Let a zero-shot detector label your dataset from its class names alone, and see whether its boxes are good enough to prelabel with and what they say about your labels.
Main idea: Grounding DINO never sees your labels, only the class names. Its boxes are then compared with your labels in three ways: how much labeling work they would save, how they score as a detector, and where they find objects your labels miss. The same comparison answers two questions at once: can this model prelabel new data, and can you trust the labels you have.
This is the prelabeling part of Dataset intelligence, on its own. Run it when you only need that answer, or before you start labeling new data of the same kind.
Overview
- Model: Grounding DINO, zero-shot, prompted with your dataset's own class names
- Task: Prelabeling check for object detection
- Input: A labeled image dataset with bounding boxes
- Pipeline: dataset -> zero-shot labeling -> prelabeling efficiency, accuracy and missed labels in parallel -> AI gate
- Output: A prelabeling efficiency report, an accuracy report, a dataset of proposed missing labels, the detector's predictions and an AI gate answer to your question
Setup
When you add the template, the setup wizard asks two questions:
| Question | What you do |
|---|---|
| 1. Which dataset do you want to check? | Pick a labeled dataset from your workspace. Its classes are what the detector looks for. |
| 2. Which model labels the images? | Pick Grounding DINO Swin-T from the model zoo, or your own fine-tuned version. |
Press Start and the check runs. Everything else is already wired.
To bring a dataset in first, use Import from Dataset Ninja, Import from Supervisely or Import images by drag & drop.
The AI gate is answered by the language model of your OvalBee instance and charged to the workspace's credits.
How it works
| Node | What happens | Docs |
|---|---|---|
| Your dataset | The dataset you picked in the wizard. | Dataset |
| Auto-label with AI | Grounding DINO labels every image from your class names. It never sees your labels. | Grounding DINO |
| Prelabeling efficiency | How much labeling work the detector's boxes would save. | Prelabeling efficiency |
| Model accuracy | Scores the detector against your labels: mAP, an error breakdown, and boxes that disagree the same way, such as a shifted box. | Evaluate object detection |
| Find missed labels | Confident detections with no label, and objects the labels and the detector put in different classes. | Annotation gaps |
| AI gate | Answers a question about the evaluation in plain words. | AI gate |
Change the AI gate's question to the one you have about your data. The default asks whether the labels are good enough to use as ground truth.
Key concepts
The detector is a second opinion. It has never seen your labels, so where it agrees with them there is nothing to discuss, and where it disagrees, the shape of the disagreement tells you who is wrong. Mistakes of the detector scatter in every direction; a disagreement that points the same way on almost every box has one cause, and it sits in the labels.
Low detector scores need context. On everyday classes, low scores against your labels usually point at the labels. On a specialised domain a zero-shot detector may not know the classes at all; pick a model fine-tuned on that domain in the wizard.
Results
- Prelabeling efficiency. How much of the labeling work the detector's boxes would take off your labelers.
- Model accuracy. mAP against your labels, with an error breakdown and the boxes behind it.
- Missed labels. Proposed missing boxes and class conflicts, in their own dataset, ready for review.
- Predictions. The detector's boxes are kept and can serve as prelabels.
- AI gate. A plain-words answer to your question, with the reasons and the files it read.
Next steps
- Dataset intelligence — the full audit: statistics, image and label quality and embedding checks next to this one, merged into one summary report.
- Labeling with review loop — send the proposed missing labels through labeling and review.
- Live training: quick start — when zero-shot boxes are not good enough, label with a model that learns from every submit.