Attributes

A Label tells what the ModelVersion found, but not always what a Computer Vision team needs to know about it: is the fruit ripe, how severe is the defect, was the image taken at night? The Attribute of a Deployment capture this kind of information during the review of the PredictedAsset, in a structured way, so that you can then filter the production images on it and measure how your ModelVersion behaves in each condition, instead of relying on free-text notes kept outside Picsellia.

They also spare you a second pass on the same images. While you review a prediction, you correct its Shape and, in the same tool, check or fill in its Attribute, both on the Asset and on each Shape. Once the PredictedAsset is reviewed, you send it to the Feedback Loop, and it reaches the DatasetVersion with its Attribute already in place. Without this, you would have to review the Shape in the Deployment, then reopen each Asset in the Annotation tool of the DatasetVersion to fill in the Attribute afterwards.

They work the same way as the Attribute of a DatasetVersion, detailed in Attributes: the same two kinds coexist, with the same configuration. A Shape Attribute is attached to a Label and filled in for each Shape of the review, and an Asset Attribute is attached to the whole PredictedAsset.

1. Configure Attribute

Attribute are configured from the Settings tab of the Deployment, in the Labels section. The section lists the Label of the Deployment. Once some Attribute are configured, its header displays the number of Asset Attribute (1 asset attribute), and each Label row displays the number of its Shape Attribute (1 attribute).

A. Shape Attribute

A Shape Attribute is defined at the Label level and is filled in for each Shape that uses this Label. Click on + Add attribute in the row of the Label, type the key of the Attribute, and click on Add to configure it. Picsellia proposes the same options as on a DatasetVersion:

• A name, and a Type (String by default).
• Whether it is Required.
• Under Advanced parameters: a predefined list of Choices (press Enter to add the next one) and Allow multiple answers to let several Choices be selected at once. For a numerical Type, a Min / Max bound is available, as detailed in Attributes.

Click on Create to save. You can add several Attribute in the same form, and do the same for any other Label that needs some.

B. Asset Attribute

An Asset Attribute is attached to the PredictedAsset as a whole, for instance the lighting conditions of the image, independently of how many Shape it contains. It is not attached to any Label: use the ... menu in the header of the Labels section, then Add attribute. The form is the same as for a Shape Attribute.

C. Update or delete Attribute

Click on the number of Shape Attribute of a Label to update their configuration: edit them, add new ones, or remove one with its trash icon and click on Update. For the Asset Attribute, the header menu now proposes Update attributes and Delete attributes, the latter deleting all of them.

2. Fill in Attribute values

Attribute values are filled in during the review of a PredictedAsset, from the Prediction Review tool, which you open with the Review button of the Details view or the pen icon of a thumbnail. They are saved with the review, which is the human point of view on what the ModelVersion predicted.

Open the Attributes tab from the right-side floating menu of the tool. It lists, for the PredictedAsset displayed:

• An Asset attributes section, with one input per Asset Attribute.
• One section per Shape, named after its Label and its index (for instance strawberry-fruit #1), with one input per Shape Attribute configured on that Label.

Each input matches the configuration of its Attribute: the Choices if there are some, marked with a red * when the Attribute is Required. A Clear button resets a value. Click on Save at the bottom of the panel to save the review: the PredictedAsset then gets the Reviewed status.

The Details view and the Table view of the Predictions tab do not display the Attribute values: you read and edit them from the Prediction Review tool.

3. Filter on Attribute values

Recording a value pays off when you can retrieve the PredictedAsset that carry it: isolate the predictions made at night to check whether the ModelVersion drops there, or gather the fruits marked as overripe before sending them back for annotation.

From the Predictions tab, Add filter, detailed in Predictions overview, proposes two sections for the Attribute:

• Reviewed shapes attributes, on the values entered during the review.
• Predicted shapes attributes, on the values carried by the Shape of the Prediction.

In each of them, either select a Label and one of its Shape Attribute, or choose one of the Asset Attribute (Attributes on asset). Then choose how to compare: Equals, Contains, Starts with, or Ends with, and pick a value, the configured Choices being suggested. Click on Apply.

The active filter is displayed next to Add filter (for instance 'lighting' equals 'day'), with a cross to remove it, and the list narrows down to the matching PredictedAsset.

The Attribute values can also be used in the Charts of the Deployment, to aggregate them.