search

jeffreyyihuang / two-stream-action-recognition

Using two stream architecture to implement a classic action recognition method on UCF101 dataset
MIT License
855 stars 252 forks source link

Motion resent101 #15

Closed LiliMeng closed 6 years ago

LiliMeng commented 6 years ago

Hi, Thanks a lot for providing the open-source code!

     May I ask how did you get your pre-trained motion ResNet101? It's pre-trained on ImageNet and fine-tuned on UCF101 or the just changing the channels of the model trained on ImageNet? 

      Highly appreciate your time and help!
jeffreyyihuang commented 6 years ago

Hi, Thanks for the good question!! I utilize the cross-modality pretraining technique to transfer the model weight pretrained on ImageNet to fit the motion ResNet101. The details are described in Readme!

Sincerely, Jeffrey

LiliMeng commented 6 years ago

Hi Jeffrey,

       Thanks a lot for your kind and quick reply. The channels in the RGB and optical flow stream are different, how did you fit the pre-trained ImageNet model to the optical flow stream? Do you have the code to load the ImageNet pre-trained model to optical flow stream? 

       Highly appreciate your time and help!
jeffreyyihuang commented 6 years ago

Hi, Please refer the link below to see my implementation details of cross-modality pretraining method!

https://github.com/jeffreyhuang1/two-stream-action-recognition/blob/master/network.py#L216

Jeffrey

LiliMeng commented 6 years ago

Great! Thanks a lot!