# YOLO Nano: a Highly Compact You Only Look Once Convolutional Neural Network for Object Detection

**Authors:** Alexander Wong, Mahmoud Famuori, Mohammad Javad Shafiee, Francis Li, Brendan Chwyl, Jonathan Chung  
**arXiv:** [1910.01271](https://arxiv.org/abs/1910.01271)  
**Date:** 2019-10-03  
**Summary page:** https://www.summarizepaper.com/en/paper/1910.01271v1/

## Key points

- Object detection is a crucial task in computer vision
- Deep convolutional neural networks have made significant progress in object detection
- Deploying these networks on edge and mobile devices poses challenges due to high computational and memory requirements
- Researchers have focused on designing efficient deep neural network architectures for edge and mobile usage
- YOLO Nano is a highly compact deep convolutional neural network designed for object detection
- YOLO Nano follows a human-machine collaborative design strategy
- It is based on the design principles from the YOLO family of single-shot object detection networks
- YOLO Nano has a model size of approximately 4.0MB, 15 times smaller than Tiny YOLOv2 and Tiny YOLOv3
- It requires 4.57 billion operations for inference, which is more than 34% lower than Tiny YOLOv2 and around 17% lower than Tiny YOLOv3
- Despite its small size and reduced computational requirements, it achieves an impressive mean average precision (mAP) of 69.1% on the VOC 2007 dataset
- It surpasses Tiny YOLOv2 by 12% and Tiny YOLOv3 by 10.7%
- The efficacy of YOLO Nano for embedded scenarios is demonstrated through experiments on inference speed and power efficiency using a Jetson AGX Xavier embedded module at different power budgets
- Overall, it provides a highly compact yet efficient deep convolutional neural network architecture well-suited for embedded object detection tasks

## Summary

Object detection is a crucial task in computer vision and has seen significant progress through the development of deep convolutional neural networks. However, deploying these networks on edge and mobile devices poses challenges due to their high computational and memory requirements. To address this issue, researchers have focused on designing efficient deep neural network architectures specifically tailored for edge and mobile usage. In this study, the authors introduce YOLO Nano – a highly compact deep convolutional neural network designed for object detection. YOLO Nano follows a human-machine collaborative design strategy that combines principled network design prototyping with machine-driven design exploration. The architecture is based on the design principles from the YOLO family of single-shot object detection networks. It stands out for its compactness with a model size of approximately 4.0MB – 15 times smaller than Tiny YOLOv2 and Tiny YOLOv3 – while requiring 4.57 billion operations for inference which is more than 34% lower than Tiny YOLOv2 and around 17% lower than Tiny YOLOv3. Despite its small size and reduced computational requirements, it achieves an impressive mean average precision (mAP) of 69.1% on the VOC 2007 dataset – surpassing Tiny YOLOv2 by 12% and Tiny YOLOv3 by 10.7%. The efficacy of YOLO Nano for embedded scenarios is further demonstrated through experiments on inference speed and power efficiency using a Jetson AGX Xavier embedded module at different power budgets. Overall, it addresses the challenge of deploying object detection networks on edge and mobile devices by providing a highly compact yet efficient deep convolutional neural network architecture that is well-suited for embedded object detection tasks.