Question about model testing.

[wzmsltw]

Hi @laura-wang , thanks for your interesting work. Since I am new to self-supervised learning, I have some questions about model evaluation.
For "action similarity labeling" and "scene recognition" task, linear SVM is used for classification. But for "action recognition", after extracting C3D feature, how classification is done?
I am looking forward to your reply.

[laura-wang]

@wzmsltw Hi, sorry for the late reply. Quite busy these days. As for the action recognition task, since the datasets UCF101 and HMDB51 are not so small, we don't extract features and then use SVM for classification. Instead, we just retain the conv weights from the pretrained network, and the finetune the whole C3D.

[ShiQingHongYa]

@laura-wang thanks for your interesting work. I find that you use the pretrained network, then finetune the whole C3D,could you please give me the network that is the pretrained?

[ShiQingHongYa]

@laura-wang Thanks for your interesting work. Could you plase privde the code of test

[laura-wang]

@alice0522 Sorry, I am quite busy these days. I'll update it as soon as possible.

[ShiQingHongYa]

@laura-wang Thank you very much !!! Please also privde the code of appearance statistics,your job is very good,thank you !

[Sissuire]

The pretrained model is expected to be released if possible, either with PyTorch or Tensorflow~ @laura-wang