Do Deep Learning-based methods introduce all that much latency? My understanding of neural network is that the training is expensive, but using them is essentially free - you have to push your pixels through a few layers of matrix multiplications and you have the results; it's just tuning those matrices that takes time.