News

World’s First TinyML Vision Challenge Crowns 2021 Winners

October 13, 2021 by Jake Hertz

Competitions are always an exciting way to test the limits and ingenuity of both technology and the designer. This week the winners of the first inaugural TinyML Vision Challenge were announced.

TinyML is a hot field of research and development right now, with many people in academia and industry dedicating time and money to the cause. The field is, however, no longer constrained to only professionals. As it continues gaining popularity, online communities have continued to emerge, providing resources, tutorials, and forums for education and collaboration. 

 

An example platform for TinyML.

An example platform for TinyML. Image used courtesy of Edge Impulse and TinyML Foundation

 

Earlier this year, the TinyML Foundation announced its first inaugural TinyML contest, where engineers worldwide were welcome to compete to create solutions to pertinent industry-grade challenges. This week the winners were announced, showcasing the potential that TinyML has to offer our world. 

This article will give some context of the competition and then look at the winning submission to get a feel for one of the many applications of TinyML. 

 

What is TinyML? 

TinyML is a subfield of Machine Learning (ML) that pursues enabling ML applications on cheap, resource-constrained devices like microcontrollers or digital signal processors (DSPs). 

The objective of the TinyML is to bring machine learning to the edge, where battery-powered, embedded devices can perform ML tasks with real-time responsivity. This effort is extraordinarily multidisciplinary, requiring optimization and maximization from fields, including hardware, software, data science, and machine learning. 

 

A suite of popular TinyML computing devices.

A suite of popular TinyML computing devices. Image used courtesy of Hackster.io 

 

The world of TinyML is becoming increasingly important as more and more interconnected devices are deployed that rely on ML tasks that require edge computing at low power.

Showing just how much ML is growing in popularity and use, this contest aims to demonstrate that with TinyML, there are many applications to focus your ingenuity. 

 

Contest Background 

For the contest, the TinyML Foundation challenged engineers worldwide to come up with a solution to some industry-grade challenge of their choosing. 

Within this, the contest focused specifically on computer vision-based applications. Contestants were allowed to utilize any hardware platform and software framework they desired, with the ultimate goal of creating a battery-operated, ultra-low-power project. 

Overall, submissions were graded based on their ability to address a problem within a given industry. 

 

A First Place Fire Detection System

The first-place winner of this year's contest was Team Sol, whose project was entitled "TinyML Aerial Forest Fire Detection." 

Inspired by the growing wildfires on the west coast of the United States, this project sought to develop an early detection system to help stop wildfires in their early stages. The final product of this project was a power-efficient, autonomous RC plane that uses an onboard camera to detect and report wildfires via satellite modem.

 

The plane is being held by a member of Team Sol.

The plane is being held by a member of Team Sol. Image used courtesy of Felleke et al and the TinyML Foundation

 

For their TinyML hardware, the team used an Arduino Nano BLE 33, an Arducam OV2640, and a LiPO battery. 

On the software side of things, the team leveraged the pre-existing "Flame Dataset," which consists of aerial photographs of landscapes with and without fires burning. The team used this dataset to train a convolutional neural network (CNN) until they reached a validation accuracy of 96%. Then, via quantization, they were able to fit the model onto the Arduino Nano 33 BLE. 

To make their project as low power as possible, the team put the Arducam OV2640, Arduino Nano 33 BLE, and satellite module to sleep for 500 ms after each photo it took. Overall, the TinyML portion of the device, including GPS, consumed ~0.3% of the plane's total power. The plane was able to last at least 90 minutes in the air at a speed of 30 mph. 

 

An Eye on the Future 

This contest produced many very impressive results, with second and third place being no less exciting. The other winners here were a computer vision-based multiparameter monitoring and diagnostics system and a low-power sewer fault detection system, respectively.

Contests like this are great for a budding field like TinyML because they stimulate interest, raise awareness, and provide examples of what the technology can do. In the future, we should expect to see TinyML start to make its way into more and more of our daily lives.

 

Featured image used courtesy of the TinyML Foundation and Felleke et al