doc-doc/HQGA

How to conduct dense and sparse sampling ?

Fly2flies opened this issue · 3 comments

Fly2flies commented

Hi, thanks you for sharing such a great work.
I would like to know how to make dense sampling and sparse sampling after uniformly sampling K clip frames.

After sampling K clip frames c1,...,cK

  • take these K frame as the middle frame and sample 16 consecutive frames forward and backward to calculate the motion features ?
  • Or uniformly 32 frames from all frames from c(i-1) to c(i+1) to calculate the motion features ?

Which of the above methods is corresponding to the paper ?

doc-doc commented

Thanks for your interest. We adopt the 1st method and sample 16 frames (8 forward / backward) centered at the key frame.

Fly2flies commented

Thanks for your interest. We adopt the 1st method and sample 16 frames (8 forward / backward) centered at the key frame.

Thank you for your reply. I would also like to know how much memory is needed to store all the TGIF-QA data and how to compress and store the extracted features ?

doc-doc commented

The raw TGIF_full dataset need about 124G. The features are store in .h5 files and need about 40G for each sub-task.