I'm planning to make two changes to how we handle the bridging keypoints that would be a breaking change and require recomputing lots of things:
we should use the precomputed bounding boxes and feed them into the bridging wrapper, rather than using the IOU with the automatically determined bounding box
we should tune how we convert the noise into confidence intervals. currently the sigmoid ends up too sharp and is mostly a binary signal.
I'm planning to make two changes to how we handle the bridging keypoints that would be a breaking change and require recomputing lots of things: