Scaling real-world video acquisition for model training