To boost NMS performance, we apply early termination here at this PR.

Notice that current detection out layer will be used internally by proposal layer.
In this case, finally, proposal will, ex output top 100 proposals among all candidates.

Typically, the # of candidates is about more than several thousands.
Currently flow seriously down-grades performance by applying NMS for all candidates.
However, as we know, to output top K NMS output, we should check if we have picked enough candidates and terminate once we have collected what we need.

As the table below, to output top K = 100 candidates, we just need to check less than one thousand bboxes
But for current flow, we check all.

Since NMS is the most time-consuming part of proposal (detection out) layer, it downgrades performance much.
Ex, the experiment shows we will speed up by > x10 times for NMS.

As the result, the inference time for proposal (detection out) layer can speedup by, ex, at x2~x3 times faster than original flow.

Finally, to output top 100, from the table below we understand that original flow does unnecessary work in calculating extra candidates (which > 100).
By saving it, we can greatly boost layer inference performance.

test image can be download at following link:

testImage.zip

test model is:
opencv/opencv_extra/testdata/dnn/mask_rcnn_inception_v2_coco_2018_01_28.pbtxt
mask_rcnn_inception_v2_coco_2018_01_28/frozen_inference_graph.pb

Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

• I agree to contribute to the project under Apache 2 License.
• To the best of my knowledge, the proposed patch is not based on a code under GPL or other license that is incompatible with OpenCV
• The PR is proposed to proper branch
• There is reference to original bug report and related work
• There is accuracy test, performance test and test data in opencv_extra repository, if applicable
  Patch to opencv_extra has the same branch name.
• The feature is well documented and sample code can be built with the project CMake