Assets and annotations

An asset is one stored item in a workspace: an image, a video, an annotation, a model checkpoint, a report. Datasets group assets, nodes produce them, and the preview widgets on the canvas show them.

Annotations

Every image or video in a dataset has one annotation, which holds everything drawn or tagged on it. An annotation is an asset of its own that points at its image, so it moves with the image when a node samples, filters or copies a dataset, and a derived dataset never loses its labels.

Annotations are written by the annotation tool, by prediction nodes and by the import nodes. The same image can carry different annotations in different datasets. That’s how a prediction dataset and a ground-truth dataset can describe the same images, which is exactly what evaluation nodes compare.

Objects in an annotation name their class by its name in the dataset’s ontology. Supported geometries include rectangles, oriented boxes, polygons, masks, polylines and points.

Tags and metadata

Assets carry tags: simple key-value pairs meant for filtering. Import nodes add some — Dataset Ninja imports tag each image with the split it came from, for example — and prediction nodes record which model produced an annotation. Nodes that filter or split data accept tag conditions written as images:key=value or annotations:key=value, and evaluation reports can break their metrics down by a tag.

Richer structured information is kept as metadata and is visible in the asset details, but it isn’t used for filtering.

Browsing assets

Storage → Assets lists every asset in the workspace, or the assets of one dataset when you open it from the Datasets tab. For each asset you see its type, author, size, ID and tags.

To actually look at images and their labels, use the dataset preview on a node in your flow. It opens a viewer with the image, its annotation drawn on top or side by side, the asset’s tags and metadata, and a tag filter to narrow down what you’re looking at.

Where the files live

Files are kept in the object storage configured for your OvalBee installation: an S3-compatible service, Azure Blob Storage or a server filesystem. Workspace members never deal with buckets directly — uploads, previews and downloads go through OvalBee. How storage is set up is covered in Platform configuration.

Related pages

• Datasets and versions — grouping assets and tracking changes.
• Ontologies — the class names annotations refer to.
• Labeling — where annotations come from.