Are optical fow compenents rescaled to [0,255] ? about two-stream-pytorch HOT 5 CLOSED

Yes, it is. But that is for training and validation. For testing, we evenly sample 25 frames/optical flow fields from the video, average to get the final result.

from two-stream-pytorch.

Yes, it is.

from two-stream-pytorch.

Hi @bryanyzhu ,

Thank you for your answer.

A question related to the inputs :

For both training and testing you have :

1. 1 single frame per video action extracted at random for the RGB stream
2. 10-frame stacking of optical flow fields (successive) extracted at random for the RGB stream

Was that your setup ?

Thank you

from two-stream-pytorch.

So , we can't extract features at frame level (for temporal stream). Let's say we have a video of 200 frames. and the temporal stream requires an input of of 10*2 (10=number of frames). It returns representation for a stacked optical flow of 10 frames. Is there any way to do that for1 frame ?

Thank you @bryanyzhu

from two-stream-pytorch.

@Benjiou You can still do that if you want. For example, when you want feature for frame I_{1}, you just take I_{1:10} to get the temporal feature. and for frame I_{2}, you take I_{2:11} to get the feature.

In this manner, you will get frame-level feature. It's just there is a lot of overlap, and maybe it doesn't give you much benefit over the clip-level feature.

from two-stream-pytorch.