How to determine True positive, False positive, False negative, True negative for object detection based on image segmentation?

If I have 100 images cracked and 100 images uncracked, I created detected model based on image segementation.

I created ground truth for cracked image, I cannt creat ground Truth for uncracked its empty nothing there.

I used IoU to measure similarity and to Determin TP, TN, FP, FN

what I know if i choose threshold for IoU=0.5, then for cracked images

• TP = number of detections with IoU>0.5
• FP = number of detections with IoU<=0.5
• FN = number of objects that not detected or detected with IoU=0

My question what about uncraked images (TN) how to deal with them?

Answer 1

[Aneela]

Hi Yasmin,

In segmentation task, we are interested in the spatial overlap between predicted segments and ground truth segments. Uncracked images don’t have a “segment” to detect, hence they are considered as True Negatives (TN).

Intersection over Union (IoU) is not applicable here as there are no detections and no ground truth annotations.

Possible ways to handle Uncracked Images:

• For an uncracked image, any detection made by the model can be considered as a False Positive (FP).
• Conversely, if the model correctly identifies that there are no cracks, this can be considered as True Negative (TN).

True Negative (TN): Number of uncracked images where no detections were made.