Overesteemed results by a model that restores video and make prediction as a whole. However, its performance significantly drops when it operates in an online regime, predicting one frame at a time.
Represent data for training in same modality as it will be inferenced in production, if each frame need to be predicted stratify train dataset to break the ability to interpolate prediction from frames inaccesible in production mode of work.
Ground truth gathering. Dataset preparation.
Сroping on frames, selection which frames to include in train dataset
kaggle "State Farm Distracted Driver Detection" competition JACOBKIE solution