I am training a 3D object detection network (Retinanet-based as of the moment) for re-detecting tracked objects. I would like to be able to add the velocity vector of the tracked object as an input to the detection network, as the velocity directly informs the direction along which the principal axis of the 3D bounding box should lie. I would like to include this information as early as possible (i.e. pass in as an input feature map) rather than simply adding it at the end with a few fully connected layers.
Is there a good or established way to encode such a vector in a feature map?