search

pha123661 / SA-DVAE

[ECCV 2024] The official repo for "SA-DVAE: Improving Zero-Shot Skeleton-Based Action Recognition by Disentangled Variational Autoencoders"
MIT License
19 stars 2 forks source link

Text features generation and further operation #5

Open ZeraPeng opened 2 weeks ago

ZeraPeng commented 2 weeks ago

Thanks a lot for sharing your excellent work! I have four questions:

  1. I noticed there are two description files: "ntu60_llm.txt" and "ntu60.csv." I’ve confirmed that "ntu60_llm.txt" is from SMIE's work. Could you share the method or source for generating "ntu60.csv"? What is the purpose of this additional file?
  2. In the "text_feats" dictionary, what do the labels "ad," "lb," and "md" signify for the npy files?
  3. When I attempted to regenerate the encoded file "resources/text_feats/ntu60/clip-vit-b-16/ad_60.npy," the original data size was (60,768), but the encoded data with "clip-vit-b-16" had a size of (60,512). Could there be any mistakes on my end? Are there any post-processing steps, such as expanding dimensions by repeating data?
  4. Could you provide the files for generating different splits? Thank you very much for your time!