Train/test data split

[xthan]

Hi, I noticed that you have two data splits for each dataset - one old and one new. May I know why you have two different versions? And which one should I use if I want to reproduce the numbers in your paper?

Thanks a lot!

[AliaksandrSiarohin]

Hi, initially I did just a random split based on closing items, and this corresponds to the old split. The scores for the old split can be found in the first version of the paper: https://arxiv.org/abs/1801.00055v1.
Then I realised that (https://arxiv.org/pdf/1705.09368.pdf) uses different split, so I adapt their split, the new split, for fair comparison. The scores is in the latest version of the paper: https://arxiv.org/abs/1801.00055,

[OwalnutO]

Hi, thanks for your wonderful code~

One problem. I find that the image naming format is different between your work and PG2 (https://arxiv.org/pdf/1705.09368.pdf), you name the images as 'fashion+men/women+cloth+id+view.jpg', but for PG2, they rename the id from 00001. So would you mind to provide the mapping between your image names and those in PG2? In other words, how do you collect the new split under your naming format?

Thanks~~

[AliaksandrSiarohin]

Hi,
I take all the images of Ma and compare them to my using SSIM. Unfortunately that was a while ago, and I not able to find the correct script now. I were compute all pairwise ssim and the image with highest ssim was considered a match.