Workflows

A workflow is an algorithm that, given an image as input, extracts the object of interest it contains and their associated label/class (and possibly other information). The framework defines two generic workflows:

• SLDC workflow: a workflow using a multi-stage procedure to extract objects and their labels
• SSL workflow: a workflow using a end-to-end segmentation procedure to extract objects and labels

The components necessary to instantiate those workflows as well as the information they return are described below.

SLDCWorkflow

SLDC stands for Segment Locate Dispatch Classify which is the multi-stages procedure applied by the workflow.

Required components:

• Segment: an instance of Segmenter
• Dispatch: either an instance Dispatcher or a RuleBasedDispatcher (the former doesn't need to be built explicitely, see Builders)
• Classify: one classifier per possible dispatch. Each must be an instance of PolygonClassifier

Generated information for each object:

• polygons (polygon): encodes the shape and location of the object
• labels (label): label/class associated to the object
• dispatches (dispatch): dispatching label for this object (in case of a rule-based dispatcher, it is the label of the rule)
• probas (proba): probability for this label to be the true label

SSLWorkflow

SSL stands for Semantic Segment Locate which are the steps applied by the workflow. Especially, the detection and classification is implemented by an end-to-end semantic segmenation procedure.

Required components:

• Semantic segment: an instance of SemanticSegmenter

Generated information for each object:

• polygons (polygon): encodes the shape and location of the object
• labels (label): label/class associated to the object