Hey,

over the course of the last month I created a dataset (images bounding box annotation) for an object detection task and it is working quite well. While evaluating classical metrics like precision and recall is easy, I wondered whether there are other useful metrics that solely rely on evaluating the dataset itself without any object detector involved. I think its general consensus that a dataset should be diverse in terms of scenes, angles etc. to mitigate any potential bias by the detector towards a specific situation. But I do not have a good idea how to evaluate the datasets "diversity", I have some additional tags for every image that I can use, but that feels really simple.

I am grateful for any suggestions towards other metrics :)