Issue with comparing gt to predicted gt
Closed this issue · 3 comments
Thank you again for sharing your code!
I am training your network with synthetic generated data. Therefore I do the following - I have a fix camera looking at a scene where I generate objects with a random pose. Now I write data in the form of the linemod dataset but instead of the camera rotation and translation as "cam_R_m2c" and "cam_t_m2c" I write down the translation and rotation of the regarding object.
The network now is training and I am getting an average distance of around 0.0105252 (whats ok I think). But when I print out the predicted rotation and translation I get totally wrong values compared to the groundtrouth of my objects (see screanshot).
I think this is because I give the network the rotation and translation of the objects and not the camera. Do I need to change my data generation or is there a simple calulation step at the end so I get the right prediction of the rotation and translation? I hope thats not a stupid question ^^
Thanks in advance!
Hi, it would be better if you could spend some time to find a more academic answer from other resources. Basically, you need to make sure that the GT pose is corresponding to the camera frame.