See a labeled detection dataset from every side in one run, and get a summary report that ranks what it found and says what to fix first.
Main idea: one dataset goes through every check OvalBee has at once: statistics, image and label quality, embedding searches, and a zero-shot detector that acts as a second opinion on the labels. Each check writes its own report, and the summary report reads them all and turns them into one verdict and a to-do list.
Run it before you spend on training or relabeling: on a dataset someone else built, on a new delivery from a vendor, or on your own data after a big change.
Overview
- Model: Grounding DINO, zero-shot, prompted with your dataset's own class names; CLIP and DINOv3 embeddings for the similarity checks
- Task: Dataset audit for object detection
- Input: A labeled image dataset with bounding boxes
- Pipeline: dataset -> statistics and quality checks, embedding checks and detector checks in parallel -> summary report
- Output: A report from every check, an AI gate answer to your question, a summary report with findings ranked by severity and a do-first list, and candidate datasets for review
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 auto-labels the images? | Pick Grounding DINO Swin-T from the model zoo, or your own fine-tuned version. |
Press Start and the audit 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 summary report and the AI gate are written by the language model of your OvalBee instance and charged to the workspace's credits.
How it works
Three branches run in parallel and do not block each other. When all of them finish, the summary report reads every report together.
| Node | What happens | Docs |
|---|---|---|
| Your dataset | The dataset you picked in the wizard. | Dataset |
| Dataset stats | Class balance, boxes per image, object sizes, class mix per split. | Compute stats |
| Image quality | Broken and blank files, blur and exposure measured against the dataset itself, resolutions. | Image quality |
| Label quality | Duplicate labels, byte-identical images, empty images, tiny and oversized boxes, where objects sit in the frame. | Annotation quality |
| Build AI index | Embeds every image and object once, for the checks below. | AI index |
| Embedding map | An interactive map of the whole dataset, grouped into clusters. | Embedding map |
| Find duplicates | Exact copies, near-identical and similar images, and the groups that can leak between splits. | Similar pairs |
| Find anomalies | Images that look unlike the rest. | Embedding anomalies |
| Find wrong labels | Objects that look closer to another class, and the least typical objects of each class. | Label diagnostics |
| Auto-label with AI | Grounding DINO labels every image from your class names. It never sees your labels. | Grounding DINO |
| 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 |
| Prelabeling efficiency | How much labeling work the detector's boxes would save. | Prelabeling efficiency |
| 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 |
| Dataset insights | The summary report: reads every report and writes the verdict. | Dataset insights |
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. The summary report then opens frames with both layers drawn to confirm it.
Not measured is not passed. When a check cannot answer on your data, for example when its threshold sits above every score, the summary report files it under "not measured" instead of calling it clean.
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
- Dataset insights. The summary report. A verdict and findings ranked Critical, Warning, Notice and Passed, with the frames behind each one; a do-first list; open questions only the dataset owner can answer; what checked out fine and what could not be measured.
- A report from every check. Each keeps its full detail, charts and galleries. Open any node to read it.
- AI gate. A plain-words answer to your question, with the reasons and the files it read.
- Candidate datasets. Duplicates, anomalies, suspected wrong labels and proposed missing boxes, each in its own dataset, ready for review. The detector's predictions are kept too and can serve as prelabels.
Next steps
- Filter-based data cleanup — apply a mechanical fix, such as a class rename, with rules and confirm it with fresh statistics.
- Labeling with review loop — send a candidate dataset through labeling and review.
- Dataset quality control — the lighter profile when you only need class counts, balance and coverage.