Written and maintained by CASRAI Editorial Board
Published
On this page
CVAT, the Computer Vision Annotation Tool, is software for labeling images and video with shapes, tags and tracks for computer-vision datasets. You can use the hosted service or run CVAT Community yourself. A first image workflow consists of creating a task, defining labels, drawing objects, saving the job and exporting annotations that your research code can read.
What it is used for in research
CVAT provides an editor for preparing labeled visual data. Its documented tools include rectangles for object detection, polygons and brush masks for segmentation, skeletons for connected keypoints, and tags for classifying a whole frame. These are different annotation representations: a class assigned to an entire image does not identify the position of an object within it. For the broader terminology and annotation-type comparison, see image annotation types and formats.
CVAT’s vocabulary separates a label, the type of object, from an attribute, a property of that object. It also distinguishes a single-frame shape from a track, which associates shapes across frames with the same object. In video interpolation mode, annotations on keyframes determine linearly interpolated intermediate shapes. Independent still images use shapes; a video identity that must persist across frames uses tracks.
The walkthrough below creates a standalone task for still images. It deliberately keeps the geometry to rectangles so the connection between the editor and the exported coordinates remains easy to inspect. A segmentation study would use the appropriate polygon or mask tools instead. CVAT also supports audio and 3D annotation, but those have different data representations and are outside this image task.
Install or access
CVAT Online is the hosted option, reached through the official CVAT application. CVAT Community is the free, self-hosted edition. The official repository lists Git, Docker Engine and Docker Compose as prerequisites and gives the following startup sequence:
git clone https://github.com/cvat-ai/cvat
cd cvat
docker compose up -d
Create an administrator account with the repository’s command:
docker exec -it cvat_server bash -ic 'python3 ~/manage.py createsuperuser'
For the default local setup, open http://localhost:8080 and sign in with that account. These commands describe the repository’s default stack; the linked official installation documentation covers platform-specific setup. They are installation instructions, not commands to run against an existing institutional CVAT server.
The repository says CVAT is primarily tested with Chromium-based browsers, including Chrome and Edge, and does not support Safari/WebKit. Use the documented browser environment when reproducing an interface problem. Keep the distinction between Community and the hosted service when following documentation: a feature mentioned for the broader platform is not a promise that every deployment enables it.
Step-by-step task: label an image set
This sequence follows CVAT’s task, Shape mode and navigation documentation. Use images you are authorized to upload. The task and label names are yours to supply; the menu labels below are the documented interface terms.
- Create the task. On Tasks, select +, then Create new task. Enter a Name. Leave Projects empty for this standalone example. Tasks assigned to a project inherit that project’s labels.
- Define the object class. On Constructor, choose Add label and enter a Label name. Select Continue to submit a label and add another. Choose names that match the objects in your own study.
- Add the images. Under Select files, use My computer to choose the image files. Select Submit and open. In the task’s Jobs section, follow a job link to enter the annotation editor.
- Draw a rectangle. Select Rectangle on the controls sidebar. Choose the correct Label and the two-point Drawing Method, then select Shape. Click opposite corners around the object. Adjust the boundary with the mouse if necessary.
- Inspect and continue. Select the object to inspect its card in the objects sidebar; its details contain attributes. Use the navigation bar to move through the images. The documented default shortcuts D and F move to the previous and next frame.
- Save the annotations. Use Save work and observe the saving indicator. Done completes an object being drawn; it is a different control from saving annotations for the job.
- Export the task. Open the task and choose Actions > Export task dataset. Select the required format, enter a name for the ZIP archive and select OK. For an inspectable native image export, use the CVAT image format described below.
Shape mode also provides undo and redo, and the object card allows correction or deletion of individual objects. Review a newly drawn object before repeating the same labeling action across the remaining images. This keeps an incorrect class or boundary from becoming a repeated annotation error.
Output and how to read it
The export is a dataset artifact, not a screenshot of the editor. CVAT’s native image format packages annotations in annotations.xml. Image records include a name, width and height; rectangle records include a label and coordinate fields named xtl, ytl, xbr and ybr. Attribute values can accompany an object. Native video XML instead represents tracks across frames.
Choose the training or analysis format deliberately. The official repository lists formats including CVAT XML, COCO, YOLO and Pascal VOC, as well as task-specific variants. A filename ending in JSON or TXT does not by itself identify the annotation convention. For background on one common target, see the MS COCO dataset and annotation structure.
Do not assume that every exported ZIP contains the image pixels. The export dialog documents a Save images option with edition-specific availability. Retain the original media separately when exporting annotations alone. The artifact table below compares native image XML, native video XML and a restorable task backup; it distinguishes their contents without treating them as interchangeable files.
Illustrative export check
This is a fictional worked example, not an observed research dataset. Suppose you annotate an image named sample_01.png that is 640 pixels wide and 480 pixels high. Give one rectangle the label specimen, with coordinates xtl=100, ytl=80, xbr=220 and ybr=200. The differences are 220 − 100 = 120 horizontally and 200 − 80 = 120 vertically.
