Mappoint / pose relation not correct? (rviz)

[flo981]

Hey,

it seems like there is a error corresponding map points/pose? I don't understand why there are so many points directly at the the current pose position. Im using a stereo camera setup so the depth information should be provided. See the pictures for the current camera frame and the visualization from 2 different potions.

[lennarthaller]

Hey, this looks like DepthMapFactor is not set correctly, meaning the scale of the points is different from the scale of the tf frame.

[flo981]

Thanks for your fast reply!:) Where does this factor come from? Camera calibration? Its not existing in my .ymal file. And what should it be?

[flo981]

DepthMapFactor is a scale factor that multiplies the input depthmap (if needed) if you are using a RGB-D camera. This is used in the TUM RGB-D dataset. (from https://github.com/raulmur/ORB_SLAM2/issues/89#issuecomment-220551549)

But I am using a stereo, and not RGB-D camera...

[flo981]

In Tracking.cc :

    if(sensor==System::STEREO || sensor==System::RGBD)
    {
        mThDepth = mbf*(float)fSettings["ThDepth"]/fx;
        cout << endl << "Depth Threshold (Close/Far Points): " << mThDepth << endl;
    }

    if(sensor==System::RGBD)
    {
        mDepthMapFactor = fSettings["DepthMapFactor"];
        if(fabs(mDepthMapFactor)<1e-5)
            mDepthMapFactor=1;
        else
            mDepthMapFactor = 1.0f/mDepthMapFactor;
    }

https://github.com/appliedAI-Initiative/orb_slam_2_ros/blob/e978a2c1d1db6abc93c9a7e0c5bbca01739d8356/orb_slam2/src/Tracking.cc#L136

[lennarthaller]

I just went through the code, you are right, in Tracker.cc, where the parameters are read, the setting is only used for RGBD cams. I have to say that I work with mono or RGBD cams most of the time and haven't really used the stereo node.

[lennarthaller]

The algorithm treats the RGBD and Stereo information the same most of the time, if you look in the paper they bring the stereo information in the same format as the RGBD image and use the same pipeline.

[lennarthaller]

So for me it still looks like its some scaling issue. Maybe you can try to apply the DepthMapFactor to the stereo image?

[flo981]

I tried changing ThDepth parameter and also adding the DepthMapFactor inside the .ymal file. The only thing changes, the more I decrease ThDepth the less features a extracted. Adding DepthMapFactor doesnt do anything, but this also visible in the code I posted above (for stereo)

But this seems to be an issue of ORBSLAM2 itself, thus I should ask Raulmur?

[flo981]

I'm also thinking of using a RGBD camera instead now... where I wont have this issue right?

[flo981]

But the original ORBSLAM2 implementation doesn't hast this issue. At least not in the map viewer...

[flo981]

And when I publish the map points and pose there, it looks also right:

[lennarthaller]

I'm also thinking of using a RGBD camera instead now... where I wont have this issue right?

At least we never encountered the issue there ;)

[lennarthaller]

And when I publish the map points and pose there, it looks also right:

Interesting, thanks you for investigating. Probably there is a bug in the stereo code of the node then. Like I said, we usually use the mono or RGBD nodes so the Stereo stuff is much less tested.

If you want to investigate further and fix the bug I am happy to merge your fix!

[flo981]

I'll take a look at it but I think this is beyond my skills... Is it possible you take a look as well? It would be crucial for my masters thesis...

[lennarthaller]

I am pretty busy atm, I wont be able to investigate that in the next couple of weeks.

[raclab]

Is there any detailed description about the usage of the package with ros? For example, what did I should do for watching of the point cloud on rviz? Thank you