# Adaptation of MobileNetV2 for Face Detection on Ultra-Low Power Platform

**Authors:** Simon Narduzzi, Engin Türetken, Jean-Philippe Thiran, L. Andrea Dunbar  
**arXiv:** [2208.11011](https://arxiv.org/abs/2208.11011)  
**Date:** 2022-08-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2208.11011v1/

## Key points

- Designing Deep Neural Networks (DNNs) for edge hardware with limited resources is challenging
- Standard designs have been adopted to facilitate deployment, but not much focus on adapting network topology to fit hardware constraints
- Authors of "Adaptation of MobileNetV2 for Face Detection on Ultra-Low Power Platform" address this issue by adapting MobileNetV2 architecture to fit within hardware constraints while maintaining high accuracy in detecting faces
- Model's performance evaluated against other state-of-the-art models and demonstrated superiority in accuracy and resource usage
- Importance of adapting neural network topologies to fit specific hardware constraints emphasized while achieving high accuracy levels
- Findings provide valuable insights into designing efficient deep learning models that can run on edge devices with limited resources without compromising performance.

## Summary

Designing Deep Neural Networks (DNNs) that can run efficiently on edge hardware with limited resources is a challenge. While standard designs have been adopted by the community to facilitate the deployment of Neural Network models, there is not much focus on adapting the network topology to fit hardware constraints. In this paper titled "Adaptation of MobileNetV2 for Face Detection on Ultra-Low Power Platform," authors Simon Narduzzi, Engin Türetken, Jean-Philippe Thiran, and L. Andrea Dunbar address this issue by adapting one of the most widely used architectures for mobile hardware platforms - MobileNetV2 - to fit within the hardware constraints while still maintaining high accuracy in detecting faces. They evaluate their model's performance against other state-of-the-art models and demonstrate its superiority in terms of accuracy and resource usage. This study emphasizes the importance of adapting neural network topologies to fit specific hardware constraints while still achieving high accuracy levels. The findings provide valuable insights into designing efficient deep learning models that can run on edge devices with limited resources without compromising performance.