Edge AI enables artificial intelligence (AI) models to operate directly on devices like smartphones, cameras, and drones. Its key advantage is that it supports faster, real-time decision-making without relying on the cloud. In fact, studies show that using AI on edge platforms can increase operational efficiency by up to 40%. 

Recent advancements in edge AI, particularly in computer vision, made it a central topic at YOLO Vision 2024 (YV24), Ultralytics' annual hybrid event that brings together AI enthusiasts and experts to explore the latest in Vision AI. One of the event's highlights was Sony's keynote presentation, where they showcased their new cutting-edge AI hardware and software solutions. The IMX500 sensor and AITRIOS platform were featured, and Sony demonstrated how these innovations are making it easier and more efficient to deploy Ultralytics YOLO models like Ultralytics YOLO11 and Ultralytics YOLOv8 on edge.

The session was led by Wei Tang, a Business Development Manager who focuses on Sony's imaging solutions, and Amir Servi, an Edge Deep Learning Product Manager with expertise in deploying deep learning models on edge devices. 

In this article, we'll revisit Sony's talk at YV24 and explore how the IMX500 sensor and AITRIOS platform optimize the use of YOLO models for faster, real-time on-edge AI processing. Let’s get started!

Sony's Vision: Democratizing AI on Edge Devices

Wei Tang opened the session by talking about Sony’s goal to make edge AI as accessible as they did with photography years ago. She emphasized how Sony is now focused on bringing advanced Vision AI to more people through edge computing. One of the driving factors behind this is the positive impact edge AI can have on the environment. By processing data directly on devices instead of relying on massive data centers, edge computing helps cut down on energy use and reduce carbon emissions. It’s a smarter, greener approach that fits perfectly with Sony’s commitment to building technology that not only works better but also helps create a more sustainable future.

Wei went on to explain how Sony Semiconductor Solutions, the division of Sony that specializes in imaging and sensing technologies, creates advanced image sensors. These sensors are used in a variety of devices, converting light into electronic signals to capture images. With over 1.2 billion sensors shipped every year, they’re found in nearly half of the world’s mobile phones, making Sony a major player in the imaging industry. 

‍

Building on this expertise, Sony is now taking things further by transforming these sensors from image-capturing devices into smart tools that can process data in real time, enabling AI-powered insights directly on devices. Before we discuss the hardware and software solutions Sony is using to support this shift, let’s understand the edge AI challenges these innovations aim to solve.

Challenges Related to AI Image Processing On Edge Devices

Developing edge AI solutions comes with a few key challenges, especially when working with devices like cameras and sensors. Many of these devices have limited power and processing ability, which makes it tricky to run advanced AI models efficiently.

Here are some of the other main limitations:

• Software complexity: Adapting AI models to work across various edge devices with different hardware configurations can be complex and require adjustments and optimizations.
• Post-processing bottlenecks: There is often a delay when transferring large amounts of data from the device to the host for post-processing. It often consumes more time than the actual AI model inference.
• Data explosion: With many IoT devices constantly generating data, the volume of data that needs to be handled locally can be overwhelming, further straining edge devices.

Getting to Know the Sony IMX500 Intelligent Vision Sensor

Sony IMX500 Intelligent Vision Sensor is a game-changing piece of hardware in edge AI processing. It’s the world’s first intelligent vision sensor with on-chip AI capabilities. This sensor helps overcome many challenges in edge AI, including data processing bottlenecks, privacy concerns, and performance limitations.

While other sensors merely pass along images and frames, the IMX500 tells a whole story. It processes data directly on the sensor, allowing devices to generate insights in real-time. During the session, Wei Tang said, "By leveraging our advanced image sensor technology, we aim to empower a new generation of applications that can enhance everyday life." The IMX500 is designed to meet this goal, transforming how devices handle data directly on the sensor, without needing to send it off to the cloud for processing.

Here are some of its key features:

• Metadata output: Instead of sending full images, it outputs metadata, significantly cutting down on data size, which reduces bandwidth use and costs.
• Enhanced privacy: By processing data on the device, the IMX500 improves privacy, especially in situations where sensitive information is involved, such as human-related computer vision tasks like person counting.
• Real-time processing: The sensor’s ability to handle data quickly means it supports fast, real-time decision-making that enables edge AI applications like autonomous systems.

