paper vs code

[ttommasi]

Hello, I have few questions on the code and on the cvpr19 paper where the method was presented

1. In eq(2) you indicate that the similarity score for a target sample is obtained by taking the maximum out of the |C_s| outputs of G_c. Instead, in the code (file step_1.py, line 122), you compute the similarity score by summing the values of all |C_s| outputs of G_c. Which of the two choices is the correct one? Does it depend on the dataset?

2. We are trying to reproduce the results you got on Office-Home (table 3). We used the provided code with the same parameters you indicate for the A-W Office case, but we guess the parameters may be different for this dataset so we kindly ask if you can give us indications about the correct parameters for Office-Home and whether they were different for each domain-pair.

3. Moreover we wonder if you saved also the OS* (as well as ALL and UNK) results for table 3: they are missing from the paper and we did not find a supplementary material. It is a bit difficult to get a fair comparison without looking at all the metrics.