When is Non-Max Suppression used in Object Detection Is non-max suppression for bounding boxes obtained from a Region Proposal Network performed during training? From what I gather, NMS is not differentiable-- in which case, it can't be performed during training (it is also mentioned in this issue). But does that mean that the boxes go into the subsequent layers such as ROI Pooling without NMS?
Where in the pipeline is NMS used and how?
 A: 1. Is non-max suppression for bounding boxes obtained from a Region Proposal Network performed during training?
Yes, according to the Faster-RCNN paper it states,

Some RPN proposals highly overlap with each
other. To reduce redundancy, we adopt non-maximum
suppression (NMS) on the proposal regions based on
their cls scores. We fix the IoU threshold for NMS
at 0.7, which leaves us about 2000 proposal regions
per image. As we will show, NMS does not harm the
ultimate detection accuracy, but substantially reduces
the number of proposals.

So, NMS on proposals is not essential, but it reduces computation without significant loss of performance.
2. NMS is not differentiable-- in which case, it can't be performed during training
According to the paper, authors use Approximate joint training which is,

Approximate joint training: In this solution, the
RPN and Fast R-CNN networks are merged into one
network during training as in Figure 2. In each SGD
iteration, the forward pass generates region proposals which are treated just like fixed, pre-computed
proposals when training a Fast R-CNN detector. The
backward propagation takes place as usual, where for
the shared layers the backward propagated signals
from both the RPN loss and the Fast R-CNN loss
are combined. This solution is easy to implement. But
this solution ignores the derivative w.r.t. the proposal
boxes’ coordinates that are also network responses,
so is approximate.

So if Approximate joint training method is used (which is common use), each proposals are treated as pre-computed proposal. This method might not be mathematically legitimate, but hey, it does the job.
