Describe the feature and motivation

While reviewing FaceDetectorYNImpl::postProcess to see how it worked, I noticed that the scoreThreshold is only used to cull faces in the non-maximum suppression call late in the function. I think an opportunity exists to just not add the box unless it meets the scoreThreshold around this line of code as dnn::NMSBoxes does that anyways later:

If I understand correctly, this would be a helpful advantage because there are typically many 1000's of raw face boxes but after thresholding on the score it could often drop to well below 100, which would help with a few things: dynamic memory allocation up front, reduced copies to faceBoxes, and reduced input to dnn::NMSBoxes.

Does this seem reasonable or am I missing something important about how it works?

Additional context

No response