AI Object Classifier

The AI Object Classifier command applies an object classifier trained via the app’s Train Classifier dialog as a classification step: every object in the pipeline cache matching Input Classes is scored independently, using the same feature recipe used at training time, then remapped through Segmentation Mapping into this project’s own classes and applied via Match Handling.

It uses the same input-selection logic as Classify Objects, but the output class comes from a trained model’s prediction instead of a fixed shape/intersection rule - use it where the populations aren’t cleanly separable by shape criteria, or to complement a Classify Objects step earlier in the pipeline.

Available in builds with the ai Cargo feature enabled - see AI Models.

Parameters

Parameter	Description
Model Path	Path to a trained object classifier .evamodel file, saved from the Train Classifier dialog
Segmentation Mapping	Maps each of the model’s predicted classes to one of this project’s object classes (despite the name, this maps object classes, not segmentation classes - a naming quirk carried over from the pixel classifier’s equivalent field)
Input Classes	Restrict scoring to objects that already carry every one of these classes. Leave empty to evaluate every object regardless of its current class - same AND semantics as Classify Objects’ Input Classes (see Object Classes)
Match Handling	What to do with an object’s classes once a mapped prediction is applied - see below

Match Handling

Mode (GUI label)	Effect
Add class on match	Add the mapped class alongside the object’s existing classes
Reclassify on match (default)	Clear every class the object carries, then assign only the mapped class

Unmapped predictions leave the object untouched

A predicted class with no matching Segmentation Mapping entry - including a mapping row still left at its default, unset target - leaves the object’s classes completely unchanged, in either Match Handling mode. This differs from AI Pixel Classifier, where an unmapped prediction resets the pixel to Background: here, “no rule matched” means “don’t touch it,” matching how a partially-configured mapping list is expected to behave mid-setup.

The command errors if Model Path doesn’t exist, or if it points at a model trained as a Pixel Classifier instead of an Object Classifier.

Example

An object classifier debris-vs-ev.evamodel was trained on two classes: Background and EV (extracellular vesicle), using shape metrics to separate genuine vesicles from imaging debris that survived thresholding.

Model Path:            models/debris-vs-ev.evamodel
Input Classes:         cy5@spot
Segmentation Mapping:  EV → cy5@ev
Match Handling:        Reclassify on match

Every cy5@spot object is scored; objects predicted as EV are reclassified to cy5@ev, while objects predicted as Background (left unmapped) keep their original cy5@spot class untouched.