problem about the result

[SimonLliu]

Hello, thanks for your good jobs!
When I use your codes, i find almost all types of loss can reach the result that rank1 is over 0.60. But when i read the paper, if the method is over 0.60(rank1), it can be seen as a good method. Even the contrastive loss can achieve rank1 over 0.60. Why? Thanks

[bnu-wangxun]

Network: Inception V2
Freeze BN
Dimension: 512
Adam 1e-5 learning rate
And utilizing all the pairs (different from the lifted structure randomly sampling)

All these together is enough to make good performance with contrasstive loss.

[SimonLliu]

Thanks very much !

[SimonLliu]

Network: Inception V2
Freeze BN
Dimension: 512
Adam 1e-5 learning rate
And utilizing all the pairs (different from the lifted structure randomly sampling)

All these together is enough to make good performance with contrasstive loss.

Wow, i used the contrastive loss in Cub200 and then got rank1 0.68(but highest baseline in your chart is 0.66). When you train the model , have you use the data augmentation 'rand_crop' which is in your codes?
Thanks for your reply!

[bnu-wangxun]

yes

[IAAI-CVResearchGroup]

Found in my experiment, why BNInception in your code If there is no freeze BN-layer's result will be about two points lower than after freeze. Why is there no such phenomenon in the InceptionBN network in other people's papers?

Looking forward to your reply.