naraysa / gzsl-od Goto Github PK

Hello @naraysa, it would be really helpful if you provide a demo code for gzsl-od case. Thanks a lot.

I want to know that the provided class attributes in the code are the human annotated attributes or word vector that described in the paper.

Hello, I would like to know the division of the type of actions you considered for seen and the actions under unseen categories to produce the results mentioned in the paper on both HMDB51 and UCF101 datasets. Will the choice of actions under the seen and unseen category affect the result?
Please give some insight on this
Thank you.

I tried reproducing the UCF101 results under zero-shot setting (not generalized) with the .mat files shared in the repo. If I am not wrong the key "att" in the .mat file corresponds to manual attributes and not word2vec vector because the key "original_att" contains one-hot vector (hence I am assuming it's attributes). But I am unable to reproduce the result of 38% accuracy from the paper. I am getting on average 25% accuracy. Is it possible that the .mat file for UCF contains word2vec representation and not the manual attributes?

while training the model, I get unseen class accuracy of 0.00 right from the first epoch until the end. Though loss values are reducing from 6 to 3 (approx). There is no change in training accuracy. what could be the reason?
Thanks

In the paper, experiments were conducted on the 30 random splits but provided splits are only ten.

Why are the only ten splits provided?

I am wondering how you annote mannual attribute on the video, can you give me some hints?