The region proposal network (RPN) in Faster-RCNN models contains a classifier and a regressor network. Why does the classifier network output two scores (object and background) for each anchor instead of just a single probability? Aren't the two classes considered exclusive?

Source: Figure 3 of the original Faster-RCNN paper


1 Answer 1


I just want to provide this intuition

  • this NN consists of a 2 steps detection pipeline (the region proposal and regression + classification in parallel) exploring a certain range of scales and aspect ratios for proposals

  • as the proposed region is rectangular and the objects of interests have not a rectangular appearance strictly speaking, but a mostly rectangular one (with different ARs), considering the resulting detection bboxes areas some pixels will belong to a certain relevant object appearance (e.g. pixels of a pedestrian, car, …) while other pixels in that BBox won’t be related to that semantic

  • as this is not semantic segmentation we can now know what pixel is actually relevant, however the objectness and backgroundness can provide a rough measure of this in an aggregated way

Let’s assume at some point in your processing pipeline, you have 2 regions containing the same car but at different scales: so ideally both of them should be associated with good confidence to CarID but it’s clear one will contain much more background pixels than the other one and you'd like this to be represented as an additional measure

  • 1
    $\begingroup$ That's a fantastic answer for a year-old question. Thanks! $\endgroup$
    – dseuss
    Commented Mar 19, 2019 at 1:44

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .