Evaluation on HumanAct12 unclear

[anthony-mendil]

First of all, thanks for your great work.
In table 3 of the respective Paper you report evaluation metrics on the HumanAct12 dataset.
I would like to compare my model results with yours, but that requires to know some aspects not mentioned in the paper.

• what weights for the different loss components produced your results?
• which pose representation are the results based on and was translation included?

Kind regards, Anthony Mendil.

[sigal-raab]

Thank you @anthony-mendil for your interest.
For the first variation, I used --lambda_rcxyz 0 --lambda_vel 0 --lambda_fc 1.
For the second variation, I used --lambda_rcxyz 1 --lambda_vel 1 --lambda_fc 0.
For your convenience, when downloading the pre-trained models, you can find the value of all the arguments in the file args.json.

[anthony-mendil]

Thanks for the info!
Is there a reason why you do not use all three together?

[sigal-raab]

You should use all three together.
While geometric losses enhance the quality of the results, they marginally compromise quantitative metrics. As such, they are not used simultaneously in our tables, in pursuit of metric superiority.