# Tiny-DSOD: Lightweight Object Detection for Resource-Restricted Usages

**Authors:** Yuxi Li, Jiuwei Li, Weiyao Lin, Jianguo Li  
**arXiv:** [1807.11013](https://arxiv.org/abs/1807.11013)  
**Date:** 2018-07-29  
**Summary page:** https://www.summarizepaper.com/en/paper/1807.11013v1/

## Key points

- The paper addresses the challenge of resource limitations in object detection methods
- Most existing methods require substantial computational resources, making them unsuitable for deployment on resource-restricted devices
- Tiny-DSOD is a novel object detection framework that considers the trade-off between resource usage and accuracy
- It introduces two innovative architecture blocks: depthwise dense block (DDB) based backbone and depthwise feature-pyramid-network (D-FPN) based front end
- Extensive experiments were conducted on PASCAL VOC 2007, KITTI, and COCO benchmarks to evaluate the performance of Tiny DSOD
- Tiny DSOD outperforms other ultra efficient object detection solutions such as Tiny YOLO, MobileNet SSD (v1 & v2), SqueezeDet, Pelee in terms of parameter size, FLOPs, and accuracy
- Achieves an impressive mean average precision (mAP) of 72.1% with only 0.95M parameters and 1.06B FLOPs
- Tiny DSOD is a lightweight object detection framework specifically designed for resource restricted usages

## Summary

The paper titled "Tiny-DSOD: Lightweight Object Detection for Resource-Restricted Usages" addresses the challenge of resource limitations in object detection methods. While deep learning has significantly advanced object detection in recent years, most existing methods require substantial computational resources, making them unsuitable for deployment on resource-restricted devices such as low-end devices and always-on devices. To overcome this limitation, the authors propose Tiny-DSOD, a novel object detection framework that considers the trade-off between resource usage and accuracy. Building upon the deeply supervised object detection (DSOD) framework, Tiny-DSOD introduces two innovative and highly efficient architecture blocks: the depthwise dense block (DDB) based backbone and the depthwise feature-pyramid-network (D-FPN) based front end. The authors conduct extensive experiments on three well known benchmarks - PASCAL VOC 2007, KITTI, and COCO - to evaluate the performance of Tiny DSOD. They compare it to other state of the art ultra efficient object detection solutions such as Tiny YOLO, MobileNet SSD (v1 & v2), SqueezeDet, Pelee etc. The results demonstrate that Tiny DSOD outperforms these solutions in all three metrics: parameter size, FLOPs (floating point operations), and accuracy. For instance, Tiny DSOD achieves an impressive mean average precision (mAP) of 72.1% with only 0.95M parameters and 1.06B FLOPs. This performance is remarkable considering its low resource requirements. In conclusion, this paper presents Tiny DSOD as a lightweight object detection framework specifically designed for resource restricted usages. Its innovative architecture blocks enable efficient utilization of resources while maintaining high accuracy levels. The experimental results validate its superiority over other ultra efficient object detection solutions in terms of both performance and resource requirements.