facebookresearch / AVT

Code release for ICCV 2021 paper "Anticipative Video Transformer"
Apache License 2.0
151 stars 28 forks source link

Data preparation (EK) #36

Closed ofir1080 closed 2 years ago

ofir1080 commented 2 years ago

Hi There! As mentioned, for faster processing all videos are resized to 256x256px. However, when running evaluation on the provided ckpt (expts/09_ek55_avt.txt), I got different results on resized/original-size versions. The reported top-{1,5} acc was applied on the original sized vidoes? My results are: original videos' size:

[2022-02-23 18:09:48,322][root][INFO] - acc1/action: 12.369477 [2022-02-23 18:09:48,323][root][INFO] - acc5/action: 29.859437

on the resized videos: [2022-02-23 18:09:56,238][root][INFO] - acc1/action: 11.646586 [2022-02-23 18:09:56,238][root][INFO] - acc5/action: 26.666666

I might have missed something? Thank you very much for your work!

rohitgirdhar commented 2 years ago

Hi @ofir1080, that is strange.. I did all the training and evaluation on the low-resolution clips. Did you use this for resize? Note that I don't resize to 256x256, but to height=256px.

ofir1080 commented 2 years ago

Thanks for your response! I will recheck it. I suspect it is a dataset version issue, as you suggested. Thanks!