Dataset - Configure your Labelmap

Now that our DatasetVersion has been created and populated with Asset, it is time to start dealing with Annotation.

Setting up a clean, well-typed Labelmap before annotating is what lets Computer Vision teams later trust their Dataset: it guarantees every Annotation follows the same structure, makes a DatasetVersion directly usable to train a compatible ModelVersion, and avoids relabeling work down the line.

1. DatasetVersion Version Type and Labelmap

Before annotating the Asset composing our DatasetVersion, we need to initialize it by defining its Version Type and its Labelmap so that Picsellia will be able to ensure consistency at each step of the process in the future.

For instance, make sure that the Version Type of a DatasetVersion is the same as the one of the ModelVersion you might be retraining in the frame of an Experiment.

A. Set DatasetVersion Version Type

To set your properly DatasetVersion, you must go to Settings of the DatasetVersion, then Configuration > Labels. From here, you can first select your Version Type and click on Set dataset version type.

The available Version Types are:

• Classification
• Object Detection
• Segmentation
• Mask
• Point
• Keypoint (skeletons)
• Line

B. Create your Labelmap

A Labelmap is the list of Label attached to a DatasetVersion. Its Label can be left flat or organized into LabelGroup, and each Label can carry its own Attribute. Independently of the Labelmap, Picsellia also lets you attach Attribute directly to the Asset as a whole, detailed below.

Once your Version Type is defined, you can start creating your Label; there are two ways to do it.

a. From Scratch

Thanks to the button + Add Label/Group, you will create a new and empty Label line. For each new line, you will be able to choose:

• The type between Label or LabelGroup (in the case of LabelGroup, you'll then have to create the Label composing the LabelGroup as detailed below)
• The name of this new Label (or LabelGroup)
• The color associated with this Label (not applicable for LabelGroup), all the Shape with that label will then be displayed with that color.

At any moment, you can rename a Label or LabelGroup by clicking on its name and confirming with the check mark icon; this updates the Label name or color of every existing Shape using it. You can also delete one or several Label from your DatasetVersion — be aware that deleting a Label deletes all the Shape carrying it from their Annotation (deleting every Label of a DatasetVersion still leaves its Annotation in place, just without any Shape), so use this feature with caution.

b. Importing them from another DatasetVersion

Thanks to the button Import labels from another version, you will be able to browse among the Label of other DatasetVersion of the current Dataset. So you just need to:

• Select the DatasetVersion name from which you want to import

• Define if you want to keep the LabelGroup and Attribute structure of the selected DatasetVersion in the current one

• Select the Label names and LabelGroup to import in the current DatasetVersion

  Once the Labelmap import is properly configured, click on the import button.

This also carries over each Label's Attribute configuration. Since no Annotation is imported at this step, you can import Label from a DatasetVersion that has another Version Type.

c. Organize Label into a LabelGroup

Besides individual Label, the + Add Label/Group button also lets you create a LabelGroup. A LabelGroup is a named container that can hold several Label, and even other LabelGroup, forming a multi-level hierarchy instead of a single flat Labelmap. In the Label list, a LabelGroup shows up as an expandable row that you can fold and unfold to reveal the Label (or nested LabelGroup) it contains.

This becomes especially valuable once a Dataset grows a large number of fine-grained Label: organizing them under a handful of broader LabelGroup (for instance, grouping car, truck, and bus under a vehicle LabelGroup) keeps the Labelmap structured and easy to navigate, both from the UI and through the SDK, which exposes the same hierarchy via create_label_group, list_label_groups, and Label.get_group() / Label.set_group().

You can create a LabelGroup as shown below:

Once a LabelGroup is created, you can add under this one either a proper Label or another LabelGroup to create a multi-level hierarchy:

Please note that at the end of the day, any Shape that will be created will have a specific Label, not a LabelGroup. LabelGroup are just there to group actual Label together, improve visibility and structure of your Labelmap.

Once created LabelGroup names can be edited, as is the case for Label, by clicking on the pen icon.

You can move a Label in or out of a LabelGroup, or change which LabelGroup it belongs to, at any time, the same way you created it.

d. Add Attribute to a Label

Each Label can also carry one or several Attribute, using the + Add attribute button next to it. Unlike Custom Metadata, which is attached to a Data/Asset as a whole, an Attribute is attached to a Label, and is filled in for each individual Shape using that Label when annotating. This is what lets Computer Vision teams capture structured information beyond the Label name itself, directly on each Shape (for instance, a color or a severity Attribute on a defect Label), instead of relying on free-text or on separate spreadsheets kept outside Picsellia.

For each Label Attribute, you'll be asked to define its type, make it required or not, and potentially define the range of possible values (depending on the type)

You can also edit or delete an existing Attribute at any time, the same way you created it. Attribute are only briefly mentioned here — detailed on a dedicated page here.

e. Asset Attribute

Beyond Shape-level Attribute, Picsellia also supports Asset Attribute, attached directly to the Asset as a whole rather than to a single Shape. This lets you capture information that describes the entire Asset rather than one specific object within it, for instance the weather conditions or the acquisition device for an image, independently of how many Shape it ends up being annotated with.

For each Asset Attribute, you'll be asked to define its type, make it required or not, and potentially define the range of possible values (depending on the type)

Asset Attribute are only briefly mentioned here — detailed alongside Shape-level Attribute in a dedicated page here.

When all your Label are created, you can visualize them and the number of Shape in the current DatasetVersion with this Label. The set of Label attached to a DatasetVersion is called the Labelmap.

2. How to create Annotation

An Annotation is a set of Shape that are linked to an Asset in a DatasetVersion. Depending on the DatasetVersion's Version Type, a Shape can be a Classification, Rectangle, Polygon, Mask, Point, Keypoints, or Line.

A Shape always has a Label among the ones defined previously, which may itself belong to a LabelGroup. Depending on that Label, a Shape can also carry values for its Attribute, and, for OCR use cases, a free text field used to transcribe the text it contains.

There are four ways to create or modify Annotation on Picsellia:

• Import Annotation file
• Annotation tool
• Import Annotations from another Dataset Version
• Processing, notably a pre-annotation Processing that runs a ModelVersion over your Asset to generate Annotation automatically

In summary, an Asset of a DatasetVersion has up to one associated Annotation. Each Annotation can have several Shape, each Shape has a Label among the Labelmap defined in the DatasetVersion Settings.

3. Visualize Annotation

Now let's consider our DatasetVersion as already annotated.

Thanks to the different views proposed in your DatasetVersion, you can see the Properties related to Annotation (number of Shape, status, annotated at/by) in Table and Details views, detailed here:

And visualize the different Shape displayed on the image of your Asset through the Grid and Details views:

4. Delete Annotations

You can delete all the Annotation related to one or several Asset, by selecting them, and clicking on More > Delete annotations.