Direct odometry methods in RTABMAP

[JudithSU]

Hi Mathieu,

First, thank you for building this great application, we have been able to learn a lot about SLAM while using RTABMAP.
We are currently researching the differences between indirect and direct visual odometry methods (such as DSO). I noticed the visual odometry integrations in RTABMAP are primarily indirect methods. Would it be possible to use a direct method such as DSO as odometry input in RTABMAP? It would be interesting to directly compare the performance of direct and indirect methods this way.
What would it take to realize this, or are there any plans for this in the future?

Again thank you for the time and effort you put into this project!

Regards,
Judith

[matlabbe]

Hi Judith,

If you can modify https://github.com/JakobEngel/dso_ros to get the output pose in nav_msgs/Odometry format, you could feed it to rtabmap node instead of using rtabmap's odometry nodes. However, looking at dso_ros, it seems a MONOCULAR odometry approach, meaning that the scale is unknown. It won't work with rtabmap is the scale is unknown. rtabmap requires stereo, VIO or RGB-D odometry approaches with scale known to work.

The only way to compare both approaches would be to record a rosbag or a dataset, then process it with each approaches separately, getting the trajectory for each approaches. Then, you would need to scale the trajectory of DSO to match the one of rtabmap to better compare them.

cheers,
Mathieu