When training a CNN or using YOLO for finding objects (e.g. a hand) you have to apply many variations to the images to generalize the concept hand, like random backgrounds, different shadows and lights, blocking etc. The question arises how big the influence of each variation type on the detection rate is? Use a hand model in Unity, create variations, train a CNN, then detect hands. Try many combinations of variations and evaluate the detection rate. Infer the influence of each variation.
If you are interested in this topic or for more information please contact Prof. Helmut Hlavacs (helmut.hlavacs(at)univie.ac.at).