rf_detector is a categorical object detector using Random Forests. It runs a ROS node that expects stereo input data (with textured light) input from the texture_light_on_off node.
rf_detector subscribes to three ROS topics:
After the rf_detector node receive all three topics, it will start detecting objects in the image.
rf_pcd_hogshape_detector_3Class.launch will start the rf_detector with some default parameter settings. There are five groups of parameters which need to be set.
1. table-top object related parameters :
PrineTableFlag -- if true, do table-top object candidate reasoning.
- max_depth -- the maximum depth that the system will search for objects.
- min_depth -- the minimum depth that the system will search for objects.
- plane_axis_x -- the x component of the surface number orientation of the table in the image_frame.
- plane_axis_y -- the y component of the surface number orientation of the table in the image_frame.
- plane_axis_z -- the z component of the surface number orientation of the table in the image_frame.
- eps_angle -- the tolerance of the different between the estimated surface norm direction and the one set by plane_axis.
2. patch sample related parameters
- Radius -- Radius of corresponding the 3d sphere to be sampled.
SampleStep -- The step size (in pixel) between two nearby sampled patches .
- winStep -- The incremental step of the patch size in pixel.
- numCellPerWin -- the number of cells in each patch for both vertical and horizontal direction.
- numCellPerBlock -- the number of cells in each Block for both vertical and horizontal direction.
- padding -- The padding in pixel for both top-left and bottom-right corner.
ObjHeight2winHeightRatio -- Object height / Patch height
3. model related parameters
- !RFmodel_filename -- the file describing the trained random forest.
FgBgProb_filename -- the Foreground/Background probability of each lead nodes in the random forest.
ClassIdBinary_filename -- the Class Id of each trained Foreground patch.
NormVotesBinary_filename --The normalized voting direction, the relative object height to patch height, the aspect ratio of the object.
MaxClassId -- the Maximum Class Id.
ClassId2ViewObjClassMapFile -- the file which map each Class Id into object Class index and Viewpoint index.
4. voting related parameters
- hitThreshold -- The minimum Foreground/Background ratio of a node to cast votes.
MeanShiftMaxIter -- The maximum mean-shift iteration to search for local mode.
ObjHeightStep -- The incremental step of the object height in pixel.
- !ObjASStep -- The incremental step of the object aspect ratio.
5. detection post processing related parameters
RectOverlapRatio -- The threshold to decide if two bounding boxes are overlapping each other.
ConfThre -- It defines the minimum score for a object detector could return.