perClass Mira 2.0 brings error visualization. It enables the user to understand where the classiifer fails and facilitates the model improvements.
In this example, we have a French fries project with a trained model. The close-up of a specific labeled area:
When we select the Errors mode in the toolbar (or press E key), the visualization will highlight correct model decisions in the labeled areas in green and incorrect decisions in red:
We can see, that there are three regions where the current model misclassifies the labeled information.
For completness, we show the decisions of the model:
It is clear that the central area is classified as rotten while our middle label stroke assigns it to the potato skin class.
The background label stroke also reaches into the potato piece resulting into incorrect classification.