Is there a way to directly read frames from a video and feed them to the model for training, instead of processing the video into individual frames and storing them locally? Doing so poses a significant challenge to disk capacity.