search

pkzhou / LGM-SLAM

The open source code of the paper: Visual Localization and Mapping Leveraging the Constraints of Local Ground Manifolds
Other
71 stars 8 forks source link

Question about the number of LGMs. #3

Open xiaobrnbrn opened 1 year ago

xiaobrnbrn commented 1 year ago

Great work! I wonder how to decide the number of LGMs in each frame? According to the paper, it seems that there are several LGMs will be generated from a single frame, but the details are not mentioned enough. Specifically, (1) how to select the region to obtain LGMs in non-level terrain, and (2) what is the boundary to differentiate consecutive LGMs? Looking forward to your insights.

pkzhou commented 1 year ago

In the current implementation, a single keyframe generates one LGM. You can also find it in the code of this project. The LGM generating is in function void LocalMapping::GroundFitting(). (1) The algorithm only runs on the road environment. Currently the ground region from the image is all used for ground fitting. (2) Boundary? The boundary of LGM depends on the used ground region in the image, and yes, the consecutive LGMs will overlap, but I don;t think it matters. I apply a smooth strategy when generating new LGM.

xiaobrnbrn commented 1 year ago

Thanks a lot for your timely reply and here is another question. Supporting that a car is driving on the road, and there is a slope in front of it. At this time, the road is actually a combination of a plane (a small range in front of the car) and a slope, if all of the undulating ground region is used to fit one LGM (actually a local plane) as you did, will it lead to a large fitting error? Or can I understand that only a small area in front of the car is fitted to generate the LGM, and the ground in the distance will be ignored? In this way, the question mentioned above may be solved.