Occlusion problems

[thomassedlmayer]

I noticed that the concaveman algorithm leads to partially transparent objects in the occlusion process. The algorithm influences the outer hull of the objects' 2d projection based on concaveman parameters only. The concaveman algorithm should really only be used for pure 2d tasks, I think. Though, the occlusion task requires the consideration of 3d.

I made the following screenshot to show the consequences:

The red dots pose the supposedly visible vertices of the white delivery van. I guessed that the red lines may aproximately pose the 2d-projected concave hull of the van. The front left vertex of the van is visible even though it is obviously occluded by the van itself. This leads to some kind of spike in the hull. It is added to the hull based on the concaveman algorithm. This means it is not considered at all that it is actually behind other surfaces of the same car. This results in the problem that the three yellow dots are detected by the sensor model.

I guess, as soon as you're working on a 2d projection you should not add vertices back to the hull. You can not know if they are visible without checking/comparing some kind of distance relation or even surface occlusion.

The mentioned problem occurs quite often and you can not rely on detections to be actually visible based on basic understanding of a 3d world. After checking some different scenarios, it seems rather lucky that vehicles occlude other vehicles the correct way. Unfortunately, we can not use the sensor model for arbitrary scenarios.

Removing the concaveman algorithm leads to more reliable results but then any detections that are not on the convex hull of the 2d projection are omitted even if they should be visible. This may lead to kind of satisfying results on the object level. On the detection level though it seems rather random which vertices are added and which are not.