Semantic Segmentation Evaluation Criteria:
Commonly used are:
- global acc (called PA in some places),
- mAcc (some places seem to be called MPA),
- mIoU
mIoU seems to be the most used, and many papers only use mIoU when showing the results.
Don’t panic if you don’t understand the picture below!
Indicator calculation method (confusion matrix):
1.global acc:
2.mAcc
is obtained by adding the five values of mAcc and taking the average.
3.mIoU
Similarly, the five iou are added and averaged to obtain mIOU
Summary:
Practical meaning:
The larger the miou, the more your predicted segmentation area coincides (similarly) with the real label area, and the better your model is, I think the same is true for mean-acc.
One thing to note: there is a confusing place, global acc is also called PA, mAcc seems to be called MPA in some places, these two seem to be a bit different
- Global Acc (PA) is global, just one, regardless of category.
- mAcc (MPA) first has different categories of acc and then sums and averages all categories to obtain mAcc
Just like a picture:
-
PA (global Acc) is: all correct pixels (regardless of category)/all pixels in a picture
-
cla-acc = the correct number of a certain category / the total number of real pixels of this category in this picture
-
cls-iou is the sum of the correct area/predicted area and the actual area of the class you predicted.
The above are the notes I made by myself. If there are any mistakes, I welcome you to correct me! !
reference:
http://t.csdn.cn/Fk7u8
https://www.bilibili.com/video/BV1ev411P7dR/?spm_id_from=333.788.recommend_more_video.0