Orthographic Projection camera - Undesirable semantic segmentation of buildings from bird's eye view (occupancy grid map)

[s-hd]

CARLA version: 0.9.13 Platform/OS: Ubuntu 20.04.4 LTS

Hello,

I would like to use a camera ~4m above the ego_vehicle with orthographic projection.

For this, I found a possibility in the following forum that seemed to work well: https://forum.carla.org/t/orthographic-camera-sensor/579/5 Thereby, the attribute "orthographic" is enabled for the "sensor.camera.semantic_segmentation" among others.

Here, some relevant screenshots of the .diff file:

Now I notice that the result for buildings unfortunately looks like this: Town02:

Town10HD:

I would expect the buildings to be semantically segmented as "building" rather than "sidewalk" even from above.

I appreciate any hints.

[s-hd]

Another example:

The Ortho-cam:

The Semantic-segmentation-cam from 1000m height:

It would be desirable to use an orthographic projection, but to classify the buildings "correctly" (as in the semantic segmentation from 1000m).