Settings for indoor/outdoor mapping with minimal memory usage

[antoinemacia]

Hello! Thanks again for this amazing library - I have a couple questions regarding tuning the app for indoor/outdoor mapping with minimal memory usage and decent point accuracy

• In data recording mode - is there a mean to keep some visual feedback of previously scanned areas (even persisting points) beyond the localisation path? I understand this mode is already configured to maximise on device capture time, but Im wondering if there are some parameters we can tweak to still have visual feedback while aiming for long sessions

• More broadly, are there some specific parameters you would recommend for ~15/20 minutes recording sessions of indoor and outdoor scans, with a decent point cloud accuracy for measurement purposes;

Thanks!

[matlabbe]

Hi,

For your first question, you may just start for the default settings, then disable manually what you don't need and also reduce significantly the point cloud resolution (Settings->Point Cloud Density) during live feedback (select point cloud view in main UI view option). If you don't care about loop closure detection, you can disable it by setting "Mapping->Max Features Extracted (Vocabulary)" to Disabled. After that, the performance would be very close to the Data Recorder mode.

For your second question, you can use any rendering options, the raw data is always saved in the database. This means that even if the live feedback uses very low point cloud density, it is possible to export the point cloud at full density later. For the point cloud accuracy, I recommend to use "Depth Confidence" set to "High", though outdoor the range may appear short. You may go Medium for outdoor. See this post for examples. Note that "Depth Confidence" is the only parameter impacting actual data saved in the database, so it cannot be changed later on unlike the density one.

cheers,
Mathieu

[antoinemacia]

Thanks @matlabbe, the post and explanations are very helpful;

I just have one more follow up question - when you mention

If you don't care about loop closure detection, you can disable it by setting "Mapping->Max Features Extracted (Vocabulary)" to Disabled.

Should we care if we're choosing to re-process on server? Are there any advantages of doing loop closure detection on device for stability, especially if we're aiming to do single long running sessions over multi sessions?

Could be completely unrelated but we had issues with a room being merged into another one mimicking the data recording settings, and I wonder the lack of loop closure detection would have an effect (re-processing the scan via the CLI did not fix the merged room issue)

For this case specifically, the outputted DB is available here and I ran the command bellow, though I am happy to open a separate issue for this

docker run -it --rm -v ~/Downloads/testing:/testing \                                                                                                                             9m 52s
   --workdir /testing --user $(id -u):$(id -g)\ 
   introlab3it/rtabmap:focal \
   rtabmap-reprocess \
   --Rtabmap/DetectionRate 1 \
   --Mem/RehearsalSimilarity 0.3 \
   --Kp/MaxFeatures 400 \
   --Rtabmap/MaxRetrieved 2\
   --RGBD/MaxLocalRetrieved  2\
   --Mem/MapLabelsAdded true \
   --Rtabmap/MemoryThr 0 \
   --Mem/STMSize 10 \
   --RGBD/ProximityBySpace true \
   --RGBD/LinearUpdate 0.05 \
   --RGBD/AngularUpdate 0.05 \
      "test1.db" \
      corrected.db

Thanks again!

[matlabbe]

You can re-enable loop closure detection on the server side, when using rtabmap-reprocess. The advantage to do it online is that you have a feedback when you are closing a loop, otherwise you would need to cross your fingers (get on your knees and pray) that in the offline reprocessing step the loop closures will be detected. Well, if you have the experience about which kind of view loop closures are usually detected, it is possible to do it blindly but the person scanning should be aware of it to scan to maximize these loop closures offline later. Note that even if we know that we did exactly the same point of view, some loop closures may be missed or rejected offline, in those cases it is possible to "recover" them by manually adding them with rtabmap-databaseViewer (with Constraints View panel).

Online loop closure detection will however need some CPU and RAM usage to work (+ graph optimizer time), so there is a trade-off. Other option is to use markers instead and disable visual loop closure detection.

I don't know if you saw that video: https://www.youtube.com/watch?v=71eRxTc1DaU, but in it we can see how often I look for loop closures, almost always when I move back to an area I previously scan, I make sure I get a loop closure before scanning another room. I think in that example I've made 4 sessions blindly, in the sense that I had loop closure feedback internally of each session, but between the sessions it was a completely different scan, but you can see that I move in a way that I "know" which point of view to take from a previous session so that offline I can make sure the 4 sessions can be connected automatically.

For your results, there is indeed a lack of loop closures. For example, having closing the loop around these areas in red circles would have helped to better globally optimize the map:

As a rule of thumb, when scanning a room, always return to initial position with same point of view to close the loop, then move to another room and do the same. This will make sure that individual rooms are geometrically correct. Then when you move close to a previous room or place that you previously scan, just scan again the entrance to get more loop closures, this will help to get constraints between all rooms, so better globally optimized map.

[antoinemacia]

@matlabbe thank you for looking into it in depth - we're still learning about the mapping process so these insights are extremely useful to us

I think this is plenty for us to make decision on UX and get us going, thanks again!