Getting started with AnnotateIt
Create a project, add images or videos, annotate them, review the dataset and export it — all on the device where AnnotateIt is running. There is no account, no project limit and no automatic sync, so the first workflow includes making a backup as well as producing a training archive.
1. Choose the project type
| Project type | What you label | Native-video workflow |
|---|---|---|
| Object Detection | Axis-aligned boxes around objects | Object tracks across keyframes |
| Instance Segmentation | Editable polygons or exact object masks | Object-mask tracks across keyframes |
| Keypoint Detection | Named joints connected by a pose template | Frame-by-frame; no tracks |
| Classification — Single Label | One mutually exclusive label for the whole media item | One label over frame ranges |
| Classification — Multi Label | Several independent labels for the whole media item | Several labels over frame ranges |
| Classification — Hierarchical | A parent or child label in a nested taxonomy | Nested labels over frame ranges |
Select Create new project, enter a unique name, choose one of the six image/video task types and add labels. These types accept images, native video and Camera media: take photos or record a silent clip, review the staged result, then select Accept. A keypoint project adds a second step for choosing or designing the skeleton. To start from existing data, use the create-menu chevron and choose a standard dataset archive, a media-only folder or an exported AnnotateIt project.
For experimental 3D point-cloud semantic segmentation, choose the separate 3D Point Cloud project type. It accepts PLY/PCD, not images or videos, and has its own manual labelling and export workflow. The image/video steps below do not apply to 3D.
2. Add media
Open the project’s Dataset and use Upload media for files or a folder, or choose Camera to take photos and record silent clips; drag-and-drop works as well. Camera media stays staged for review until you select Accept. The progress dialog reports each imported file and lets you retry failures. Classification projects can assign labels to the whole upload before it starts. Images that exceed the canvas limit are marked for Prepare rather than opening as a blank canvas.
3. Annotate and review
Open Annotate from the dataset toolbar or one media item’s menu. Draw manually, use an on-device smart tool, or switch to AI Annotation Mode once a prediction source is ready. Moving to another item saves ordinary work silently. Batch auto-annotation is different: its drafts remain Pending AI Review and stay out of exports, versions and ground-truth statistics until accepted.
4. Check the dataset
- Quality reports missing annotations and required attributes, invalid shapes, exact duplicates, class imbalance, dimension outliers and video coverage; the optional deep pass also decodes media locally.
- Splits creates deterministic train, validation and test assignments with a seed, optional stratification, manual overrides and locks.
- Versions freezes the current dataset for comparison, download or restore. A version is working history in the same installation, not a backup.
- History records completed project, dataset, media, annotation, label, split and version actions.
5. Export — and back up
Export dataset creates a training archive such as COCO, YOLO, VOC, Datumaro, Supervisely Video, MOT, KITTI, MOTS or Plain ZIP, depending on the task and media. Export project creates a lossless AnnotateIt archive for one project. Settings → Backup protects the complete local profile, including every project, imported model and app setting. Nothing performs automatic cloud backup or sync.
Feature availability follows the device, not the licence. iPhone and iPad keep the annotation, dataset and format workflow but omit the Models hub, custom model import, Semantic search and Text prompt. The platform comparison lists the exact differences.