Researchers Built AI Model to Fasten High-resolution Computer Vision

• Blockchain Council
• October 23, 2023

In the heart of a bustling city, where autonomous vehicles weave through complex intersections, the importance of precise object recognition cannot be overstated. Imagine an autonomous vehicle identifying everything from idling delivery trucks to cyclists hurtling toward intersections. The key to this seamless operation lies in a cutting-edge computer vision model and it goes by the name EfficientViT.

“EfficientViT offers a game-changing reduction in computational complexity,” declares Song Han, an associate professor in the Department of Electrical Engineering and Computer Science at MIT and senior author of the groundbreaking paper. “Our work shows that it is possible to drastically reduce the computation so this real-time image segmentation can happen locally on a device.”

Also Read- A Complete Guide To Generative AI

Semantic segmentation, the task at hand, is the process of categorizing every pixel in a high-resolution image, ensuring that every object is recognized accurately, even in the most demanding scenarios. While previous models were accurate, they stumbled when dealing with high-resolution images, burdened by quadratic computations that made real-time processing impossible on devices like sensors and mobile phones.

The MIT researchers have developed a model that transcends these limitations. EfficientViT introduces a new building block for semantic segmentation models, retaining the precision of existing models while drastically reducing computational complexity. The result? EfficientViT can process high-resolution images up to nine times faster than its predecessors on mobile devices while maintaining or surpassing accuracy.

EfficientViT is not just about efficiency; it’s about striking the perfect balance between efficiency and accuracy. By simplifying the construction of the attention map, a core component of computer vision models, the researchers have reduced the overall computational load without compromising the model’s ability to grasp the global context of an image.

The impact of EfficientViT reaches far beyond autonomous vehicles. Its potential for high-resolution computer vision tasks, including medical image segmentation, is nothing short of revolutionary. Imagine a world where medical professionals can perform intricate image analysis in real time, thanks to this groundbreaking technology.

Furthermore, EfficientViT’s hardware-friendly architecture makes it adaptable to various devices, from virtual reality headsets to the onboard computers of autonomous vehicles. Its applications span image classification and an array of computer vision tasks.

Experts from AMD, Inc., and Oracle recognize the monumental impact of EfficientViT. Lu Tian, Senior Director of AI Algorithms at AMD, Inc., notes that “Their research not only showcases the efficiency and capability of transformers but also reveals their immense potential for real-world applications, such as enhancing image quality in video games.”

Jay Jackson, Global Vice President of Artificial Intelligence and Machine Learning at Oracle, adds, “Oracle Cloud Infrastructure has been supporting his team to advance this line of impactful research toward efficient and green AI.”

EfficientViT signifies a paradigm shift in high-resolution computer vision. It’s not merely about achieving accuracy; it’s about achieving it efficiently. This model heralds a new era for AI applications, making real-time, high-resolution computer vision a reality. Whether it’s guiding autonomous vehicles through bustling city streets or revolutionizing medical diagnostics, EfficientViT is poised to redefine how we perceive the world through the eyes of AI.