Background
Deep neural networks have become the de-facto standard in almost all
computer vision tasks. Ever since deep networks have outperformed conventional
computer vision approaches (Krizhevsky et al. 2012), there has been a
constant improvement in their performance, which has hinged on more powerful
hardware, greater funding of this research direction and wider involvement
of researchers world-wide (
FT ). To a large extent, the research has been focused on accuracy of these deep networks (as seen through the challenge tracks in most of CV competitions) and only recently (e.g., co-design of DNN, Kwon et al. 2018, deployment on edge devices Shafiee et al. 2017) more effort has been devoted to efficient deployment, which is often the more important aspect for industrial applications. However, ensuring the reliable performance of the network has remained comparatively less explored.
