Dataset - DatasetVersion overview

The Assets overview is a shared place to access and organize the Asset of a DatasetVersion. It provides a simple way to view and manage your Asset, their Annotation, and associated information.

For Computer Vision teams, being able to quickly browse, sort, and filter the Asset of a DatasetVersion is what makes it possible to spot mislabeled or missing Annotation, review edge cases, and curate the right Asset before training a model.

A DatasetVersion is composed of four tabs: Assets (covered on this page), Analytics, Charts, and Settings.

Your Dataset can have several DatasetVersion. In the header, you'll retrieve the name of the current Dataset and the name of the current DatasetVersion. You can click on the Dataset name at the top of the page to access the full list of DatasetVersion available for it.

With Picsellia, you have three different viewing options within a DatasetVersion:

• Grid View: A visual layout that displays multiple Asset at once.
• Table View: A view that focuses on the information associated with each Asset.
• Details View: A comprehensive view that shows all the elements of a specific Asset, including the file itself and its related details.

You can easily switch between these views.

1. Understanding Asset in Picsellia

Asset refers to an image or a video that is part of a DatasetVersion. As explained here, Data is an element of the Datalake that represents an image or video physically stored through a Cloud Storage Integration, while an Asset is an element of a DatasetVersion.

An Asset is always related to Data, which means that if a file is used in several DatasetVersion, we will have several Asset related to a unique Data in the Datalake. This allows storing the file only once even if it is used in many DatasetVersion.

The Asset in Picsellia's DatasetVersion includes both the file and its Properties, inherited from its Data. These Properties can be viewed using the Table button, which opens the Table view, detailed here.

In addition to these Properties, each Asset also inherits the DataTag, Metadata and Custom Metadata of its Data. As a reminder, two Metadata systems coexist:

• A fixed Metadata schema: the same predefined list of keys is available to every Data (for instance, GPS coordinates or acquisition date).
• A Custom Metadata system: each user can define their own key/value fields.

Unlike Properties, neither is visible from the Table view, only from the Details view. Creating, editing, and deleting Metadata (fixed or custom) is detailed here.

The key difference between an Asset and its underlying Data is that an Asset can carry an Annotation, whereas Data can't.

2. Asset visualization

Inside the DatasetVersion, you can switch between views using the appropriate button.

As is the case for Datalake, thumbnailed versions of the images are used for display to ensure smooth navigation with reduced loading time, even in case of large volumes of images.

A. Grid View

The Grid View lets you visualize multiple Asset from the DatasetVersion at once. You can adjust the zoom and change the aspect ratio (Square, Landscape, Portrait, 16:9) to match your Asset's format.

In case the DatasetVersion is annotated, you can filter the Shape to display whatever their type: Classification, Rectangle, Polygon, Mask, Point, Keypoints (skeletons), or Line. You can hide/display Shape by Label. For instance, hide all the Shape with the Label car.

The Showing dropdown also allows you to display a given Projection name in the grid instead of the thumbnail created for the original image; you must first register the Projection name in the Organization settings.

B. Table View

The Table View primarily focuses on displaying all the information associated with each Asset from the DatasetVersion.

The Table view provides access to the following Properties:

• preview: thumbnail of your Asset
• filename: complete filename of your Asset
• width/height: dimensions in pixels of your Asset
• number of shapes: amount of Shape contained in the Annotation of the Asset
• status: review status of the Annotation (e.g. Pending, Accepted, Rejected)
• annotated at / annotated by: when and by whom the Annotation was created or last modified
• source: source of your Asset (e.g., direct_upload, AWS_bucket, serving, etc.)
• created at / created by: when and by whom the Asset was created on Picsellia
• asset tags: the AssetTag attached to the Asset, detailed here
• data tags: the DataTag attached to the Data the Asset is related to.

Please note that some of the displayed Properties are related to the Asset, not the linked Data. Neither the fixed Metadata nor Custom Metadata are visible from the Table view, only from the Details view.

