# Pelee: A Real-Time Object Detection System on Mobile Devices

**Authors:** Robert J. Wang, Xiang Li, Charles X. Ling  
**arXiv:** [1804.06882](https://arxiv.org/abs/1804.06882)  
**Date:** 2018-04-18  
**Summary page:** https://www.summarizepaper.com/en/paper/1804.06882v3/

## Key points

- Growing demand for efficient Convolutional Neural Network (CNN) models for mobile devices with limited computational power and memory resources
- Emergence of efficient architectures like MobileNet, ShuffleNet, and MobileNetV2 relying on depthwise separable convolution
- Introduction of PeleeNet architecture by researchers led by Robert J. Wang, Xiang Li, and Charles X. Ling using conventional convolution instead of depthwise separable convolution
- PeleeNet achieved higher accuracy and ran over 1.8 times faster than MobileNet and MobileNetV2 on NVIDIA TX2 hardware while being only 66% of the size of MobileNet
- Development of real-time object detection system named Pelee combining PeleeNet with SSD method optimized for speed
- Impressive results: Pelee achieved mean average precision (mAP) of 76.4% on PASCAL VOC2007 dataset and 22.4 mAP on MS COCO dataset at speeds of 23.6 FPS on iPhone 8 and 125 FPS on NVIDIA TX2 hardware
- Outperformed YOLOv2 in precision with a computational cost that was 13.6 times lower and a model size that was 11.3 times smaller
- Demonstrates potential impact of efficient model design in enabling real-time object detection systems to operate seamlessly on mobile devices without compromising performance or accuracy levels

## Summary

In the realm of deep learning, there is a growing demand for efficient Convolutional Neural Network (CNN) models that can run on mobile devices with limited computational power and memory resources. This necessity has spurred research into designing models that are both accurate and fast. Recent years have seen the emergence of several efficient architectures such as MobileNet, ShuffleNet, and MobileNetV2, all of which heavily rely on depthwise separable convolution. However, this particular type of convolution lacks efficient implementation in most deep learning frameworks. To address this challenge, a team of researchers led by Robert J. Wang, Xiang Li, and Charles X. Ling proposed an innovative architecture called PeleeNet. Unlike its counterparts, PeleeNet is built using conventional convolution instead of depthwise separable convolution. The researchers conducted experiments on the ImageNet ILSVRC 2012 dataset and found that PeleeNet not only achieved higher accuracy but also ran over 1.8 times faster than MobileNet and MobileNetV2 on NVIDIA TX2 hardware. Additionally, PeleeNet boasted a model size that was only 66% of MobileNet's size. Building upon the success of PeleeNet, the researchers went on to develop a real-time object detection system named Pelee by combining PeleeNet with the Single Shot MultiBox Detector (SSD) method and optimizing the architecture for speed. The results were impressive - Pelee achieved a mean average precision (mAP) of 76.4% on the PASCAL VOC2007 dataset and 22.4 mAP on the MS COCO dataset while running at speeds of 23.6 frames per second (FPS) on an iPhone 8 and 125 FPS on NVIDIA TX2 hardware. Notably, Pelee outperformed YOLOv2 in terms of precision while offering a computational cost that was 13.6 times lower and a model size that was 11.3 times smaller. These findings demonstrate the potential impact of efficient model design in enabling real-time object detection systems to operate seamlessly on mobile devices without compromising performance or accuracy levels. This groundbreaking research by Wang et al., accepted at NeurIPS 2018, showcases how innovative architectural choices can lead to significant advancements in deep learning applications for mobile platforms.