The reason for scaling the bounding box size when computing the coverage in detectnet?

[szm-R]

Hi, I've been using detectnet for a while and as I was getting good results with the default settings (almost the default settings) I didn't give them much thought until recently.

Now I've been reading the detectnet_transform_layer.cpp and detectnet_coverage_rectangular.cpp when I came across this peice of code (line 182 of detectnet_coverage_rectangular.cpp of function coverageBoundingBox ) Size2v shrunkSize = boundingBox.size() * (Dtype)this->param_.scale_cvg();

Where the actual bounding box size of the objects is scaled with scale_cvg parameter (which has the value of 0.4 in the prototxt provided in NVIDIA Caffe and a default value of 0.5 in caffe.proto). I was wondering what is the need for such scaling when computing the coverage region?

[mablem8]

Somewhat unsatisfying answer here: https://github.com/NVIDIA/caffe/issues/236