Learning to Identify Critical States for Reinforcement Learning from Videos