About Dotmask work

[jasonlee]

Sorry to bother you . After I read "Dynamic Object Tracking and Masking for Visual SLAM "
https://arxiv.org/pdf/2008.00072.pdf , I have some question that I want to ask.

The paper used  depth image processing to make a mask on the map .
I want to ask why it turn some specific pixels with depth value to zero , then it won't show feature points on image that showed on RTAB-Map left-up screen .  

What do the points actually mean on image that showed on RTAB-Map left-up screen ?  Aren't they feature points that extract from RGB image ?

As I knew , the visual odometry extract feature points  on RGB image first , then use the points to get depth value with depth image .  Sorry , I am the newer about this part .

Thanks for your reading .
Have a nice day.

[matlabbe]

Hi,

Note sure which left-up screen you are referring, but RTAB-Map uses the parameter Vis/DepthAsMask=true by default, which means that when we do feature extraction (using opencv), we only extract features where there is a valid depth. Dotmask clears depth values on dynamic objects, so implicitly visual odometry won't track features on dynamic objects. On mapping side, the dense point cloud is created from the depth images, so if there is no valid depth, dynamic obstacles won't appear in the resulting point cloud of the map.