After export, inspect the corresponding image record and rectangle. The image name should identify the intended source file, the dimensions should match that file, and the label should still read specimen. Check each value against the object you drew. A syntactically readable XML file is not enough if its objects refer to different images or an unintended class.
Repeat that inspection for a few deliberately different examples, such as a small object and an object near an image edge. This is an editorial validation exercise: it demonstrates what to compare before passing the exported annotations to another program, without asserting that those checks establish annotation accuracy for an entire study.
Troubleshooting
Objects do not behave as expected across frames
Check whether you created a Shape or a Track. CVAT defines a shape as a single-frame annotation and a track as linked representations of the same object. Interpolation belongs to video tracking. Drawing unrelated still-image objects as though they shared one continuing identity changes what the annotation means.
A drawing tool or label is missing
Check the task’s Label shape setting. It can restrict a label to a geometry, and the controls sidebar reflects those restrictions. For a task inside a project, check the project labels rather than expecting the standalone label-construction steps to replace them.
An object is finished, but the job still needs saving
The navigation documentation distinguishes Done from Save work. Completing a shape is not the same action as saving the current job’s annotations. The separate Finish the job action saves annotations and changes its state to Completed. Use these controls according to the state you intend to record.
The exported task contains unexpected annotations
Use Info in the editor to inspect annotation statistics, including counts by label and shape or track type. It also reports manual and interpolated frames. Compare those categories with the task you intended to create. These statistics describe the annotations present; they do not certify whether the labels are scientifically correct.
A rectangle needs correction
Shape mode supports changing the rectangle’s position and boundaries with the mouse. Its object card exposes attributes, and undo can reverse a mistaken edit. CVAT also has an occluded property for an object that is not fully visible. Record the appropriate property rather than assuming that moving a boundary alone captures every aspect of the annotation.
Free alternatives
The following alternatives refer to their open-source distributions. Their maintainers document different interfaces and supported tasks; neither row implies that all hosted services or separately packaged applications are equivalent to those distributions.
| Tool | Documented interface and data | Source installation |
|---|---|---|
| Label Studio | Browser interface for labeling images, audio, text, video and time series; the repository uses Apache-2.0 licensing. | The official repository documents pip install label-studio and starting the local service with label-studio. |
| Labelme | Python and Qt graphical image annotation with polygons, rectangles, circles, lines and points; the repository uses GPL-3.0 licensing. | The official repository documents pip install labelme. This row refers to the source package. |
Label Studio’s repository describes a configurable labeling interface across several data types. Labelme’s repository centers on a graphical annotation application and includes image-classification flags and examples for detection and segmentation. Compare those documented workflows with the geometry and output your project actually requires; the presence of an image-labeling interface alone does not establish that its export matches an existing training pipeline.
Preserve a restorable annotation task
For restoration or transfer to another CVAT server, the documentation provides task and project backups. A task backup contains task specification and annotation files, named task.json and annotations.json, with a data area. Use Backup Task or Backup Project from the action menu, and Create from backup on the relevant listing page to restore an archive.
Cloud-backed data introduces an important qualification: lightweight backups preserve metadata and annotations without copying the raw media. After restoring one, attach cloud storage on the task page before CVAT can fetch the data. A small archive can therefore depend on storage outside that archive.
As an editorial preservation practice, retain the label definitions, source-media identifiers, export format and annotation instructions alongside the files you hand to collaborators. The datasheets for datasets framework provides a place to document how an annotation dataset was constructed and what it contains.
Sources
- CVAT: Overview. Checked 2026-11-16.
- CVAT Community: official repository and installation commands. Checked 2026-11-16.
- CVAT: Vocabulary. Checked 2026-11-16.
- CVAT: Tasks. Checked 2026-11-16.
- CVAT: Shape mode. Checked 2026-11-16.
- CVAT: Menu and Navigation Bar. Checked 2026-11-16.
- CVAT: Export annotations and data. Checked 2026-11-16.
- CVAT: Native image and video format specifications. Checked 2026-11-16.
- CVAT: Backup Task and Project. Checked 2026-11-16.
- HumanSignal: Label Studio repository. Checked 2026-11-16.
- Labelme: official repository. Checked 2026-11-16.
Data: CVAT native exports and task backup artifacts
Original compilation of native CVAT image and video output structures and task-backup contents, combining the official format, export and backup documentation. Media inclusion depends on export options; lightweight cloud backups omit raw media.
| Artifact | Annotation structure | Specification or metadata | Media qualification | Official documentation |
|---|---|---|---|---|
| Native CVAT image export | annotations.xml with image records and object shapes | Image records include name, width and height; object attributes supported | Export dialog documents a Save images option with availability restrictions | CVAT format specifications; export documentation |
| Native CVAT video export | annotations.xml with tracks; shapes exported as single-frame tracks | Track attributes supported | Export dialog documents a Save images option with availability restrictions | CVAT format specifications; export documentation |
| CVAT task backup | annotations.json | task.json | Data directory; lightweight backups of cloud-backed data omit raw media | Backup Task and Project |
Sources
- CVAT native image and video format specifications (accessed 2026-11-16)
- CVAT export annotations and data (accessed 2026-11-16)
- CVAT Backup Task and Project (accessed 2026-11-16)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