The IMX500 is not just a camera sensor - it’s a powerful sensing tool that transforms how devices perceive and interact with the world around them. By embedding AI directly into the sensor, Sony is making edge AI more accessible for industries like automotive, healthcare, and smart cities. In subsequent sections, we'll dive deeper into how the IMX500 works with Ultralytics YOLO models to improve object detection and data processing on edge devices.

‍

Sony’s AITRIOS Platform: Simplifying Edge AI

After introducing the IMX500 sensor, Wei Tang expressed that while hardware is crucial, it's not enough on its own to address the full scope of challenges involved in edge AI deployment. She talked about how integrating AI on devices like cameras and sensors requires more than just advanced hardware - it needs smart software to manage it. This is where Sony’s AITRIOS platform comes in, offering a reliable software solution designed to make deploying AI on edge devices simpler and more efficient.

AITRIOS acts as a bridge between complex AI models and the limitations of edge devices. It provides developers with a range of tools for quickly deploying pre-trained AI models. But more importantly, it supports continuous retraining so that AI models can stay adaptable to real-world changes. 

Wei also highlighted how AITRIOS simplifies the process for those who don’t have deep AI expertise, offering flexibility to customize AI models for specific edge AI use cases. It also tackles common challenges like memory constraints and performance drops, making it easier to integrate AI into smaller devices without sacrificing accuracy or speed. 

‍

Optimizing YOLO Models on the IMX500

In the second part of the talk, the mic was passed to Amir, who dove into the technical side of how Sony optimized YOLO models on the IMX500 sensor. 

Amir started off by saying, “YOLO models are edge-enabling and are fairly easy to optimize, thanks to Glenn and the team. I will convince you of that, don't worry." Amir then explained that while a lot of focus typically goes into optimizing the AI model itself, this approach often overlooks a crucial concern: post-processing bottlenecks.

Amir pointed out that in many cases, once the AI model completes its task, the process of transferring data and handling post-processing on a host device can cause significant delays. This back-and-forth data transfer between the device and the host introduces latency, which can be a major obstacle to achieving the best performance.

‍

To tackle this, Amir emphasized the importance of looking at the entire end-to-end system, rather than just focusing on the AI model. With the IMX500 sensor, they discovered that post-processing was the main bottleneck slowing everything down. He shared that the real breakthrough was unlocking on-chip non-maximum suppression (NMS). 

It allowed post-processing to happen directly on the sensor, eliminating the need to transfer large amounts of data to a host device. By running NMS directly on the IMX500, Sony broke through what Amir called the “post-processing glass ceiling,” achieving far better performance and latency reduction.

‍

Next, we’ll take a look at how this innovation helped YOLO models, especially YOLOv8 Nano, run more efficiently on edge devices, creating new opportunities for real-time AI processing on smaller, resource-constrained hardware.

YOLOv8 Models Achieve 4x Speed Boost with Sony’s IMX500

Wrapping up the talk on a high note, Amir demonstrated how they were able to quadruple the performance of the YOLOv8 Nano model by running NMS on edge. He showcased this on a Raspberry Pi 5, which was integrated with the IMX500 AI sensor. Amir compared the performance when post-processing was handled on a host device versus on the IMX500 chip. 

The results clearly showed a major improvement in frames per second (FPS) and overall efficiency when the processing was done on-chip. The optimization made object detection faster and smoother and also demonstrated the practicality of real-time AI processing on smaller, resource-constrained devices like the Raspberry Pi.

Punti di forza

Sony’s IMX500 sensor, the AITRIOS platform, and Ultralytics YOLO models are reshaping edge AI development. On-chip AI processing reduces data transfer and latency while boosting privacy, security, and efficiency. By focusing on the entire system, not just the AI model, these innovations make edge AI more accessible to developers and those without deep AI expertise. As edge AI technology continues to advance, it will likely enable smarter devices, faster decision-making, and stronger privacy protections across a wide range of industries and applications.

Resta in contatto con la nostra comunità per continuare a imparare sull'IA! Dai un'occhiata al nostro repository GitHub per scoprire come possiamo utilizzare l'IA per creare soluzioni innovative in vari settori come l'agricoltura e la produzione. 🚀