search

akanazawa / hmr

Project page for End-to-end Recovery of Human Shape and Pose
Other
1.55k stars 392 forks source link

When I trained the model with my own 3d motion data, the logged images upside down. #67

Closed Superlee506 closed 5 years ago

Superlee506 commented 5 years ago

I get the g3d using motion capture, and I'm sure the orientation of Y-axis is the same with SMPL, but when I trained the model with my own dataset, the logged images is becoming worse and upside down. I'm not sure if it's related to the Issue 50. Any suggestion would be appreciated. Thanks.

Superlee506 commented 5 years ago

After training.png individualimage At the beginning of the train At the beginning of the train

MatthewD1993 commented 5 years ago

Yes, it's related to the issue #50.

akanazawa commented 5 years ago

I'm not too sure if it's related to that issue, that definitely can contribute, and you can try to rectify the data as I describe there, but even with that one parameter being upside down it works bc everything else points to the right orientation.

I think the issue is that the you shouldn't align the y-axis with SMPL, but make sure that they are aligned with the image coordinate system, so head Y < toe Y. Because the projection we use is a simple scaled orthographic model. Make sure that when you project the 3D joints with a very simple camera like [1, 0, 0] (1 scale, 0 translation), it looks up right in the image.