When we switch to Current image confusion matrix, we can easily introspect how individual confusion matrix entries map to image pixels. The confusion matrix is calculated on the pixels labeled in the selected image. Note that we may miss some of the classes again, the respective rows of the confusion matrix will then remain empty.



The Current image confusion matrix is fully interactive, when we hover over the confusion matrix entries, errors at the pixel level are visualized over the image.

For example, moving the mouse over the "background" class error will make perClass Mira switch to the Errors mode. Here we can see only the labeled pixels falling into the specific field of the confusion matrix. The pixels correctly classified by our model are rendered in green and the misclassified pixels in red. In our example, we can see that from all pixels labeled as "background", some are misclassified into the "virus" and "leaves" classes.



We may wish to adjust transparency of Foreground and Background layers using the Alpha toolbar button. This allows us to see more clearly what structures the errors represent.