Offline image annotation
How to label computer vision datasets when the images can't leave your machine.
Many datasets can't go to the cloud: medical scans, footage from a factory floor, images that belong to a client, anything under a contract or a data protection rule. You still need bounding boxes and masks to train a model. This page explains what labeling offline actually requires and what to check before you trust a tool with private images.
Three ways to label private images
| Approach | What it means for your data |
|---|---|
| Cloud annotation service | Images are uploaded to the provider. Convenient, but often ruled out by contracts or policy. |
| Self-hosted web tool | Open-source tools such as CVAT and Label Studio can run on your own server. Images stay in-house, but someone has to install, secure and maintain that server. |
| Desktop application | The tool runs on the workstation that already has the images. No upload and no server; collaboration happens by sharing project files and exports. |
Self-hosted tools are a good fit for large teams that need shared review queues and can run infrastructure. For one person or a small team, a desktop tool is usually less work.
What to check in an "offline" tool
"Offline" is used loosely. These questions separate tools that really keep data local from those that only work offline for a while.
- Where do images and labels live? They should stay in folders you control, not in a hidden sync or account workspace.
- Where does AI assistance run? Auto-labeling that calls a cloud API sends your images out, even if the editor itself is local.
- Does the license need a constant connection? Some tools stop working when they can't reach a server.
- Can you export to standard formats? You should be able to leave with YOLO, COCO or Pascal VOC files at any time.
How SwiftLabel answers them
| Images and labels | Images stay where they are on disk. Annotations are saved in a local .swiftlabel project file next to them. |
|---|---|
| AI assistance | YOLO detection and segmentation, SAM click-to-segment, body and hand pose, and face landmark models all run locally through ONNX Runtime. You can use the bundled models or your own ONNX files. |
| License | The 14-day trial needs no account. A purchased license needs a one-time internet connection to activate; after that the app works offline, revalidating in the background when a connection is available. |
| Export | YOLO (v3 to v12), COCO JSON, Pascal VOC, CVAT XML, LabelMe, CreateML and more, with optional train / valid / test split. |
| When internet is used | Buying a license, receiving the key by email, activating it, and reading the website. Not for labeling. |
More detail on data handling is on the Security page.
Get started
The step-by-step YOLO labeling guide takes you from a folder of images to a training-ready export. SwiftLabel runs on Windows 10/11 and Ubuntu 22.04+, and the trial includes every feature.