Choosing appropriate anchor sizes

[holtvogt]

In my dataset, I will be working on only a limited point cloud range. In this area, very frequently a certain vehicle class will never be completely displayed, or will be truncated.

• In this context, should the anchor size still refer to the actual average size of the vehicle, e.g. a bus?
• Does the anchor size affect the training if the ground truth boxes are also only as large as the area allows?

Example training frame: