Map cloud rotation drift

[marlu]

Hi community,

I try to build a map with default launch files from RGBD Handheld tutorial. After a short mapping process, the map cloud drifted (as seen in the picture below), probably due to declined loop closures. Is the any recommended procedure or settings to avoid this behaviour? Could have it been caused by too fast camera movement? Besides this, are there any resources for a map post-processing within RTAB-Map? Thank you.

[matlabbe]

With rtabmap-databaseViewer, you can browse data in the database, compare odometry and loop closure links with Constraints View. That can help to see how bad odometry or loop closures can be. Links can also be refined with the same tool.

What kind of camera are you using? The accuracy of tracking will differ depending on the camera used.

[marlu]

Thank you for the reply,

I will go through the data and compare the odometry and loop closures. I use Luxonis Oak-D-Pro camera with Xsens IMU. Comparing this camera to Realsense D435i it seems to perform worse in terms of map quality. Do you have some extended experience with OAK cameras and their optimal settings?

[matlabbe]

If the OAK-D has good calibration and correct rectification (I saw on OAK-D cameras, depending on the calibration, that we can see artifacts in the images, don't know if they fixed that), the pose accuracy in stereo mode (Left-Right) would be similar than realsense stereo mode (Left-Right IR cameras).

For point cloud reconstruction, realsense D435 uses a IR dotted pattern that helps estimating depth on textureless surface, but you cannot use stereo IR at the same time. Using D435 in RGB-D mode has better point cloud reconstruction (because of the dotted projector), but motion estimation will be bad (RGB camera has smaller FOV and has more motion blur than the IR cameras).

I don't have experience with Oak-D-Pro, does it have also an IR dotted projector? I guess what I just said about D435 would also apply to Oak-D-Pro. If OAK-D has better RGB camera than D435, then maybe it could be better (in RGB-D mode). Note that realsense and depthai uses different disparity approaches.

Mathieu