Data driven adaptation for QoS aware embedded vision systems

We present a data driven method for high efficiency in a neuro-inspired vision pipeline. Our goal is to reduce low-utility computation arising from duplicated processing. In this paper, we examine two forms of redundant information in image data, spatiotemporal redundancy and channel redundancy. To maximize efficiency, the paper presents a dynamic, configurable approach that limits the computational cost of hardware by reusing previous results and sharing data paths. Our technique reduces redundant computation from both spatiotemporal and channel redundancy.