why I got very low IoU compaired with ssim?

Question

yasmin ismail 2023-10-19

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2036001-why-i-got-very-low-iou-compaired-with-ssim

回答： Maneet Kaur Bagga 2024-10-11

%%SSim

A = imread('new7027-157.png');

BW_groundTruth =imread('Label7027_157.png');

[ssimval,ssimmap] = ssim(A,BW_groundTruth);

imshow(ssimmap,[])

title("Local SSIM Map with Global SSIM Value: "+num2str(ssimval))

0.93

%%IoU

A = logical(imread('new7027-157.png'));

BW_groundTruth =logical(imread('Label7027_157.png'));

similarity = jaccard(A, BW_groundTruth)

similarity =

0.2822

I created GroundTuth by using Image segmenter App and export it to work space and i got high result 0.93 when I applied ssim , then i used jaccrad index(IoU) and I got 0.2822 , why I got very law result although both images seem same by eye vision, how to fix it, is it related to the way of creating groundtruth or I missed something to do after labeling before calculationg the similarity? can you help me to fix it please?

I attached the original image (70027-157.jpg) which used in Image segmenter App for labeling, the labeled image is(label7027-157), the the model output(new7027-157) which is needed to be evaluated

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Maneet Kaur Bagga 2024-10-11

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2036001-why-i-got-very-low-iou-compaired-with-ssim#answer_1530555

Hi,

After inspection of your images, I can see that both the "ground truth" and "model output" are in "grayscale" format and not in "binary" form that is the pixel values are not strictly "0" and "1". The "Jacard Index" requires binary images (logical arrays) and therefore this difference in pixel intensity is resulting in low similarity score using "IOU".

To resolve the issue you need to apply a threshold to both the model output and the ground truth to convert them into binary masks before computing IOU. You can set the pixel value greater than "0" as "1" (foreground) and others to be set as "0" (background).

Also, the low IOU score with the high SSIM suggests that the overall structure and visual similarity of the two images are good but the actual pixel-wise overlap is not close. This could be due to small misalignments or differences in the segmentation boundaries between the ground truth and model output.

Please refer to the following MathWorks documentation for more understanding:

SSIM : https://in.mathworks.com/help/images/ref/ssim.html

IOU : https://in.mathworks.com/help/images/ref/jaccard.html

Hope this helps!