clip_timestamp in calculating FVD about common_metrics_on_video_quality HOT 5 CLOSED

in my project and most paper experience, you should sample a clip during preprocessing (in your customized video dataset get_item function.

like in mcvd

https://github.com/voletiv/mcvd-pytorch/blob/master/datasets/ucf101.py#L53-L86

in this file

time_idx is the first frame id which means the video clip will start from here
self.frames_per_sample is the sample video length

from common_metrics_on_video_quality.

thanks

from common_metrics_on_video_quality.

hello

1. you mean clip_timestamp is only used for video1 not video2? here we make sure video1 and video2 have same shape, so videos should have same length. clip_timestamp will take the first 10 frames 11 frames .... len(video1) frames into account.

2. if you want to test on video dataset, as i known, all length should be equal. it video length is not equal, you should random sample to get a clip for each video which has the same length.

if you have other qusetion, please reply me again

from common_metrics_on_video_quality.

Thank you for your reply.
That means if there are many videos and each video has a different length , should I need to randomly sample a continuous frame from all the videos during preprocessing? Or just take the first n frames?(n = 10 or 11 ...

from common_metrics_on_video_quality.

have a good research

from common_metrics_on_video_quality.