# TinyissimoYOLO: A Quantized, Low-Memory Footprint, TinyML Object Detection Network for Low Power Microcontrollers

**Authors:** Julian Moosmann, Marco Giordano, Christian Vogt, Michele Magno  
**arXiv:** [2306.00001](https://arxiv.org/abs/2306.00001)  
**Date:** 2023-05-22  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.00001v2/

## Key points

- TinyissimoYOLO is an object detection network designed for low-power microcontrollers
- Highly flexible and ultra-lightweight with a quantized and memory-efficient architecture
- Optimized for real-time object detection on embedded microcontrollers with power requirements in the milliwatt range and limited memory capacity of less than 0.5MB
- Deployed on the MAX78000 microcontroller, achieving high frame rates and ultra-low energy consumption
- Inference efficiency surpasses 106 MAC/Cycle, showcasing exceptional performance capabilities
- Can be trained for multi-object detection tasks but demonstrated object detection with up to 3 classes due to compact size
- Utilized quantization-aware training techniques and implemented 8-bit quantization when deploying on various microcontrollers including STM32H7A3, STM32L4R9, Apollo4b, and the MAX78000's CNN accelerator

## Summary

In this paper, authors Julian Moosmann, Marco Giordano, Christian Vogt, and Michele Magno introduce TinyissimoYOLO - a cutting-edge object detection network designed for low-power microcontrollers. This innovative network is highly flexible and ultra-lightweight with a quantized and memory-efficient architecture. It is specifically tailored for applications where power consumption and memory constraints are critical factors. TinyissimoYOLO has been optimized for real-time object detection on embedded microcontrollers with power requirements in the milliwatt range and limited memory capacity of less than 0.5MB for storing convolutional neural network (CNN) weights. The network boasts 422k parameters and has been evaluated to take advantage of CNN accelerators for enhanced performance. The authors successfully deployed TinyissimoYOLO on the MAX78000 microcontroller, achieving impressive results such as a high frame rate of up to 180fps and an ultra-low energy consumption of only 196μJ per inference. The network's inference efficiency surpasses 106 MAC/Cycle, showcasing its exceptional performance capabilities. While TinyissimoYOLO can be trained for multi-object detection tasks, the authors demonstrate object detection with up to 3 classes due to the network's compact size. They utilized quantization-aware training techniques and implemented 8-bit quantization when deploying the network on various microcontrollers including STM32H7A3, STM32L4R9, Apollo4b, and the MAX78000's CNN accelerator. Overall, this paper presents a comprehensive evaluation of TinyissimoYOLO's performance across different platforms and highlights its potential for enabling efficient object detection in resource-constrained environments. The research contributes valuable insights into developing advanced machine learning solutions tailored for low-power microcontroller applications.