GRAVA: An Architecture Supporting Automatic Context Transitions and Its Application to Robust Computer Vision