C. Details View

The Details view aims to show all the elements of a specific Asset from the DatasetVersion. You can choose the aspect ratio (Square, Landscape, Portrait, 16:9), zoom in/out on the Asset and modify the display of the Shape (display only Shape with a given Label, Hide/Unhide all the Shape, Hide the Label etc...)

Properties, Tags (DataTag and AssetTag), Metadata and Custom Metadata related to the Asset, as well as its potential Annotation, are displayed on the right side of the view. Custom Metadata creation and editing are still available here as long as you have the necessary rights on the Datalake that the Data related to the current Asset belongs to.

3. Exploration modes and filtering

By default, the DatasetVersion is in Query Language mode: the main area shows the Grid, Table, or Details view described above, and the search bar lets you filter Asset using the Query Language, detailed here. This search bar stays available regardless of the exploration mode selected.

In addition to the Query Language, Asset can also be filtered by date using the Select Date button, or by Custom Metadata using the Add filter button.

Next to the search bar, a mode selector lets you switch to another exploration mode, which replaces the main view:

• Embeddings: mainly displays a UMAP projection of your Asset embeddings, with clustering, letting you explore and visualize different sets of points, in addition to similarity or text-to-image search. Detailed here.

  

4. Asset ordering

From any Assets overview view, you can select the sorting button to access more sorting options, and the arrow button next to it to switch between ascending and descending order.

You can sort the Asset in the DatasetVersion by:

• Creation Date
• File Name

5. Asset selection and actions

To select Asset from the DatasetVersion for further actions, you can:

1. Manually select by checking the tick box on each Asset whatever the current view (Details, Grid or Table)
2. Select all or the first x displayed Asset from your DatasetVersion.

The number of selected Asset against the total number of Asset currently displayed/filtered in the DatasetVersion will be shown.

Once Asset are selected, action buttons let you:

• Launch a Processing, detailed here.
• Attach or detach AssetTag, detailed here.
• Duplicate the selected Asset within the same DatasetVersion.
• Create a new DatasetVersion, or copy the selected Asset to an existing one, detailed here.
• Delete the Annotation of the selected Asset, or delete the Asset themselves, detailed below.

6. Asset or Annotation deletion

To delete Asset from your DatasetVersion, select the Asset to be deleted, then click on More > Delete assets; in that case, the selected Assets and associated Annotations (if existing) will be deleted from the current DatasetVersion.

By clicking on Delete Annotation, you will delete the Annotation related to the selected Assets after that, selected Assets will be raw images or videos, free of Annotations.

7. Settings

From the DatasetVersion Settings tab, you can:

• General: rename the DatasetVersion.
• Asset Tags: manage the AssetTag of the DatasetVersion, detailed here.
• Labels: manage the Label (and their Attribute) of the DatasetVersion, detailed here.
• Import Shapes: import Annotation from a file, detailed here.
• Embeddings > Image Embeddings: activate Visual Search for the DatasetVersion, detailed here.

From the General section, you can:

• Edit the DatasetVersion name.
• Edit the DatasetVersion description.
• Lock the DatasetVersion
• Transfer it to another Dataset
• Delete it (for security reasons, you will be asked to type the DatasetVersion name in the deletion pop-up).

A. Lock a DatasetVersion

Once locked, a DatasetVersion cannot be modified anymore: its Asset, Annotation, Label, and general information become read-only, and a Locked badge is displayed next to its name everywhere it appears.

For Computer Vision teams, locking a DatasetVersion is what turns the versioning system's reproducibility promise into an enforced guarantee rather than a convention: once a Training or Evaluation has run against a given DatasetVersion, locking it protects against anyone (including yourself) accidentally editing it later, so you can always be certain of exactly which data produced a given Model.

A locked DatasetVersion can be unlocked at any time from the same Danger zone, which restores the ability to edit it.

Lock is detailed alongside Transfer here.

B. Transfer a DatasetVersion

A DatasetVersion can be transferred to another Dataset of the same Organization, among the ones you have the rights to create a version of. This is detailed here.