Question about the sampling rate calculation.

[hollowgalaxy]

First question, in the csv with duration and frames. Is the frames number the total number of frames in the video, or is it the number of frames created by the _bn.shape[0]?

Second questions. For the calculation in here. Why is are you flooring the rate as opposed to using ceil for the yc2 dataset?

Third question. Why do (duration/frame)*ceil(0.5*frame/duration), if the model is outputting position in frame numbers.... Then given that units of 0.5 is frames/seconds -> Wouldn't changing the equation to get the correct time to this instead to do 0.5*(duration/frame) meaning you convert to duration, to match the units

[LuoweiZhou]

@hollowgalaxy please see the followings.

1. The total number of frames in the video.
2. Please refer to our feature extraction code here. ANet is processed with an older version of the repo and yc2 is processed with the latest code which had a minor change regarding the sampling approach.
3. We formulate the frame rate to match the frame rate we used in feature extraction.