Guide

How to label a YOLO dataset offline

From a folder of images to a data.yaml you can train on, without uploading a single image.

This guide walks through labeling an object detection dataset for YOLO on your own computer. It uses SwiftLabel, a desktop annotation tool, but the steps (classes, boxes, pre-labeling, export) apply to any YOLO workflow. Everything runs locally; your images never leave your machine.

What you need

A folder of images (JPG or PNG), or a video you want to label frame by frame.
SwiftLabel for Windows 10/11 or Ubuntu 22.04+. The 14-day trial has every feature and needs no account or credit card.
Optional: a YOLOv8 or YOLO11 model exported to ONNX, if you want AI pre-labeling.

1. Create an Object Detection project

Open SwiftLabel and choose File → New Project (Ctrl N).
Select Object Detection as the project type. This enables the bounding box tool and the YOLO export formats. If you need masks instead of boxes, choose Instance Segmentation.
Save the project with File → Save Project (Ctrl S). Saving it in the same folder as your images keeps relative paths working if you move the folder later.

2. Import your images

Use File → Import Images and select your image folder. For video, use File → Import Video; each frame becomes an entry in the file list. If you already have labels in YOLO, COCO, Pascal VOC or another supported format, File → Import Dataset loads them and registers the classes automatically.

3. Define your classes

Click + Add in the Label Classes panel and create one class per object type, for example person, forklift, pallet. The class order becomes the class index in your YOLO labels, so decide on it before you start and keep it stable.

4. Draw bounding boxes

Press B for the bounding box tool, then click and drag around each object. The active class is assigned automatically.
Press D for the next image and A for the previous one.
Press V to select, move or resize a box, and Delete to remove it.
The file list shows labeled images in green, so you can see what's left at a glance.

Box consistently: cover the whole visible object, keep boxes tight, and label every instance of a class in an image. Inconsistent boxes cost more accuracy than a smaller dataset does.

5. Speed it up with AI pre-labeling

Once you have a model, labeling becomes reviewing. SwiftLabel runs YOLO detection models locally through ONNX Runtime, with no internet connection.

Set your model in Settings → Models → Detection. Any YOLOv8 or YOLO11 detection model in ONNX format works, and class names are read from the model's metadata.
Click AI Auto-Label in the side toolbar.
Adjust the confidence and NMS IoU thresholds, then run it on the current image or on all images.
Review the predictions, fix or delete wrong boxes, and add missed ones.

A practical loop for custom classes: label a first batch by hand, train a small model on it, export that model to ONNX, and use it to pre-label the rest. Each round makes the next one faster.

6. Export for YOLO training

Open File → Export Dataset.
Choose YOLOv8 or YOLOv11. Both write Ultralytics-style normalized labels and a data.yaml.
Tick Split into train / valid / test and set the ratios.
Keep Copy image files to output directory enabled so the export folder is self-contained.
Pick an output folder and click Export.

Then train with the Ultralytics CLI, pointing it at the exported file:

yolo detect train data=path/to/export/data.yaml model=yolo11n.pt

The same project can also be exported to COCO JSON, Pascal VOC, CVAT XML and other formats if you train with a different framework.

Why label offline?

Your images and annotations stay on your disk, there's no server to set up and no account to manage, and the AI models run on your machine. That matters when the images belong to a client, a patient or a production line. Read more in Offline image annotation.

Try it

Download SwiftLabel and label your first dataset during the free 14-day trial. A license is a one-time $29 for one user on two devices.

Download the free trial