About extracting features on UCF101, by using C3D model fine­tuned on UCF101

[misunusc]

Question 1: I am confused about the different names of fc layers in the prototxt files, for different models training on different dataset. I follow the user guide to fine­tune C3D model. After that, I want to use the prototxt named "c3d_sport1m_feature_extractor_frm.prototxt", for extract features. But, I see that the names (fc6-1, fc7-1, fc8-1) of fc layers in it, are different from the names (fc6, fc7, fc8) of corresponding fc layers in "c3d_ucf101_finetuning_train.prototxt". Based on my understanding, I should change the names of fc layers in "c3d_sport1m_feature_extractor_frm.prototxt" to be the same as the ones in "c3d_ucf101_finetuning_train.prototxt", right? And then, I should extract features from fc6, instead of fc6-1, right?

Question 2: Except names, I should also change "num_output" to be 101 in "c3d_sport1m_feature_extractor_frm.prototxt", for extracting features from UCF101, right? One more thing is, should I put "use_temporal_jitter: false" in "c3d_sport1m_feature_extractor_frm.prototxt", for extracting features from UCF101? Since "c3d_ucf101_finetuning_train.prototxt" has "use_temporal_jitter: false".

Look forward to your reply. Thank you very much.

[hjjpku]

So, what has happened after you changing the .prototxt file as you described in Question 1? Does it extract the feature correctly?

[longmao-yiran]

I have same question.when I add a layer in c3d1.0 and finetune the network,i think it will be error ,but it he still runs.I think the author's caffemodel doesn't correspond to his protext name. Look forward to your reply. Thank you very much.