Challenges in Learning Optimum Models for Complex First Order Activity Recognition Settings