About Table 1 in the paper

[ZhangF233]

Other segmentation-based methods, such as CSAD, also do not use the ground truth segmentation masks provided by PSAD.
Why is CSAD labeled as Supervised Masks in Table 1?

[MaticFuc]

Hi,

We have added the checkmark for CSAD as it uses different procedures for different categories (fine-grained vs coarse-grained segmentations), and it requires manual filtering of image tags.

I do agree that "supervised masks" is not the best word for this, but we could not think of a better term, and we spent a lot of time on this.

Hope this clears up the confusion.

[ZhangF233]

Thank you for your reply!

I compared the results from the CSAD paper with the results provided in this paper. The CSAD metric in Table 1 comes from the original paper, not from reproduction under different experimental conditions.

CSAD uses textual descriptions of components during segmentation but does not use Supervised Masks.

Therefore, I believe it is more reasonable to classify CSAD as 'w/o Supervised Mask' in Table 1, which is more comparable to SALAD's 96.1.