I examined your tutorial setups with stereo cameras and I was wondering about what happens in an alternative setup.
Consider a robot working both in indoor and outdoor environments and suppose it cannot be equipped with a stereo camera, but only with RGBD camera. Indoor RTAB-Map works very well fusing visual odometry with map building, but if the robot exit from the room and goes outside, the RGBD obviously does not work for building the map.
1) Is the odometry still valid outdoor since it relies on RGB visual features?
2) Can the loop closures still be detected from visual words even if the map is not built outdoor?
3) When the robot comes back indoor from outdoor, could there be any problem in receiving again a valid depth map and keeping the whole graph consistent?
4) If the robot was equipped only with a stereo camera, how would it perform indoor in comparison with an RGBD camera since the CPU has to compute the disparity map from left/right fields, too?
1) No, the depth is required to estimate the motion.
2) The loop closures can be detected, but they won't be accepted as the motion between the images could not be estimated by the lack of depth.
3) If you are using RGB-D visual odometry, the robot would get lost outside (particularly on a sunny day), stopping the mapping until the robot comes back to where odometry got lost. If you are using robot odometry, the graph will continue to grow without to be able to accept loop closures. When the robot comes back inside and if it detects a loop closure, the graph will be corrected and it will be consistent.
4) With a calibrated stereo camera, the trajectory would look the similar to a RGB-D camera. Computing the disparity doesn't add a large overhead (on my laptop though). However, the 3D cloud of the environment will be more sparse (particularly on surfaces where there is no texture), thus not recommended for reconstruction. The big advantage of a stereo camera over a RGB-D camera is that it will work outdoor.