How would you design a computer vision pipeline for a real-time application (e.g. video surveillance) balancing accuracy and latency?