One of the challenges of research in Machine Perception today is filling the gap between researchers and engineers, general algorithms and development of working systems. Key performance issues like robustness, adaptability, flexibility and real-time must be pursued together with simplicity of design. These requirements have accentuated the interest in multi-modular systems and data fusion. The literature has proposed two main paradigms of multi-modular fusion: redundancy (or combination) and selection (or switching). We compare the relative advantages and complementary uses of the two paradigms. Then we focus on selection, which seems to attract less attention than redundancy--at least in Machine Vision, where we operate. After summarizing our work on a selection-based fusion scheme named Context Commutation, we describe a methodology in ten steps to design selection-based systems. With this paper, we hope to contribute to the understanding of selection as an additional tool for researchers and engineers working on machine perception.