Drop in your images, label them batch by batch, and let a model that learns from every submit pre-label the next batches once it is good enough.
Overview
- Model: Grounding DINO - trained live on your team's submits and used to pre-label the next batches
- Task: Object detection labeling (bounding boxes)
- Input: Images from your computer and the classes you want to find
- Pipeline: import images -> take the next batch -> check how good the model is -> pre-label the batch if it is good enough -> label and submit -> measure what pre-labeling saved
- Output: A labeled dataset, new model versions, a live training report and a prelabel efficiency report
When you need it
You have images to label and want the labeling to get faster as you go, but without handing labelers pre-labels that cost more to fix than to draw. Here a live training session learns from every image your team submits, and a quality gate decides batch by batch whether the model has earned the right to pre-label. Early batches go to labelers as they are; once the model does enough of the work, each batch arrives with boxes and your team moves from drawing to correcting. The whole loop is one graph in OvalBee, built from three templates you can also use on their own: Import images by drag & drop, Live training module and the labeling step of Basic labeling pipeline.
What you need
- Images to label. Drop them onto the import node; every drop adds a new version of the input dataset.
- A Grounding DINO model in your workspace. Copy one from the Model zoo and pick it on the Live training node.
- The classes and the labelers. Set them on the Labeling node.
- A GPU. The live training session runs on one for as long as it is up.
- Labelers who turn on Live Training in the annotation tool, so their submits reach the model.
How it works
| Step | What happens |
|---|---|
| Import | Images dropped from your computer land in the input dataset |
| Next batch | Takes the next 25 images and never sends the same image twice |
| Live training session | A Grounding DINO model that keeps training on every submit and reports its prelabel quality |
| Quality gate | Checks whether prelabel quality is at least 60% |
| Pre-label | Above the bar, the session draws boxes on every image of the batch |
| Move to labeling | Moves the batch, with or without boxes, into the labeling pool |
| Labeling | Labelers draw or correct boxes and press Submit. Each submit also trains the session |
| Labeled data | Submitted images with their boxes. This is your result |
| Prelabel efficiency | Compares the pre-labels with what labelers submitted |
Each run of the flow sends one batch. Run it when your labelers need more images, or set a schedule on the batch step. When every image has been sent, the flow stops.
How labeling works
| Step | What happens |
|---|---|
| Connect | Open the annotation tool, turn on Live Training and pick the session |
| First batches | Prelabel quality reads 0% until the session has scored 10 submits, so the first batches arrive without boxes and are labeled by hand |
| Gate | Before each batch, the quality gate reads the session's prelabel quality over its latest scored images |
| Pre-label | Once quality clears the bar, the batch arrives with the session's boxes |
| Correct and submit | Labelers fix the boxes and submit. Every image still goes through a person |
What a run gives you
- A labeled dataset with every submitted image and its boxes, ready for training or for a review step.
- New model versions. The session saves its weights as your team labels, so you end up with a Grounding DINO model trained on your own data.
- A live training report that updates while the session runs and shows when the model is good enough to trust.
- A prelabel efficiency report once labelers have submitted pre-labeled images: it grades the pre-labels from Minimal assist to Production-grade and shows the share of correct, shifted, wrong-class, extra and missed boxes.
Related templates
- Live training module - the same live training and quality gate, to plug into a labeling pipeline you already have
- Basic labeling pipeline - add review so every image is checked before it reaches the final dataset
- Live training: quick start - the smallest live training graph, pre-labeling every image still waiting, without batches or a quality gate
- Active learning with live training - let the model pick which images to label next