Re: Info on some parameters

Posted by matlabbe on
URL: http://official-rtab-map-forum.206.s1.nabble.com/Info-on-some-parameters-tp4568p4572.html

Hi Manuel,

1) T_ratio can be used to reject loop closure hypotheses that are decreasing. For example if set at 0.9, then if the last best hypothesis was 0.6, then the current one should be at least 0.9x0.6=0.54 to be considered as a loop closure hypothesis. By default it is disabled, as in RGB-D SLAM mode, the wrong hypotheses can be easily rejected on the motion estimation step.

2) When creating a new node in the map, the laser scan can be filtered. If downsampling is set to 2, half of the rays will be filtered. If voxel size is set, the density of the scan will be uniformly filtered to have 1 pt/voxel size. The normal K and radius are parameters used to compute normals in the scan, if they are not already computed. Those parameters are also in ICP parameters, but it is better do filtering when creating the node instead of using those in ICP to avoid refiltering each time we want to do ICP with those scans. These parameters are not used if you don't have lidar.

3) T_similarity is explained in section III-B of this paper: Appearance-Based Loop Closure Detection for Online Large-Scale and Long-Term Operation
For the "weight->ID" thing, the difference is the ID vs Age. Age is updated when a node is brought back in Working Memory, as ID not. In the papers, the "weight->Age" approach is used.

4) During quantization to visual word vocabulary, features matching a word in the vocabulary will be linked to descriptor of this word. With this option on, we still keep the descriptor of that feature for motion estimation. With raw descriptors saved, when we need to estimate motion estimation between two images of a loop closure, we recompute the matches between them using original descriptors of all features (ignoring quantization to vocabulary), generating more accurate correspondences.

cheers,
Mathieu