YOLO Annotator: An Offline Labeling Tool for bbox, OBB and Panoptic Segmentation Under GPL-3.0

Data labeling usually eats more time than model training, and the author of YOLO Annotator confirmed this firsthand. On a road infrastructure defect detection project, the team moved from detectors to segmentation models and could not find a suitable tool. So they wrote their own: a serverless desktop annotator that covers bbox, oriented bounding boxes, polygons, masks, and semantic and panoptic segmentation. The code is published under the GPL-3.0 license, and the export is designed for YOLO, Detectron2 and Mask2Former.

Why a "middle ground" was needed

The project started with detection. The team inherited models and experience from earlier work, but these were not enough for road-defect tasks. Cracks, potholes and pavement damage are poorly described by a rectangle, so the team turned to segmentation. The question of what to label with came up immediately.

The author splits existing solutions into two camps. Server-based tools such as CVAT and Label Studio can do almost everything, but you have to deploy and maintain them. They are not easy to learn and demand a lot of hardware. Desktop tools like labelImg install in a minute, but mostly draw only rectangles. The team needed a program that runs on a laptop, requires no infrastructure, and still supports polygon labeling.

How the tool grew its features

The first version handled bbox and polygons. Requirements then piled up as the project progressed. First came oriented bounding boxes (OBB), then masks, then a multi-user mode. Semantic and panoptic segmentation were added last. Each extension raised the question of export: data has to reach training without manual converters.

The announcement on Habr lists the main features:

• bbox and oriented bounding box (OBB) labeling;
• polygon labeling and masks;
• semantic and panoptic segmentation;
• multi-user mode;
• export to formats that YOLO, Detectron2 and Mask2Former accept without modification (the headline claims support for YOLO and COCO);
• runs on a laptop, with no server deployment;
• open source, GPL-3.0 license.

The author covers the program's internals in the full article. The announcement gives no architecture details, and it includes no labeling-speed measurements or comparison with CVAT.

What to consider before choosing it

The project looks like a tool that grew out of a real task, and the feature set shows it. It needs both OBB and panoptic. These are the formats that often break a pipeline when the annotator supports one thing and the training framework expects another. If your team is small, works on laptops and faces a similar gap between labelImg and CVAT, the annotator is worth trying on a pilot sample.

Check three things before adopting it.

• License. GPL-3.0 requires you to open-source derivative works when you distribute them. For internal use this is usually not a problem, but if you plan to embed the tool in a commercial product, clear it with your lawyers.
• Export format. Run a small batch through your training pipeline before labeling thousands of frames. This applies especially to panoptic: the category and instance ID schemes of Detectron2 and Mask2Former differ in the details.
• Multi-user mode. It is implemented without a central server. Find out how the program merges the labels of several people and resolves conflicts, especially if you need quality control and consistency between annotators.

For large teams that need role-based access, review and an audit trail of edits, server-based systems remain the logical choice. For everyone else, there is now one more lightweight option.