# Tiny SSD: A Tiny Single-shot Detection Deep Convolutional Neural Network for Real-time Embedded Object Detection

**Authors:** Alexander Wong, Mohammad Javad Shafiee, Francis Li, Brendan Chwyl  
**arXiv:** [1802.06488](https://arxiv.org/abs/1802.06488)  
**Date:** 2018-02-19  
**Summary page:** https://www.summarizepaper.com/en/paper/1802.06488v1/

## Key points

- Object detection is a challenging problem in computer vision
- Deep neural networks are effective but difficult to deploy on embedded devices
- Tiny SSD is a small deep neural network architecture designed for embedded object detection
- It combines the efficiency of SqueezeNet's Fire microarchitecture with the performance of SSD's single-shot detection macroarchitecture
- Tiny SSD minimizes model size while maintaining object detection performance
- The resulting model has a size of 2.3MB, 26 times smaller than Tiny YOLO
- Tiny SSD achieves an mAP of 61.3% on VOC 2007, higher than Tiny YOLO's results
- It enables efficient object detection on resource-constrained devices like smartphones or IoT devices

## Summary

Object detection is a challenging problem in computer vision that involves identifying and localizing objects within an image. Deep neural networks have been shown to be effective for addressing this challenge but their high computational and memory requirements make them difficult to deploy on embedded devices. To overcome this limitation, researchers have developed small deep neural network architectures for object detection that are more suitable for embedded scenarios. In this paper, the authors introduce Tiny SSD - a single-shot detection deep convolutional neural network designed specifically for real-time embedded object detection. Tiny SSD combines the efficiency of the Fire microarchitecture from SqueezeNet with the object detection performance of the single-shot detection macroarchitecture from SSD. The network consists of a highly optimized non-uniform Fire sub-network stack and a non-uniform sub-network stack of highly optimized SSD-based auxiliary convolutional feature layers. The key objective of Tiny SSD is to minimize model size while maintaining object detection performance. The resulting model has a size of 2.3MB which is approximately 26 times smaller than Tiny YOLO - another popular embedded object detection architecture. Despite its small size, Tiny SSD achieves an mAP (mean average precision) of 61.3% on VOC 2007 which is approximately 4.2% higher than Tiny YOLO's results. These experimental results demonstrate that it is possible to design very small deep neural network architectures for real-time object detection that are well suited for embedded scenarios without sacrificing performance. By reducing both computational and memory requirements, Tiny SSD opens up new possibilities for deploying efficient object detection systems on resource constrained devices such as smartphones or IoT devices.