Thanks for the notebook, will look through it.
Re: your question, I don’t quite understand what you mean by:
The larger dog prediction box (teal color, confidence of 0.38) should only count as 1 FP in the dog category and not be counted at all in the sofa category. This is what I’m understanding your first statement (“I was thinking it would count as, Dog: 1TP, 1FP and Sofa: 1TP”) to mean and I agree with that.
Maybe you have an error in keeping the classes separate when doing your TP/FP counts? Or do you mean that some variant of the mAP metric is supposed to count detections across different classes as FPs?