search

NVlabs / PoseCNN-PyTorch

PyTorch implementation of the PoseCNN framework
Other
186 stars 47 forks source link

The noise in ground truth dataset #14

Open alanxuefei opened 3 years ago

alanxuefei commented 3 years ago

According to the paper, the depth video is used to obtain ground truth for avoiding manual labeling. I guess that it could be scan matching and global optimization.

I tried to reproduce the scenario in a 3D visualization tool (Rviz2) using ground truth data. 1, Shift the camera axes using rotation_translation_matrix. 2, Re-project 3D point cloud into camera axes.

Peek 2021-05-27 10-56

The items in the 3D world jump at the centimeter level. The length of axes in the simulator is 1 meter. Blue is the Z axe. In theory, ground truth may also have noise. Is it caused by the noise from the ground truth or my re-projection method?

liguanglin commented 2 years ago

I find the given ground truth of camera pose have some errors with the true ground truth. Try to use object pose ground truth to get camera pose.