I presume that you guys will be using some learning algorithm to find balls in your image. Congrats if you get it to work, that is no easy task. While I don't have a dataset to give you of my own, you might be interested in taking your learning algorithm a step further and computing the "
DeepPose" of your balls from images. Remove the middle man if you will. Your algorithm last year ran at a speed that left some to be desired. The jetson boards can handle CNN models like no other (thanks to their cuda cores), with inference that is, and not necessarily with training. If you have a CNN with, say, 3 convolution layers, it would run at well over 30 fps.
Just a thought you might want to consider.