about the depth data error

[mantouRobot]

Hi~
      I use a RGBD camera and it's error is like this: 1m->no error, 2m->0.2m offset, 3m->0.4m offset. I think this precision can't fusion object well. What's u opinion?

[matlabbe]

Hi,

yes indeed, values are less precise farther from the sensor. Generally, I set the maximum point cloud range to 3-4 meters for sensors like Kinect. RTAB-Map is not meant for very precise object reconstruction like Signed Distance Functions (SDFs) based approaches (e.g., Kinect fusion, Reconstruct me), but for large scale mapping. RTAB-Map relies on basic approaches from PCL to to filter/assemble the point clouds.

cheers,
Mathieu

[mantouRobot]

Hi~
      In my understand, the graph optimization can optimize the odometry offset(namely optimize x(k)=f(x(k-1), u, w) of noise w), but can't optimize the measurement noise(namely z(k)=g(x(k), v) of noise v). So this approach can't fusion object very well. What's u opinion?
      Thank u so much for discussion~
       

[matlabbe]

Yes, for the graph optimization part, only the odometry links are optimized on loop closure (visual words are not used like it would in a SBA approach). This is because RTAB-Map is independent of the odometry source, which could be wheel odometry for example. There is a sparse bundle adjustment (SBA) option that can be done in post-processing, taking the visual word correspondences for more constraints to optimize. Note that I am currently improving the SBA code (not yet committed), and it works a lot better than before.

cheers,
Mathieu

[mantouRobot]

Wow, so expect it~
      Thanks so much~