Towards Closing the Energy Gap Between HOG and CNN Features for Embedded\n Vision

摘要

Computer vision enables a wide range of applications in robotics/drones,\nself-driving cars, smart Internet of Things, and portable/wearable electronics.\nFor many of these applications, local embedded processing is preferred due to\nprivacy and/or latency concerns. Accordingly, energy-efficient embedded vision\nhardware delivering real-time and robust performance is crucial. While deep\nlearning is gaining popularity in several computer vision algorithms, a\nsignificant energy consumption difference exists compared to traditional\nhand-crafted approaches. In this paper, we provide an in-depth analysis of the\ncomputation, energy and accuracy trade-offs between learned features such as\ndeep Convolutional Neural Networks (CNN) and hand-crafted features such as\nHistogram of Oriented Gradients (HOG). This analysis is supported by\nmeasurements from two chips that implement these algorithms. Our goal is to\nunderstand the source of the energy discrepancy between the two approaches and\nto provide insight about the potential areas where CNNs can be improved and\neventually approach the energy-efficiency of HOG while maintaining its\noutstanding performance accuracy.\n