When using TensorFlow Object Detection on an image with for example a crowd of people, the objects get grouped together like so:
This is of course good for some use cases, but I would like to know if there is a way to not group objects together.
Is there a way to lower the threshold, for when objects are grouped together?
the reason for this I believe is that sometimes in the COCO training set, crowds of people are annotated as "person". Consequently, the detectors also learn to label crowds of people as "person". There is a way to ignore these "is_crowd" person annotations, but we haven't done this yet.