AnnotateIt

From raw media to a training-ready dataset, all on your machine.

Label images, video and 3D point clouds. Let local AI draft annotations, review the results, then check quality, create splits, version and export your dataset.

Open the app · Get the desktop app · Documentation · GitHub

The current web app: interactive previews, ready-to-open demo projects, tutorials and a task-based model catalog.

Build and review datasets locally

• Annotate: bounding boxes, instance masks, keypoints, image classification, video tracks and 3D point clouds.
• Get AI assistance: one-click masks, text prompts, pre-trained detectors, pose models and semantic search, where supported by your device.
• Review and organize: model evaluation, dataset quality checks, train/validation/test splits and dataset versions.
• Prepare and export: image editing, capture tools, and formats including COCO, YOLO, VOC, Datumaro, MOT, MOTS and KITTI.

Core annotation and supported on-device model inference run locally. No account is required for the core app. Model and demo downloads need a connection. The optional AI Assistant connects to the provider you configure; platform and model requirements vary.

Models for local auto-annotation

Download a model once, then set it up inside a matching project. The current catalog includes 14 detection checkpoints and 10 instance-segmentation checkpoints from these families:

Current development build. Availability in installed or deployed versions depends on the release. Large variants need more memory and compute.

Small datasets to try the workflow

Ready-to-import samples: Beans · EuroSAT · Fashion-MNIST · dSprites · Open Images V7.

Each dataset card identifies its upstream source, license, sample contents and export formats.

Model provenance and validation

AnnotateIt distributes independent ONNX conversions and mirrors with upstream attribution. Each model card documents its source, license, preprocessing, tensor contract and validation limitations. App downloads use pinned revisions and SHA-256 verification. Published experimental repositories are separate from the supported model catalog.

See each model's reports for what was actually tested: a successful compatibility check is not a full benchmark or a guarantee of identical accuracy across runtimes. Original model authors retain credit; independent conversions do not imply their endorsement.