Faster RCNN Box detection per image

Hi,

For the Faster RCNN model and e.g. MaskRCNN there is a parameter called box_detections_per_image and the default value is 100. When I set it to one however I still get more than one prediction, so what does the parameter actually do?

Is it the number of boxes inside the region proposal?

What should the default be if you are only looking for one object in the image?

This has now been solved