[autoware.universe] Class Priors for Pointpainting Fusion

[calibratorterminator55]

Hi,

I've been playing with Pointpainting Fusion but I discovered by coincidence that there's little to no difference between using it with and w/out appending class priors to feature of points. The results are pretty much the same when passing a non-painted cloud or a painted one to the detector network.

Another point I would like to ask is about the handling of overlapping FOVs. For a point, corresponding classes are summed as 2s exponent. But this causes change in proposed class if multiple fovs overlap around a lidar point.

e.g.

FOV of 2 cameras overlap around 3D volume V. 
Assume there is a CAR object in V. Lets denote points created by object contained in V as P.
CAR_obj_fac = 1
BIKE_obj_fac = 2
PED_obj_fac = 4

- Fusion Node will fuse rois from 2 different cams by calling fuseOnSingleImage on the \
same cloud with respective rois messages as below.

fuseOnSingleImage(painted_cloud, roisFromCam1); // add 1(CAR_obj_fac) to P's points' class fields
fuseOnSingleImage(painted_cloud, roisFromCam2); // add 1(CAR_obj_fac) to P's points' class fields

- As every point in P is going to be accessed and added by 1 twice in this case. 
Cluster P which normally must be painted as a CAR will now have class code 2(BIKE_obj_fac) which denotes BIKE.