Comparison with CVPR2020 VAL on FashionIQ

[BrandonHanx]

Hi all, thanks for your work.

I wonder why not compare your result with CVPR2020 paper named VAL?

It seems like VAL is published earlier and performed better than you on FashionIQ. Does your work have a different setting with VAL?

Thanks in advance. Please point out if I am wrong.

[ecom-research]

We did not know this work. We searched for works on FashionIQ but did not find any relevant research papers. Perhaps, it was done parallel to our work. Nonetheless, the work seems interesting.
One different setting in comparison to our experimental settings, which I quickly found out was that VAL are using "side-information" for training FashionIQ. To quote from Sec 4.4: "We use the side information from Fashion200k as auxiliary supervision for pretraining in VAL (Lvv+Lvs)".

Perhaps this is the reason for not only the superior performance of VAL but also TIRG model.