Analysis of efficient CNN design techniques for semantic segmentation

Majority of CNN architecture design is aimed at achieving high accuracy in public benchmarks by increasing the complexity. Typically, they are over-specified by a large margin and can be optimized by a factor of 10-100x with only a small reduction in accuracy. In spite of the increase in computational power of embedded systems, these networks are still not suitable for embedded deployment. There is a large need to optimize for hardware and reduce the size of the network by orders of magnitude for computer vision applications. This has led to a growing community which is focused on designing efficient networks. However, CNN architectures are evolving rapidly and efficient architectures seem to lag behind. There is also a gap in understanding the hardware architecture details and incorporating it into the network design. The motivation of this paper is to systematically summarize efficient design techniques and provide guidelines for an application developer. We also perform a case study by benchmarking various semantic segmentation algorithms for autonomous driving.