Next, two heads
are added on top for object detection: a linear layer for classifying each object query into one of the objects or "no
object", and a MLP to predict bounding boxes for each query.