Label images while a model learns from every image you submit and draws boxes on the ones still waiting, so the work gets faster as you go.
Overview
- Model: Grounding DINO, trained live on your submits and used to pre-label the rest
- Task: Object detection labeling (bounding boxes)
- Input: Images to label and the classes you want to find
- Pipeline: label the first images by hand → the model trains on each submit → it pre-labels the images still waiting → correct and submit → repeat
- Output: A labeled dataset, a live training report and the trained model
When you need it
You have a pile of images to label, and drawing every box by hand is slow. Here a model starts learning from your very first submits. After a few images it suggests boxes on the rest, and your team moves from drawing to correcting. Every correction goes back into training, so the suggestions get better while you work. This is the smallest live training graph in OvalBee: one labeling step, no review. Start here, and grow it into a full pipeline with review when you need one.
What you need
- A Grounding DINO model in your workspace. Copy one from the Model zoo before you start. Setup asks you to pick it.
- A GPU. The live training session runs on one GPU for as long as it is up.
- Images to label. Pick a dataset in setup, or start empty and add images later.
Setup also asks for the classes and the labelers. When you press Start, OvalBee creates the datasets and starts the live training session for you.
How it works
| Step | What happens |
|---|---|
| Import | Drop images from your computer to add them to the labeling pool |
| Labeling pool | Images wait here for labelers |
| Live training session | A Grounding DINO model that keeps training for as long as it runs |
| Labeling | Labelers draw boxes and press Submit or Skip. Each submit also goes to the session |
| Prelabeling | Every 15 minutes, draws the session's boxes on every image still waiting, in the pool and in labelers' queues |
| Labeled | Submitted images with their boxes. This is your result |
| Skipped | Images a labeler set aside |
How labeling works
| Step | What happens |
|---|---|
| Connect | Open the annotation tool, turn on Live Training and pick the session |
| Start by hand | Label the first images. The session learns from each submit within a couple of training steps |
| Pre-label | Prelabeling gives every image not labeled yet boxes from the newest weights. It runs every 15 minutes, or press Run to get them now |
| Correct | Labelers fix the boxes and submit. Each fix trains the session further |
| Refresh | Each new Prelabeling run replaces only boxes the model drew, never boxes a labeler saved |
What a run gives you
- A labeled dataset with every submitted image and its boxes, ready for training or the next step of your pipeline.
- A live training report that updates while the session runs: the training loss, how close the model's boxes come to what labelers submit, and a timeline of submits and checkpoints. It shows you when the model is good enough to trust.
- The trained model, saved to your model registry every few minutes, so you can use it after the session ends.
Related templates
- Basic labeling pipeline - add review so every image is checked before it reaches the final dataset
- Active learning with live training - let a detector pick which images to label next
- Train EdgeCrafter Detector on Images - train a detector on your labeled dataset