matlabbe

Would need to implement a stereo disparity approach working with not-rectified images. [Current implemented approaches](https://github.com/introlab/rtabmap/blob/64f79813cd2c3a7e3f916a7ef2637e93f1f13a5f/corelib/src/util2d.cpp#L122-L736) try to find the corresponding pixel of a feature in the right image assuming it...

Set large variance for z, roll and pitch, then the function will work properly (at least for the minimum variation, the default). Note also that these functions are generally used...

If you use default `RGBD/OptimizationMaxError`, it should be already able to reject those wrong loop closures, unless your odometry covariance is too large. How is odometry computed?

The pose covariance would increase out of bound, and twist covariance would be ~constant. The propagated covariance is indirectly done during graph optimization, where each link contains twist covariance. For...

You can look for `opt_poses` in Admin table to get already optimized graph, for convenience using `DBDriver` class: https://github.com/introlab/rtabmap/blob/64f79813cd2c3a7e3f916a7ef2637e93f1f13a5f/corelib/include/rtabmap/core/DBDriver.h#L108

Do you mean the oocupancy grid? it is also in Admin table (see [here](https://github.com/introlab/rtabmap/blob/e2f037189a6963296dc3752ee51437d0ea5676dc/corelib/src/resources/DatabaseSchema.sql.in#L130C2-L133)), though only created automatically on ROS when there is a subscriber on the map. Otherwise, you...

Is this scenario in mapping or localization mode? In localization mode, `/rtabmap/localization_pose` will be published with very large covariance >99 until the first localization. Also added [pub_loc_pose_only_when_localizing](https://github.com/introlab/rtabmap_ros/issues/979) parameter for rtabmap...

You would also need `Mem/StereoFromMotion` set to true, so that loop closures can be closed on rtabmap side. For rtabmap's VO, it requires some of the features with depth.

Some features of VPI that could replace opencv functions used in rtabmap: - [Stereo Disparity Estimator](https://docs.nvidia.com/vpi/algo_stereo_disparity.html) - [KLT Feature Tracker](https://docs.nvidia.com/vpi/algo_klt_tracker.html) - [Harris Corner Detector](https://docs.nvidia.com/vpi/algo_harris_corners.html) - [Pyramidal LK Optical Flow](https://docs.nvidia.com/vpi/algo_optflow_lk.html) -...

The "not moving" effect at 2:10 on the first video is similar to KITTI dataset sequence 1 on the highway. To avoid this, We should reduce `Vis/CorGuessWinSize` to 20. Based...