search

yabufarha / ms-tcn

Other
214 stars 58 forks source link

How to get features? #29

Closed chenluochuan closed 3 years ago

chenluochuan commented 3 years ago

I am trying to train the network with other dataset. But how can i extract frame-wise features from an example include multiple actions by using I3D network.

chenyr0021 commented 3 years ago

I meet the same problem. I have the pre-trained i3d model, but the features I extracted on my own are in 1024xT dimension. While the features I download from the website are 2048xT. I dont know if there is anyway to increase the channels.

chenluochuan commented 3 years ago

I meet the same problem. I have the pre-trained i3d model, but the features I extracted on my own are in 1024xT dimension. While the features I download from the website are 2048xT. I dont know if there is anyway to increase the channels.

Could you please tell me which dataset you used to pre-train i3d model? isolated action dataset or continuous action dataset?

yabufarha commented 3 years ago

Hi @chenluochuan, Hi @chenyr0021

we used I3D pre-trained on Kinetics to extract the features. Here is the code we used: https://github.com/ahsaniqbal/Kinetics-FeatureExtractor

@chenyr0021 the 2048 dimensional features are obtained by concatenating two 1024 features vectors from the I3D: one form the RGB stream and the second one from the flow stream.

chenyr0021 commented 3 years ago

Hi @chenluochuan, Hi @chenyr0021

we used I3D pre-trained on Kinetics to extract the features. Here is the code we used: https://github.com/ahsaniqbal/Kinetics-FeatureExtractor

@chenyr0021 the 2048 dimensional features are obtained by concatenating two 1024 features vectors from the I3D: one form the RGB stream and the second one from the flow stream.

I see. Thanks a lot!

chenluochuan commented 3 years ago

Hi @chenluochuan, Hi @chenyr0021

we used I3D pre-trained on Kinetics to extract the features. Here is the code we used: https://github.com/ahsaniqbal/Kinetics-FeatureExtractor

@chenyr0021 the 2048 dimensional features are obtained by concatenating two 1024 features vectors from the I3D: one form the RGB stream and the second one from the flow stream.

well, thank you very much!