search

ronekko / deep_metric_learning

Deep metric learning methods implemented in Chainer
MIT License
158 stars 44 forks source link

CUB200 cannot match any papers' result, although cars196 performs better. #3

Open unluckydan opened 7 years ago

unluckydan commented 7 years ago

Do you have any idea about this? And I notice one thing, your crop is 224, but these four papers use 227. So I have no idea about what kinds of thing make the difference in CUB200.

ronekko commented 7 years ago

Thank you for sharing your results. And I also observed similar results.

One of the purpose of this repository is to compare the performances among various methods in the same condition, rather than to reprecate as possible as their individual results. So there are some difference of configurations and results between the papers and this repository's.

In particular, for all methods the experimental configuration is roughly according to the paper of N-pair loss, i.e. the pre-trained network is BVLC's GoogLeNet and its input size should be 224.

I'm wondering that overall results reported in the paper of Clustering loss are significantly lower than other papers. I suspect one of the causes comes from that they used GoogLeNet with batch normalization (in contrast to BVLC's one does not use BN).

unluckydan commented 7 years ago

I will try to figure it out, if there is any news I will let you know.

ronekko commented 7 years ago

That's grateful. Thank you!

unluckydan commented 7 years ago

I modified something like dropout and using 227 cropped images, but nothing improves. That is very wired.

qk-huang commented 6 years ago

Hello, can you share your results? Thanks

huanhuancao commented 6 years ago

Hello, can you share your results? Thanks

So do you get good performs on the three datasets? I got these [best] soft results by main_n_pair_mc.py.