BOW for complex environments

[alexr]

Hi Mathieu,

I have a question about applying RTABMAP in complex environments.

I understand RTABMAP uses the BOW to make a visual dictionary of features and then the features are matched using the SURF or SIFT features to recognize the location.

How well does the BOW method work in complex environments say forests, or agricultural fields. As the environment has lots of vegetation, can you answer how well the BOW method can retrieve different plant species with the SIFT and SURF? How well the RTABMAP can handle such complex environments?

Thanks

Alex

[matlabbe]

Hi Alex,

Not sure if I ever see (in papers) someone trying loop closure detection in the forest or agricultural fields, loop closure datasets are generally in human-made environments. BOW won't distinguish between species (unless maybe with a more complex approach looking for similar SIFT/SURF features extracted from a particular species, in a supervised way). Highly repetitive environments such as large fields would trigger false loop closure detections. Beside increasing loop closure detection threshold (Rtabmap/LoopThr), RTAB-Map has also some mechanisms (see this Robust Graph Optimization page for examples) to reject these wrong loop closures, as long as the odometry is relatively good and consistent.

cheers,
Mathieu