search

layumi / Person-reID_GAN

ICCV2017 Unlabeled Samples Generated by GAN Improve the Person Re-identification Baseline in vitro
https://arxiv.org/abs/1701.07717
MIT License
321 stars 89 forks source link

How to apply LSRO to verification models? #7

Closed FlyHighest closed 6 years ago

FlyHighest commented 6 years ago

When computing verification loss, the last layer's result is [1,0] or [0,1], depending on whether the input images are of the same class.
I wonder how to handle the following two conditions:

  1. Two images are generated by GAN
  2. One image is from the dataset, the other is generated
layumi commented 6 years ago

Hi @FlyHighest Sorry for late response. I will add the training code (for the ident+verif model), and you may check it. We do not use the generated images for calculating verification loss. The main reason is too much GAN images in the dataset. The verification result will always be false (Or be confused at the early time which actually comprise the result). Because Generated&Generated image pair should be false match. Real& Generated image pair should be false match.

But we still use the generated images for calculating identification loss.

FlyHighest commented 6 years ago

Thank you for your clarification