Not Enough Correspondences about rtabmap

[matlabbe]

Because the visual features are far (>3-4 meters), try 3D -> 2D motion estimation like in the screenshot below. If this doesn't help, set the Logger level to debug (Preferences->Logging) and copy/paste the output here.

[omair18]

Thank you for the help. I have made the above mentioned changes but i'm still stuck with the same issue. I have tried changing odometry strategy and visual word type to see what changes it does with output. Log file for Rtabmap with SURF features and Optical flow is here. Log file with SIFT & BOW is here.

[matlabbe]

It looks like there is a problem with the calibration. Around 1200 features are extracted by the odometry but only ~100 can be triangulated in 3D. In these 100, maybe only 10 would be inliers so not enough to correctly find a transformation between the frames.

Can you do a small recording? Click Tools->"Data recorder...", start a new session, then on next "Stop", the recorded database will be saved.

cheers

[omair18]

I have done a small recording for 50-55 images. You can see the file here.
output.db.zip

[matlabbe]

Did you follow this tutorial to calibrate the cameras?

The stereo rectification is wrong, see this image (I opened the database with rtabmap-databaseViewer):

These are the left and right images superposed. Normally it should only have a disparity on x-axis, not both axes.

Example of good stereo rectification:

Note that you can also use the Calibrate tool in RTAB-Map (Preferences->Source->Calibrate) to calibrate the cameras (it is mostly the same interface as the ROS tutorial). There is an example here (you can do it with the live images of the camera, no need to use a video file).

cheers