Can you offer some samples about manual attribute?

[fjchange]

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

[GayatriPurandharT]

I think UCF101 dataset come along with manual attribute description. I'm not sure of the format they used, in order to be used as semantic embeddings. I would like to know about it.

[devraj89]

I believe UCF101 is available here: http://www.thumos.info/download.html
Regarding the attributes for the Olympic dataset, I think it was in some pdf file.
https://github.com/gzoumpourlis/olympic_sports_tools
The pdf is linked there.

[fjchange]

Thanks, @devraj89 , there are the manual attributes, but what is word2vec attributes?

[GayatriPurandharT]

Is it possible to define manual attributes for the hmdb51 dataset? if yes, can you describe how?
Thank you very much.

[devraj89]

@fjchange the word2vec attributes are generated by sending the category names through the word2vec model.
@GayatriPurandharT I really do not know how the manual attributes were even constructed for the other two datasets. I just used the ones that were provided to us. Maybe you can ask the original authors regarding how they generated the attributes ? This might itself be a research topic.

[GayatriPurandharT]

@devraj89 Thanks for your reply. I downloaded attributes provided by UCF101. Each action has got attributes with truth-value 0/1. Can I add a new action of my own? I would want to know how the video features and attributes have been merged into data_action/ucf101/split_1/att_splits.mat can. So that I train my own videos with new actions. Thanks

[devraj89]

@GayatriPurandharT regarding adding new action classes I am not sure. You have to figure out what the attributes should be. As I mentioned previously, contact the original authors of UCF101 dataset. They might be able to tell you on this more.

att_splits.mat --> for this please go through this paper : https://arxiv.org/abs/1712.00981 and the code: https://www.mpi-inf.mpg.de/departments/computer-vision-and-machine-learning/research/zero-shot-learning/feature-generating-networks-for-zero-shot-learning/ to understand how the .mat file was created.

To generate those files I had to write a separate code which I dont have access to right now.

[GayatriPurandharT]

@devraj89 Thanks a lot for the information. I will go through the code. Also, it would be really helpful if you publish the code to create .mat files for video files and attributes.

[prashastbindal]

@GayatriPurandharT : Were you able to create attributes file for UCF 101 ? If yes, can you please share it ? Please also share the code to create the files if you have it.