Convolutional neural networks have brought groundbreaking accuracy and generalization to the field of computer vision. New processors promise to lower cost and power requirements for deploying these algorithms. Optimizations such as binary networks have reduced computational complexity, but simpler, more “conventional” computer vision techniques continue to offer an attractive cost-to-performance ratio for easier tasks, with the ability to run on older, less expensive edge SOCs. Such algorithms also usually require orders of magnitude fewer training examples. However, conventional algorithms often need hand-tuning of parameters and do not generalize well to previously unseen environments. By combining CNN vision with simpler algorithms into a layered intelligent vision pipeline, and by understanding the constraints of the problem, the weaknesses of simpler algorithms can be offset by the strengths of CNNs, while still preserving their cost-saving benefits.