why mAP in your paper is lower than others?

[llf1234]

recently, i recover your excellent work hashNet,and find that the mAP in cifar10 is near 0.8 which is higher than that report in your follow-up work hashGANs
but not merely so，
In your paper, you report CNNH achieve 0.5696 in NUS-WIDE(16bits) But In there paper is 0.611 in NUS-WIDE(12bit) In there original paper,there network is composed of three conv-pooling layers, one fully connected layer and an output layers, which is even smaller than AlexNet.
In other hash method，they use CNN-F model as backbone whose network is similiar with AlexNet ，and they achieve mAP about 0.8898 in cifar 10(16bit) ,but your paper is much lower.
So, can you tell me why your mAP is lower than other? It is very import for me to do future work。Thanks！

[caozhangjie]

I think we do not report any results on Cifar10 but use a more difficult dataset, ImageNet.
In the NUS-WIDE dataset, many image urls are invalid. Thus, the NUS-WIDE downloaded by us may be a little different than theirs. And they use the images associated with the 21 most frequent concept tags (classes), but we use all the images in the NUS-WIDE, which are associated with 81 concepts. Thus, our task is a little difficult than theirs, which causes the performance to be worse.

[llf1234]

Thanks for your interpretation！but In your professional brothers Yue Cao‘s paper:"HashGAN: Deep Learning to Hash with Pair Conditional Wasserstein GAN" they report the results on cifar10 using hashNet。the results（0.643（16bits)） is much lower than others（above 0.8）。the only differernt is that their backbone is CNN-F but not AlexNet。I recover your work with tensorflow， it also run up to 0.8。I feel so puzzled。

[caozhangjie]

You can ask them about the details of their experiments. Yue Cao's email is caoyue10@gmail.com