Hi @kaykobad ,
Thanks for your great work! I'm currently studying the Kinetics-Sound (KS) dataset processing in your paper and have a question about the dataset setup.
In your paper, you note that KS is a subset of Kinetics400 with 31 classes (14,739 training samples and 2,594 test samples). However, the original KS paper (Arandjelovic & Zisserman, 2017) describes the dataset as containing 34 classes.
I’ve also reviewed the dataset processing details provided in your codebase, but didn’t find explicit mentions of filtering or selecting specific classes from the original 34 to arrive at 31.
Could you kindly provide more detailed processing steps for the Kinetics-Sound dataset?
Thanks in advance for your time and clarification!
Hi @kaykobad ,
Thanks for your great work! I'm currently studying the Kinetics-Sound (KS) dataset processing in your paper and have a question about the dataset setup.
In your paper, you note that KS is a subset of Kinetics400 with 31 classes (14,739 training samples and 2,594 test samples). However, the original KS paper (Arandjelovic & Zisserman, 2017) describes the dataset as containing 34 classes.
I’ve also reviewed the dataset processing details provided in your codebase, but didn’t find explicit mentions of filtering or selecting specific classes from the original 34 to arrive at 31.
Could you kindly provide more detailed processing steps for the Kinetics-Sound dataset?
Thanks in advance for your time and clarification!