Concerns about dissimilar real image and ref image pairs

[Huage001]

Hi,

This is really an impressive project. But I have some concerns about deepfashion_ref.txt, where a small number of dissimilar real image and ref image pairs are included in training and evaluation, according to the code in function 'get_ref_video_like' of deepfashionHD_dataset.py if I understand correctly. An example used in the evaluation in deepfashion_ref_test.txt is

real_image: img\MEN\Denim\id_00000089\02_7_additional.jpg
ref_image: img\WOMEN\Dresses\id_00001748\03_4_full.jpg.

I wonder if it would hurt the training and evaluation?

Thanks so much.