leesb7426 / cvpr2022-task-discrepancy-maximization-for-fine-grained-few-shot-classification Goto Github PK

Official PyTorch Repository of "Task Discrepancy Maximization for Fine-grained Few-Shot Classification" (TDM, CVPR 2022 Oral Paper)

Python 76.02% Shell 23.98%

cvpr2022 deep-learning few-shot-learning fine-grained-classification image-classification oral tdm

cvpr2022-task-discrepancy-maximization-for-fine-grained-few-shot-classification's People

Contributors

Stargazers

Watchers

Forkers

cv-ip airgods ljm198134 shaohua-long ruizhaoz lushuo57 zjn0613

cvpr2022-task-discrepancy-maximization-for-fine-grained-few-shot-classification's Issues

The function of fully connected blocks

Hi. Why use fully connected blocks to transform those two scores to two weights? There are some other easy ways like softmax. How do you design this architecture?
Thanks a lot.

Where is the paper?

Qusetions for the combination of query weights and support weights.

please ignore this issue, I ignored torch's broadcasting principle

Regarding the fourth formula

By reading your article, the channel wise representativeness score is calculated through variance。But images often have a non Gaussian distribution, is it appropriate to use variance for measurement?

About Formula 4

Hi! What is the meaning of the double vertical symbol in Formula 4, and can you provide an explanation? Is it a European distance?Is this value directly proportional to discrimination?

Recommend Projects