Re: rtabmap does not find loop closures

Posted by matlabbe on
URL: http://official-rtab-map-forum.206.s1.nabble.com/rtabmap-does-not-find-loop-closures-tp6586p6597.html

Hi,

1) Look at Section 3.5 of this paper:

RTAB-Map integrates three graph optimization approaches:  TORO [Grisetti et al., 2010], g2o [Kummerle et al.,2011] and GTSAM [Dellaert, 2012].  g2o and GTSAM converge faster than TORO, but are less robust to multi-session mapping when multiple independent graphes have to be merged together.  TORO is also lesss ensitive to poorly estimated odometry covariance.  However, for single map, based on empirical data, g2o and GTSAM optimization quality is better than TORO, particularly for 6DoF maps.  GTSAM is slightly more robust to multi-session than g2o, and thus is the strategy now used by default in RTAB-Map contrarily to our previous works using TORO.

RGBD/OptimizeMaxError is also explained in that section. Note that you can do "rtabmap --params"  to have a description of all parameters. g2o/Optimizer is the optimization approach used in g2o, which is [0=Levenberg 1=GaussNewton]. RGBD/LoopClosureReextractFeatures to true can be used to re-do old approach on loop closures, where we re-extracted features. In recent versions, we don't need it as we have already extracted all features we needed to compute loop closure transformation.

2) It depends on the sensor used, but as long as the camera can see the same features, it could theoretically close the loop.

3) Low odomety drift, frequent loop closure detection.

4) Not necessary the whole trajectory again, but at least if you know that you are passing by a nearby already visited location, you can go there, close the loop, then return back exploring.