Quality of Output

[cdor]

Hi,

I'm using RealSense D435 + T265 with RTAB-MAP standalone on Windows. I am scanning a small room and the quality of  the results is not very good (quiet bad I would say, not being able to recognize what's in the room and from what I've seen the results should be better than what I'm getting). I wanted to know if someone can have a look and let me know if I have maybe poor settings for the export, if the results could be due to a bad scanning method or what else could be causing this.
I'm including screenshots of the exported mesh and my export settings

Here is a link to the database: https://drive.google.com/file/d/1K7Lx6luLb_b0cc9a_T2K7yEfIVs1DqRf/view?usp=sharing

Thank you!

[matlabbe]

Based on the raw cloud in your database, I am surprised that the resulting mesh is as good.



Before trying to estimate a mesh from the point cloud, we should make sure that the point cloud is good. We can easily see here that the walls seem almost 1 meter thick! The cloud above is not good enough. I see either a lot of drifting from T265 or the static transform between the D400 and the T265 is wrong. Note also that D400 is a not a TOF or LiDAR camera, the resulting point cloud will always be a little wavy on textureless surfaces. You can also limit the range of the depth image under 3 meters (D400 is quite noisy over 3 meters).

[cdor]

Hi,

Thank you for your answer.

Can I share with you another database? (not sure why in the previous one the walls were that thick). I tried to scan that room multiple times and I still don't think I'm getting as good an output mesh as I should.
I've also checked and I'm using the same mount as the one that the settings point to when setting the transform between the cameras.
I'm including a link to the database, I would really appreciate if you can tell me what else could be wrong or if this should be the expected result for this type of room?
https://drive.google.com/file/d/1K6q8xPREErfUYtpE93B2NC-MSF2WavRv/view?usp=sharing

Thank you very much.

[matlabbe]

Are you using T265 pose estimation? It seems to drift a lot in that room. The textureless walls could explain that drift. Here is an example comparing the D435 point clouds on right between two poses separated by only 5 seconds:


We can already see a large drift. This explains why the walls are so thick. As the point cloud is not created from a TOF/LiDAR camera, we cannot really refine the T265 odometry with geometry. This is kind of the limit of that hardware. Try in a room with more visual textures, the T265 should work better, and the clouds may be less noisy too.