search

chaneyddtt / Generating-Multiple-Hypotheses-for-3D-Human-Pose-Estimation-with-Mixture-Density-Network

Code for our CVPR2019 paper: Generating Multiple Hypotheses for 3D Human Pose Estimation with Mixture Density Network
MIT License
97 stars 9 forks source link

How to get the unique pose from the generated five hypotheses during testing? #19

Open wyx1997 opened 3 years ago

wyx1997 commented 3 years ago

Hi, thanks for your great idea and work!

I am now working on my own data. I want to get a unique pose from the generated five hypotheses when testing, just like the normal test procedure. However, After carefully reading your paper and code, I have not found the relevant details about it yet. So If I have missed some points?

Or I think maybe it can be realized by summing the weighted gaussian kernels as your paper says, but I wonder that if the five gaussian kernels are independent of each other? If not, how to deal with it?

By the way, my test set does not have 3d labels.

Looking forward to your reply! Thanks a lot!

chaneyddtt commented 3 years ago

Hi, thanks for your interests. There are some options for you select one specific solution. For example, you can select the most consistent one by using multi-view information when multi-view information is available (I used this one in the paper). Or you can select the one that satisfies the smooth constraint when you have video for the test set (I have tried this one previously but does not work very well in my case). You can also try the methods that are described in the original MDN material when you do not have multi-view or video information. image

wyx1997 commented 3 years ago

Thank you very much, I would try these methods.