search

zhengyuf / STED-gaze

Code for paper 'Self-learning Transformations for Improving Gaze and Head Redirection'
GNU General Public License v3.0
89 stars 23 forks source link

Number of training which the result is averaged over #9

Closed san9569 closed 3 years ago

san9569 commented 3 years ago

Hi. I'm very interested in your work.

So I've trained your training code several times. I could confirm that the variation of results exists somewhat. Can I know the number of training done to get the results in your paper?

zhengyuf commented 3 years ago

Hi San,

By variation of results, do you mean the variation of evaluation metrics? If so, how big is the variation? Also, which table are you trying to reproduce?

Cheers, Yufeng

san9569 commented 3 years ago

Hi, Thank you for your quick response.

I was about to reproduce Table 2. For example, in the case of gaze redirection, the difference between the first and second trials is about 0.6. Is the results are reasonable?

Thank you

zhengyuf commented 3 years ago

Hi,

Table 2 in the main paper is only based on 1 training trial. But if you go to table 1 in the supplementary material, you can see that the gaze redirection error is below 2.4 even when the loss weights are not optimal, and we didn't cherry-pick experiment trials. So I think a difference of 0.6 is a bit big (I assume that you are talking about our final method, not the ablation ones in the first three rows of table 2, main paper.)

Cheers, Yufeng