Evaluation of Segmentations

Semantic Evaluation:

In the semantic evaluation all pixels are considered equal. Values are calculated in a case-wise manner and are then aggregated over the whole dataset.

Cases (considered per class):

All values are pixel-wise TBD

Instance Evaluation:

Through connected components all pixels are grouped into instances. For each case the instance-wise values are calculated, then averaged over the case and then over the whole dataset.

Cases (considered per class):

All values are pixel-wise

1. No GT Segmentation in Image and No PD Precision = Recall = DICE =
2. No GT Segmentation in Image but PD Precision = Recall = DICE =
3. GT Segmentation in Image but No PD Precision = ? Recall = 0 DICE = ?
4. GT Segmentation in Image and PD Precision = (TP / (TP + FP)) Recall = (TP / (TP + FN)) DICE = (2 * TP) / (2 * TP + FP + FN)