only
- Predicted
Annotated image depth overlay
Loading picture
Annotated image loading
- Predicted
- Ground truth
Annotated image failed
Usage rules
- Coordinates are the picture's own pixels; give its natural
widthandheight. The box keeps that aspect ratio, so the marks line up at any size. - Line style says where a mark came from (predicted solid, ground truth dashed, previous dotted); the class is the word on it. Never a colour per class, never status colours.
- Keep the legend (or an AnnotationLegend nearby) whenever more than one kind is on screen.
- Crowded picture, or a picture shown phone-wide:
showLabels="selected". The words are still read out in the summary or on each annotation's button. - Selectable only when something happens on select (a details panel, a review); otherwise the picture is one image with a summary.
Changelog
- 0.18.02026-11-19First release: AnnotatedImage (src or image, width, height, annotations of box, point, polygon and keypoints) with an SVG in the picture's pixel space and non-scaling strokes; predicted solid, ground truth dashed, previous dotted and faded, each on a halo; selected 3 px with handles; class words (all, selected or none); a role=img summary, or with onSelectedChange one Tab stop, arrows, Enter and Esc; overlays; loading and failed states; and AnnotationLegend for the kinds present.