Non-submodular Visual Attention for Robot Navigation