bad lidar annotations in some scenes

[dorn-gerhard]

Hi, I examined that some frames have bad (wrong) lidar annotations regarding measurement points, which are invalid (for instance the laser beam didn't come back as for very remote targets or when hitting very reflective surfaces such as car windows.

The NuScene data set is dealing with invalid data (e.g. measurements which did not return) by labelling them with 31 (ego_vehicle). In fact points belonging to those invalid measurements have a radius close to zero.

Here is an example from "scene-1073", sample 4:

• Upper image: plotting the log10 of the radii of lidar measurements using phi-theta coordinates (1084-1087 phi points, 32 theta points), Black spots correspond to invalid data as radius is below 10^-4 meter
• Lower image: annotation of ego (1) and vegetation (2) is plotted).

From the image it is clear, that in the upper part lidar is shooting in the sky black spots. The labels here should not be vegetation but Ego vehicle (or the class corresponding to invalid data).

This is sample 1 of the same scene, providing a correct (consistent) annotation Note that the radii of non-valid points (ego_vehicle) are not scaled down to 10^-5 in this sample

[whyekit-motional]

@dorn-gerhard if a lidar measurement does not have a return, there would not be any point to label in the first place

Thus, it is not clear how you are inferring that

invalid data (e.g. measurements which did not return) by labelling them with 31 (ego_vehicle)