Making Curiosity Explicit in Vision-based RL