search

VIS4ROB-lab / ccm_slam

CCM-SLAM: Robust and Efficient Centralized Collaborative Monocular SLAM for Robotic Teams
544 stars 106 forks source link

Camera Pose in world reference #49

Closed DanielRoeder1 closed 10 months ago

DanielRoeder1 commented 2 years ago

I am trying to perform 3D reconstruction using the depth maps of the TUM datasets in conjunction with the poses of CCMSLAM. I know that the camera poses of CCMSLAM are not scaled correctly but I take care of this separately.

Now to perform this reconstruction correctly from my understanding I would need the Twc (camera pose in world reference) to transform the point clouds resulting from the depth maps. But how would I get the camera pose so that it aligns correctly with the typical xyz coordinate frame?

Currently, the poses are coherent to a somewhat twisted coordinate frame and then transformed by this static transformer: https://github.com/VIS4ROB-lab/ccm_slam/issues/44

I found this issue in the ORBSLAM2 repo describing how to correctly visualize the poses: https://github.com/raulmur/ORB_SLAM2/issues/832

While the approach from the ORBSLAM2 repo seems to help the poses are still not fully aligned.

Could someone give a hint about the transformation that is needed or some explanation on the transformation between the ORB frame and world frame?

patriksc commented 2 years ago

H Daniel, is this still a problem, or did you already solve it (sorry for the late reply here)? If yes, could you maybe share a bit more information where you are stuck, I'm not sure whether I can exactly follow what actual problem is here. Particularly, is your dense 3D reconstruction from depth data globally consistent, but just shifted/twisted w.r.t. the sparse 3D points? Or are the dense points clouds correct withing a single camera frame, but not correctly aligned, so that the 3D reconstruction itself is inconsistent? If you could maybe also share an image, that would probably also help.

patriksc commented 10 months ago

Closed - inactive