PROB adapts the Deformable DETR model by adding the proposed 'probabilistic objectness' head. In training, we alternate between distribution estimation (top right) and objectness likelihood ...