Testing Localization.

[tejasps28]

Hi Mason, Thanks for sharing this work.
I was able to generate map with this using the ROS wrapper(roman_ros) on KITTI. Can you share information on how to test global localization on the map which is generated?

[mbpeterson70]

Hi @tejasps28, setting up the global localization pipeline should be fairly straightforward given 1 or more roman map files generated from the mapping ROS node.

You'll basically want to run the demo but provide the pre-mapped roman map files. This is a little hacky right now, but you'll need to create a new params directory, and to start with all you should need is a data.yaml file. It should only need a list of runs:

runs: [<roman_map_prefix1>, <roman_map_prefix2>]

You'll then need to create an output directory, with a map directory inside of it. Inside the map directory, you should have your roman map files from running ROS: <roman_map_prefix1>.pkl, etc.

Finally, to get meaningful results you should provide ground truth through a pose_gt.yaml file in your params directory. This will load pose data using this from_yaml function from robotdatapy here.

Then you should be able to run the demo using the -p argument to point it to your params directory and -o argument to point it to your output directory.

I just added a bit more documentation here that should also hopefully be helpful.

[AdwaithChegu]

Hey @mbpeterson70, @tejasps28
I used the ROS wrapper and successfully generated a .pkl file. How can I obtain the ground truth data corresponding to this file?

[mbpeterson70]

Hi @AdwaithChegu, did you use the subset of the Kimera-Multi data? Or what was the original data you used the ROS wrapper with?

[AdwaithChegu]

Hi @mbpeterson70 ,
I am using the RealSense camera with the RomanROS instead of a subset of the Kimera-Multi data.

[mbpeterson70]

@AdwaithChegu cool! So if I understand right, you're looking to get ground truth pose/trajectory information for some real-time data?

Unfortunately, obtaining ground truth for evaluation is out of the scope of ROMAN. Often localization experiments will get ground-truth from GPS for outdoor experiments, or from a motion capture system indoors. There are certain things you can do to get closer to ground-truth, like using fiducial markers with known locations. Of all of these options, GPS is probably the easiest/cheapest option.

Let me know if I misunderstood your question!

[AdwaithChegu]

@mbpeterson70 Thank you for your reply. I want to test the localization in real time (with out using any bag files)on that map which is generated. How can i achive this ?

[mbpeterson70]

Hi @AdwaithChegu, unfortunately determining ground truth data is out of the scope of this project. Additionally, online loop closures is not currently supported, but we plan to release this capability in coming months.

However, one way to qualitatively test the results is to take the saved map files and run the offline submap alignment and to qualitatively check the offline SLAM results. For instance, even without ground truth localization, you can qualitatively tell that the following trajectory estimates above look much more consistent than the trajectory estimate on the bottom that gets only few loop closures:

I mentioned how to take a ros-generated map file and test localization earlier:

You'll basically want to run the demo but provide the pre-mapped roman map files. This is a little hacky right now, but you'll need to create a new params directory, and to start with all you should need is a data.yaml file. It should only need a list of runs:

runs: [<roman_map_prefix1>, <roman_map_prefix2>]

You'll then need to create an output directory, with a map directory inside of it. Inside the map directory, you should have your roman map files from running ROS: <roman_map_prefix1>.pkl, etc.

[notu97]

Hi Mason, Thanks for sharing this work. I was able to generate map with this using the ROS wrapper(roman_ros) on KITTI. Can you share information on how to test global localization on the map which is generated?

@tejasps28 Do you have the steps as to how did you run the KITTI dataset. Did you use kitti2bag to generate the Rosbags for KITTI?