Vastly different behaviour FasterRCNN v1 and v2

torchvision.models.detection.faster_rcnn — Torchvision main documentation on this page there are two different implementations:

  • fasterrcnn_resnet50_fpn (I get clutter overlapping predictions and low map50 performance), but importantly no NMS is applied which is really what I’m after, I want control of that step.

  • fasterrcnn_resnet50_fpn_v2. NMS seems to be applied internally, and I get beautiful clean bounding boxes, but cluttered scenes fall apart and I get one big bounding box for the whole frame (it’s actually more of a counting problem, so that’s about the worst possible). Much higher map50s though.

Basically I’m after the performance of the latter with the behaviour of the first.
Lots of the parameters of both should be responsible for tweaking the internal NMS, but so far I’ve had very little luck with them.

nms_thresh=1 should turn off NMS right?