How did you decide how to merge the detection classes?

On this page users can suggest improvements to the nuScenes dataset, homepage, devkit etc.

Thanks for the great effort you put into this dataset!

Could you elaborate the reason behind having a separate construction_vehicle class in the detection challenge and not assigning it e.g. to truck?

Similar question for ambulance and police vehicles:
Why did you ignore them instead of assigning them to truck and car respectively?

Caring about construction vehicles (as a separate class) and ignoring ambulances and police cars in the evaluation seems a bit inconsistent (and gives me the impression that you care more about construction vehicles than ambulances and police cars).

Not meant as a critique, asking more out of curiosity.

1 Like

you can check construction_vehicle and police cars’ instance count.

Hi @MartinHahner.
There are a number of factors that we took into account for keeping/merging certain classes, such as:

  1. The frequency of a class. Very rare classes may have too little data to train and evaluate performance on them. According to https://www.nuscenes.org/data-annotation we have 14671 construction vehicles, but only 49 ambulances. This does not mean that they are not important (ambulance and police vehicles definitely are).
  2. The importance of a class. E.g. bicycle racks are only a tool used to filter out parked bicycles, we don’t care about the racks themselves.
  3. Whether they can be merged with other classes. E.g. trucks have a typical size and velocity range. If we add construction vehicles, which are much larger and typically stationary, the various statistics may become diluted.
1 Like

@holger-motional: thanks for your fast and detailed answer!

1 Like