3D reconstruction issue

[mpillet]

Hello,

I'm using 2 cameras to reconstruct the 3D kinematics of a human, with the 133 body model keypoints. I'm using my own triangulation to reconstruct the 3D, based on the 2D of the 2 cams. However, I'm facing an issue, depending if the person is facing or facing away the cameras. The body is more forward-oriented in a case rather than in another one. I think that the problem is probably due to the 2D marking, which is probably different depending if the cameras see the front of the body or the back and this small difference in the 2D pose estimation is enough to modify the 3D posew. Did you already face this issue previously ? Otherwise, I'd like to try the triangulation that you implemented to see if it's better but I really don't understand how to use it, could you maybe explain me more in details ?
Another option is that 2 cameras are maybe not enough to have a really precise reconstruction, independantly of the body direction ?

Another question is, does the model estimate the 2D pose only based on the current frame OR it also uses time information (the information of the previous frames) ? If no, is there a way to introduce this ?

Thanks in advance for the help.