Is it possible to train the ffn with several distinct pieces of data with each having its own bounding box?