# Fast YOLO: A Fast You Only Look Once System for Real-time Embedded Object Detection in Video

**Authors:** Mohammad Javad Shafiee, Brendan Chywl, Francis Li, Alexander Wong  
**arXiv:** [1709.05943](https://arxiv.org/abs/1709.05943)  
**Date:** 2017-09-18  
**Summary page:** https://www.summarizepaper.com/en/paper/1709.05943v1/

## Key points

- Object detection is a complex task in computer vision
- Deep neural networks (DNNs) have shown superior performance in object detection, with YOLOv2 being one of the state-of-the-art models
- The authors propose a new framework called Fast YOLO to address real-time object detection on embedded devices with limited computational power and memory
- Fast YOLO leverages the evolutionary deep intelligence framework to evolve the YOLOv2 network architecture and create an optimized version called O-YOLOv2
- O-YOLOv2 has significantly fewer parameters while maintaining minimal drop in accuracy
- Fast YOLO introduces a motion-adaptive inference method to reduce power consumption by analyzing temporal motion characteristics
- Experimental results show that Fast YOLO can reduce deep inferences by 38.13% and achieve a speedup of ~3.3X compared to original YOLOv2 on Nvidia Jetson TX1 embedded system
- Fast YOLO enables real-time object detection on resource-constrained devices without sacrificing performance

## Summary

Object detection is a complex task in computer vision that involves identifying and localizing objects within an image or video. Deep neural networks (DNNs) have shown superior performance in object detection, with YOLOv2 being one of the state-of-the-art models in terms of speed and accuracy. To address the issue of applying YOLOv2 to real-time object detection on embedded devices with limited computational power and memory, the authors propose a new framework called Fast YOLO. This framework leverages the evolutionary deep intelligence framework to evolve the YOLOv2 network architecture and create an optimized version called O-YOLOv2. This optimized architecture has significantly fewer parameters (2.8X fewer) while maintaining a minimal drop in Intersection over Union (IOU) accuracy (~2%). Additionally, Fast YOLO introduces a motion-adaptive inference method to reduce power consumption on embedded devices by analyzing temporal motion characteristics to determine when deep inference with O-YOLOv2 is necessary, reducing its frequency. Experimental results demonstrate that Fast YOLO can reduce the number of deep inferences by an average of 38.13% and achieve an average speedup of ~3.3X compared to the original YOLOv2 on a Nvidia Jetson TX1 embedded system, resulting in an average frame rate of ~18FPS for real-time object detection in videos. Overall, this paper presents a novel approach to accelerate object detection using modified versions of YOLOv2 and motion-adaptive inference, enabling real-time performance on resource constrained devices without sacrificing performance.