about B_in_camera about iros20-6d-pose-tracking

[wenbowen123]

1. this looks correct to me, if your pkl saves the pose of B in camera (z axis forward, y axis down).
2. it looks like your generated A and B pair has too large motion. What's your setting about max_rotation?

[zhuyazhi]

thank you~

1. yes, we use the opencv camera.

2. I set the same max_rotation as yours。strangely, i found that the pose I extract from my .pkl as B_in_cam is not the same as the B_in_cam saved in .npz after running produce_train_pair_data.py.
   Does there exist some transformation before saving the pose B_in_cam I extract from my .pkl into .npz?

[wenbowen123]

The B_in_cam should be same. The original produce_train_pair_data.py may not change it. You can check where that change happens.

[zhuyazhi]

ok, thank you, i will check it.
thanks again for your kindly reply.

[zhuyazhi]

hello, wenbo，I want to know if the speed 90 fps in the paper includes the time of rendering?
I think the whole process includes render *A with the last frame's pose, crop two images as input, propagate to the se3-tracknet and get the relative pose.

[wenbowen123]

Yes, it includes all the steps.

[zhuyazhi]

thank you～

[zhuyazhi]

when doing inference， I think using the model textured_simple（1w+ faces）for rendering is suitable, is it right?
because the textured.ply includes 50w + faces, and it's time consuming when use this model to render at the last frame's pose, which can not achieve the speed reported in the paper。

[wenbowen123]

Yes, that could affect the computation time. Besides, the image to render is small (176x176) and the whole process was optimized in C++. It also depends on what hardware you are using.