Postprocessing Peoplenet

Hi I am deploying Peoplenet on Triton Inference server. The model returns, for each class bounding boxes of shape [4, 60, 34] and confidences of [60, 34] (I imagine these to be 16x16 px subdivisions of the input image). How can we obtain the detections from this result?

For postprocess, please refer to /opt/nvidia/deepstream/deepstream/sources/libs/nvdsinfer_customparser/nvdsinfer_custombboxparser.cpp