2017/09/18 by Mohammad Javad Shafiee, Shafiee, Mohammad Javad, Brendan Chywl +5 · 3 citations
Computer Science · Engineering · #Advanced Neural Network Applications #Artificial Intelligence (cs.AI) #CCD and CMOS Imaging Sensors #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Neural and Evolutionary Computing (cs.NE) #Visual Attention and Saliency Detection
paper · pdf · doi:10.48550/arxiv.1709.05943
openalex publication_date 2017/09/18 · openalex created_date 2022/10/06 · openalex updated_date 2026/07/28
Object detection is considered one of the most challenging problems in this\nfield of computer vision, as it involves the combination of object\nclassification and object localization within a scene. Recently, deep neural\nnetworks (DNNs) have been demonstrated to achieve superior object detection\nperformance compared to other approaches, with YOLOv2 (an improved You Only\nLook Once model) being one of the state-of-the-art in DNN-based object\ndetection methods in terms of both speed and accuracy. Although YOLOv2 can\nachieve real-time performance on a powerful GPU, it still remains very\nchallenging for leveraging this approach for real-time object detection in\nvideo on embedded computing devices with limited computational power and\nlimited memory. In this paper, we propose a new framework called Fast YOLO, a\nfast You Only Look Once framework which accelerates YOLOv2 to be able to\nperform object detection in video on embedded devices in a real-time manner.\nFirst, we leverage the evolutionary deep intelligence framework to evolve the\nYOLOv2 network architecture and produce an optimized architecture (referred to\nas O-YOLOv2 here) that has 2.8X fewer parameters with just a ~2% IOU drop. To\nfurther reduce power consumption on embedded devices while maintaining\nperformance, a motion-adaptive inference method is introduced into the proposed\nFast YOLO framework to reduce the frequency of deep inference with O-YOLOv2\nbased on temporal motion characteristics. Experimental results show that the\nproposed Fast YOLO framework can reduce the number of deep inferences by an\naverage of 38.13%, and an average speedup of ~3.3X for objection detection in\nvideo compared to the original YOLOv2, leading Fast YOLO to run an average of\n~18FPS on a Nvidia Jetson TX1 embedded system.\n