Image annotation workflow from upload to export
This is the image annotation workflow inside ImageLabeling, step by step. You upload raw images, an AI model draws the first labels for your classes, your team corrects them in a keyboard first editor, reviewers approve or send work back, you freeze a dataset version and export it in the format your training code reads. Each step below shows what you do and what the workspace does for you.
Step 1. Upload your images
Drag and drop images or a zip file into a project, or paste a cloud bucket URL and the workspace imports the files from there. JPG, PNG and WebP images go straight into the project queue. If you already have labels from an earlier project, import them in COCO JSON, YOLO txt, Pascal VOC XML or CSV and they appear on the images as editable objects.
- Drag and drop, zip archive or cloud bucket URL
- Existing labels imported in the same open formats you export
- Video teams upload extracted frames as images
- Storage from 10,000 images on Starter up to 5,000,000 on Enterprise
Step 2. Let AI pre-label every image
Before anyone opens the editor, the AI model reads each image against your label schema and proposes boxes, polygons, segmentation masks, keypoints or an image class. Every object carries a confidence value, so you know where to look first. The details of prompts, confidence and quota are on the auto labeling page.
- Draft labels for the classes you defined, not a generic list
- Confidence shown per object, sortable in the object list
- One AI pre-label per image counts toward the monthly AI quota
Step 3. Fix the draft in the editor
Labelers open the queue and walk through images with the keyboard. A accepts an object, B draws a box, P a polygon, K a keypoint, D moves to the next image. Wrong classes are fixed from the class picker, attributes such as occluded or truncated are set per object, and undo history covers every change. The full shortcut list lives on the image labeling tool page.
- Accept, edit or reject each AI object in one click
- Brush and polygon tools for segmentation masks
- Zoom for small objects and dense scenes
Step 4. Review and QA before anything ships
A reviewer sees each labeled image after the labeler. Approve it, or send it back with a comment that points at the exact object. On Team and higher, consensus runs on a sample of images: two labelers label the same image and disagreements land in the issue queue. Business adds a QA score per labeler, so you see who needs coaching and who can review.
- Labeler then reviewer, with comments on send back
- Consensus on a sample of images (Team and higher)
- QA score per labeler (Business and higher)
- Issue queue for disputed or unclear images
Step 5. Freeze a dataset version
When a batch is approved, freeze it as a version. A frozen version never changes, so a model trained on it can always be traced back to the exact labels it saw. Compare two versions and the workspace lists added, changed and removed labels. Set the train, validation and test split once, and it stays stable across later versions.
- Frozen versions you can retrain on months later
- Version compare by added, changed and removed labels
- Train, validation and test split stored with the version
Step 6. Export in the format your code reads
Export a frozen version as COCO JSON, YOLO txt per image with a data.yaml, Pascal VOC XML or CSV. Re-exports never count toward the AI quota, so you can export the same version for two frameworks at no extra cost. Format specifics are on the YOLO labeling tool and COCO annotation tool pages.
- COCO JSON, YOLO txt plus data.yaml, Pascal VOC XML, CSV
- Exports by version, so the file always matches a known state
- API export on Team and higher, webhook events on Business
How the image annotation workflow changes with team size
A solo ML engineer on Starter often skips formal review and acts as both labeler and reviewer. A startup on Team splits the roles and uses consensus to calibrate new labelers. A company with an in-house labeling team on Business adds QA scores, audit log and webhooks that tell the training pipeline when a batch is reviewed. If you would rather not staff labelers at all, data annotation services add vetted annotators who follow the same workflow inside your workspace. Plans and limits are on the pricing page.
FAQ
How long does one pass through the workflow take?
It depends on image count, annotation type and how much review you require. Boxes on clear images move fastest because most AI objects are accepted as they are. Dense segmentation takes longer because masks need brush work on edges.
Can I skip the review step?
Yes. Review is a setting per project. On Starter one reviewer can approve work, and on small projects the labeler can be the reviewer. For datasets that feed production models, keep review on.
Can I go back and relabel an old version?
A frozen version stays as it is. You relabel in the working set and freeze a new version, then compare the two to see exactly what changed.
Does the workflow work for video?
Upload frames as images and they go through the same steps. There is no native video timeline.
Run the workflow on your own image
Upload one image, pick a task and see AI labels drawn on it. Then create your workspace for the full workflow.