About reported performance on MPI-INF-3DHP in paper

[syguan96]

About reported performance on MPI-INF-3DHP in paper

[syguan96]

You should carefully tune loss weights and learning rate of BOA on 3DHP. During experiment, we also found that in-the-wild videos and 3DPW shares similar hyper parameters, but 3DHP not. The reason maybe that the labeled 3D keypoints and 2D keypoints are not so accurate, and the posture of 3DHP is much more hard than 3DPW. Hope this helps.

…

2023年2月23日 19:26，Daniel Jung ***@***.***> 写道： In the Table 4 of BOA paper, it is reported that BOA achieves PA-MPJPE of 77.4. However, when I run the MPI-INF-3DHP code from this repository, I get PA-MPJPE of 92.61. Is there reason for such difference in performance between main paper and this code? <https://user-images.githubusercontent.com/56655511/220892976-43f36e78-77e4-443f-9ff9-e97b930d5d1a.png> — Reply to this email directly, view it on GitHub <#7>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ATO3YWZKRBFYJQU5OEWR5MLWY5CMXANCNFSM6AAAAAAVFQDA7I>. You are receiving this because you are subscribed to this thread.

[dqj5182]

Is it possible I can get loss weights and learning rates of BOA on 3DHP (reported as in main paper)? It seems that there are too many hyperparameters for me to change in the system.

Thanks.

[syguan96]

I accidentally deleted it… sorry… One experiment only takes few minute. You can use grid search as I did. I remembered that betas of Adam also was tuned.From my phone在 2023年2月23日，20:54，Daniel Jung ***@***.***> 写道： Is it possible I can get loss weights and learning rates of BOA on 3DHP (reported as in main paper)? It seems that there are too many hyperparameters for me to change in the system. Thanks. —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you commented.Message ID: ***@***.***>

[dqj5182]

Oh I see. No problem! I'll do the search on my own