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
320 stars 89 forks source link

the label of generated DCGAN pictures #11

Open qiaoguan opened 6 years ago

qiaoguan commented 6 years ago

hi~, In your paper, for the generated DCGAN pictures, the probability for each class is the same ,i.e. 1/k（k is the numer of the class）, while in the training stage , we have to assign a label for any training sample, so i do not know how to assign label for the generated DCGAN pictures, however, i looked at your source code in prepare_gan_data.m, it seems that the label for the generated DCGAN samples is 0, i'm a little confused, can you explaining why? thanks! （chinese is also OK if it is easier to explain）

layumi commented 6 years ago

Hi @qiaoguan, Sorry for the late response. Yes, the input label for the generated images is 0 (when preparing dataset). In fact, I write the "1/k" in the loss function. If the input label is 0, the code know the input is generated sample and then assign a different loss.