Assume an Object Detection task. The private dataset has its bounding-box (bbox) characteristics, and assume only width-height measures are relevant.
Assume we know what is that width-height distribution of the private dataset.
How is it possible to approximate this distribution based on an existing dataset, which is available?
I though of clustering the private dataset's distribution (for, say $k=10$ clusters, $M \in R^{(k\times2)}$), then greedily picking the closest samples from the available dataset.
I'm wondering what are the available tools, because this first intuition feels way off the optimal
