search

alibaba-mmai-research / HyRSM

Code for our CVPR 2022 Paper "Hybrid Relation Guided Set Matching for Few-shot Action Recognition".
Apache License 2.0
26 stars 5 forks source link

About the Class Activation Maps(CAM) in your paper #5

Open eveleslie opened 2 years ago

eveleslie commented 2 years ago

Hello, I am glad to see your open-source code. Your work inspires me a lot. I want to ask how the class activation diagram in the paper is drawn. Because I know that the class activation map function provided in Torchcam allows only one input, while few-shot work often has multiple inputs. How do you handle this, and is it convenient to provide the code for me to learn? I am looking forward to your reply!

wangxiang1230 commented 2 years ago

Hello, I am glad to see your open-source code. Your work inspires me a lot. I want to ask how the class activation diagram in the paper is drawn. Because I know that the class activation map function provided in Torchcam allows only one input, while few-shot work often has multiple inputs. How do you handle this, and is it convenient to provide the code for me to learn? I am looking forward to your reply!

Hi, thanks for you attention. For class activation map, we input an episodic task to the model. Note that the support and query videos need to be entered into ResNet-50 together when visualizing (i.e. support and query are concatenated and then entered into the network), rather than separately. Once the loss has been calculated, the Grad-CAM visualization can be done in a similar way to the common image. Sorry that code cannot be shared at this time as a very complex open source approval process is required. Thanks again.