Influence of annotation accuracy

[mar-du]

Hi there,
I am currently doing some experiments on the influence of annotation accuracy on yolov5 model performance.
I know the recommendation is the most accurate bounding box (BB) possible, but sometime this is hard to achieve.
That's why I artificially manipulated my BB to evaluate the impact of annotation inaccuracies. It seems like the model adapts quite well to bigger BB much more than I expected. I wanted to discuss if you see any reasons for that. Is there a hyperparameter that adapts to mean BB size that I am not aware of or something like that?
Are there other things of model architecture that need to be considered in this question?
I welcome all thoughts and would like to get more understanding of yolo architecture.
Thanks a lot.

[glenn-jocher]

In AI in general you will usually see graceful degradation of outputs on degrading inputs. There are usually no hard fails that are not explicit coding errors.