-
Notifications
You must be signed in to change notification settings - Fork 391
Parametrizing SegMap
This page presents how the different parameters influence the SegMatch performances and provides ideas on how to select adequate values for obtaining the best results in your own applications. For each relevant parameter, we included a link to where this value can be set in the configuration file.
The radius of the local map directly influences the number of local segments which can be used for localization. However, with the current batch segmentation implementation (more to come..), larger radii highly increase the computational power required by the segmentation step. In our experiments and for the KITTI dataset, we noted that a radius between 40 and 60 meters, which yields in average between 30 and 40 local segments, resulted in good localization performances.
When creating the target segment map online, we remove duplicate segments by ensuring that no pair of segments have their centroid closer than a minimum centroid distance. This step is crucial as is considerably minimizes ambiguity in the geometric verification step.
The voxel grid resolution and the Euclidean region growing search radius will influence the segmentation process, whether it will over or under segment. If you get too few segments, you should consider reducing the growing search radius and verify that the ground is correctly removed. For the demonstrations based on the KITTI dataset, this is performed by filtering points based on height (see here).
In order to filter measurement noise, one can consider increasing the minimum number of point for activating a voxel.
Finally, we also use a curvature-based region growing segmenter for which we will add a demonstration shortly (see here).
Currently, the eigen value-based descriptor does not require any parametrization.
Retrieval is parametrized by the number of nearest neighbors to search for and on the feature space dimension. For the latter, this should be 7, the dimension of the eigen value-based descriptor.
When using a random forest to refine the correspondence candidates (i.e. enable two stage retrival is true) one can adjust the threshold on the random forest output.
In general, having more correspondence candidates increases the chances to localize but increase the computing needed by the subsequent geometric verification step which then needs to be more strict.
The geometric verification step is driven by two parameters. First, the resolution represents the maximum distance in meters for considering two correspondences as consistent. Then, we require a minimum number of these correspondences to be alltogether pair-wise consistent. This is represented by the minimum cluster size. In order to reduce the potential occurrence of false localizations, one can either reduce the local map radius, reduce the maximum distance for consistency, or increase the minimum cluster size. In general, localizations based on higher number of consistent correspondences yield more accurate results.
This page will be updated as we add new features to the repository. More to come! :)