tiangeluo / fsl-global Goto Github PK

Dear Tiange:

Thank you for your impressive paper.

Just curious, your baseline method is PN, MN, RN, which also utilize the novel classes when training. Due to the extremely data imbalance, PN, MN, RN might fail to function well in novel classes(in Table 3), you feed the network with few(5 images per class) novel class images though.

Did you try to compare with the baseline Low-shot Visual Recognition by Shrinking and Hallucinating Features(ICCV17)?

As they share the same training paradigm as your paper and manage to utilize the few novel class images when training.

Looking forward to your reply:)

hello, the work in your paper is very good. I'm a student in the college.After reading this paper, I just wonder how the global class representations update in the training? Do they update in each training episode or in other way?I can not understand this point, can you tell me something about this?
Looking for your reply, thank you.

I wondered that if you include novel classes in the training episode in the standard FSL setting, shouldn't you train a new model from scratch every time you have a new test episode? Say the dataset is MiniImageNet, normally you need to test 600 episodes to get the result, so can I say this means you need to train 600 models to handle these different episodes separately? if so, how long would it take?

hi,can you share the gen model named Hallucinator,thanks.