Question about the training and evaluation protocol for Table 1 and Table 2 #5

Hi authors,

Thank you for sharing this interesting work and the code.

I have a question about the reported Real3D-AD results in Table 1 and Table 2.

In the paper, the main text says that Table 1 follows a one-class training protocol, where each model is trained on annotated data from a single category and evaluated on the remaining categories with averaged performance.

Table 2 also reports cross-category P-AUROC results, where each row corresponds to training on one category and testing on the other categories. I recalculated the row-wise means in Table 2, and they match the reported Mean column. However, when averaging all cross-category P-AUROC values in Table 2, I obtain about 81.39, while the BTP point-level mean reported in Table 1 is 84.5.

I also checked that the Table 1 BTP point-level mean 84.5 is the average of the 12 per-category values in Table 1.

Could you please clarify how the Table 1 BTP results were obtained?

Specifically, I would like to know:

• For each category result in Table 1, which source category was used for training?
• Is Table 1 averaged over multiple one-class training settings, or does it use a selected/best source category for each test category?
• Are Table 1 and Table 2 produced under the same checkpoints and evaluation settings?

This clarification would be very helpful for reproducing the results correctly.

Thank you very much.