Data Curation
Catch duplicates, anomalies and labeling errors, then send them to the next step of your flow.
Report
A confident detector finds 27 two-wheelers that no existing box covers, mostly small ones at the edge of the frame. The 52 images went to labeling; the original labels are unchanged.
Compare with original labels
motorcycle 0.88
motorcycle 0.83
car 0.71Report
Validation scores on these images measure memory, not generalization. 296 extra copies went to the Duplicate assets dataset, one keeper per group.
Most similar pairs
Duplicate groups
Report
Darkest frames
Even the brightest part of the frame stays dark, so there is little left to label.
Blown highlights
Part of the frame has lost its detail at the top of the scale.
Report
92% of Resestor objects sit nearer the Resistor class than their own. Merge the two classes or relabel the 3 images behind them.
Suspect objects
Report
Most of them are rainy drives, a condition the training set barely covers. They went to the Anomalies dataset, and the other 2,313 images to Typical. Nothing was deleted.
What the anomalies are
Most typical
The images the reference explains best, as a sanity check for the baseline.
Other tools stop at a chart. In OvalBee each check produces datasets the rest of your flow can act on.
Every check is a node with dataset outputs. Send duplicates to removal, flagged assets to review and clean data to training.
Thresholds, reference datasets and embedding models are node inputs, so each check fits your data.
Every run keeps its report, with the assets behind each number. Send a link to the team or download it.
Rerun the same checks on a schedule or when new data arrives. Each result is a new dataset version.
Dataset exploration
Search your datasets in plain language, explore visual clusters on an embedding map, and save selected assets as a new dataset version.
See every dataset and version in one viewer, and narrow it down by class, tag, object size or metadata.
Describe what you need and get the matching assets, or ask for the most diverse ones.
An embedding map groups similar assets into named clusters, so outliers and rare scenes stand out.
Turn a search, a filter or a map cluster into a new versioned dataset, ready for labeling or training.
Assets
4,212 images · 6 datasets
Create dataset
Name
6 images from the embedding map selection. It starts as version v1.
Data management
Connect your storage, import any format or a public dataset, and keep every version traceable from the first upload to the trained model.
Connect Amazon S3, Google Cloud Storage or Azure Blob once, with securely stored credentials, and work with data where it already lives. Results go back to the same buckets.
Amazon S3s3://raw-images
Google Cloudgs://field-data
Azure Blobaz://captures
Bring COCO, YOLO, Pascal VOC, LabelMe, Cityscapes or Supervisely data as it is, with no proprietary format in between. Every node reads the labels in the format it needs.
Start from hundreds of public computer vision datasets, already labeled, each with its license and statistics, and import one in a single step.
Imports, cleanups and relabeling each add a dataset version. Each one records the run that made it, so any model leads back to its exact data.
Merge, filter, sample, split or remap classes with ready nodes. Select one, set its options, and every asset that flows through is handled the same way.
Keep images where
Define classes, shapes and tags once and share them across datasets. The ontology is versioned like the data, so every dataset knows which classes it was labeled with.
Classes
Tags
Capabilities
Dark framesBlown highlightsBlurry imagesBroken filesExact duplicatesNear-duplicatesOutliers
Annotation gapsLabel errorsClass balanceLabel coverageOut-of-frame labelsSplit leakageWritten findings
Search in plain wordsObject-level indexEmbedding mapClusters named by a VLMRare subtypes in a tagDiverse picks by cluster
ClassesTagsImage sizeObjects per imageObject areaBox aspectGeometryKeep or remove labels
MergeCopy and moveRandom sampleTrain, validation and test splitRename and drop classesRemove assetsNext batch
A version per changeRestore any versionTrash for 30 daysRun behind each versionShared ontologies
Drag and dropS3-compatible bucketsImages and videoDataset NinjaSuperviselyExport with annotations
COCOYOLOPascal VOCLabelMeCityscapesSuperviselyImageFolderDetected on import
SchedulesChecks on every new batchAI gateResults as datasetsShareable reportsAI agents over MCP
FAQ
It is the work of deciding which data a model should learn from: removing duplicates and broken files, catching label errors, finding gaps and rare cases, and keeping train, validation and test sets clean. A smaller, cleaner dataset usually trains a better model than a larger noisy one, and costs less to label.
No. Image quality, duplicate search, outlier detection, search in plain words and the embedding map all work on unlabeled data. Checks that compare labels, such as label diagnostics or annotation gaps, need annotations, and you can add them later in the same flow.
Checks never edit the dataset they read. They write their findings to new datasets, such as duplicates or anomalies, and every change you make on purpose becomes a new version. Removed items stay in the trash for 30 days, so an earlier state is always within reach.
Run Similar pairs with your test set as the reference. It finds exact and near-identical copies across the two sets and collects them in a dataset you can review or remove, so validation scores measure generalization rather than memory.
Yes. Dataset insights reads the reports of other checks along with the dataset statistics, and writes one summary in plain language, with a severity and a supporting chart for each finding. The individual reports stay available next to it.
You can choose between CLIP, SigLIP 2, PE-Core and DINOv3. The index stores a semantic embedding for search by meaning and a visual one for similarity, for whole images and for individual labeled objects, and it updates as the dataset changes.
All the common computer vision modalities: images, video, 3D point clouds, multispectral imagery and multimodal data that combines them. Annotations come in as COCO, YOLO, Pascal VOC, LabelMe, Cityscapes, Supervisely or ImageFolder, and can be exported as COCO, YOLO or Supervisely.
Yes. Put the checks in a flow and run it on a schedule. An AI gate can read the reports and decide whether the flow continues, so a bad batch stops before it reaches labeling or training instead of being noticed later.
