Re: Loop closure registration strategies - Help

Posted by matlabbe on
URL: http://official-rtab-map-forum.206.s1.nabble.com/Loop-closure-registration-strategies-Help-tp6421p6432.html

Hi,

Reg/Strategy = 1 (Geometry ICP)

Visual transformation is still done for only loop closure transform estimation. As loop closures are detected visually, there is always a first transformation computed using visual features, then ICP is done to refine that transformation. If "Reg/Strategy" = 2 or 0 is set, we can set "RGBD/LoopClosureReextractFeatures" to true to re-extract visual features using parameters in "Vis/" group.

When refining odometry links ("RGBD/NeighborLinkRefining"=true), as odometry has already a guess transform, ICP is done directly without computing a visual one.

Loop Closure Detection and Visual Registration
We can set "Mem/DepthAsMask" and "Vis/DepthAsMask" to false to extract features everywhere independently of the depth image values. If some features don't have depth, structure from motion is done to estimate their depth based on odometry transformation (assuming that odometry is quite accurate!).

I don't know what kind of camera you are using, but if you can use it in stereo mode instead of RGB-D, rtabmap will extract features everywhere by default, and estimate feature depth even if disparity is very small (infinity). For example, it will be able to initialize feature's depth at some kilometers. Those far features are very useful for more accurate orientation estimation for visual odometry, and also for loop closures.

Depth by structure from motion
See above.

cheers,
Mathieu