How to train a CNN on 1 million images when your data is continuous and weakly labeled towards large vocabulary statistical sign language recognition systems