Autonomous vehicles must be able to quickly and accurately identify objects they encounter, such as delivery trucks parked around corners or cyclists approaching intersections. To achieve this, autonomous cars might use a powerful computer vision model to classify each pixel in high-resolution scene images, ensuring they don't overlook objects that might be obscured in low-quality images. However, this task, known as semantic segmentation, is highly complex, requiring extensive computation, especially with high-resolution images.
According to international media reports, researchers from the Massachusetts Institute of Technology (MIT), the MIT-IBM Watson AI Lab, and other institutions have collaborated to develop a more efficient computer vision model, significantly reducing the computational complexity of the aforementioned task. This model can perform real-time, accurate semantic segmentation on devices with limited hardware resources, such as on-board vehicle computers, enabling autonomous cars to make instantaneous decisions.
Current state-of-the-art semantic segmentation models can directly learn the interactions between every pair of pixels in an image, meaning their computation grows quadratically as image resolution increases. While such models are very accurate, their processing speed is too slow to handle high-resolution images in real-time on edge devices like sensors or smartphones.
Researchers at MIT designed a new component for the semantic segmentation model, boasting capabilities equivalent to the most advanced models of its kind but achieving only linear computational complexity and facilitating efficient hardware operations.
The result of the researchers' work is a novel series of models for high-resolution computer vision. When deployed on mobile devices, these models operate 9 times faster than their predecessors. Importantly, in comparison to alternative solutions, this new model achieves similar, if not better, accuracy.
Not only can this technology assist autonomous cars in making real-time decisions, but it can also enhance the efficiency of other high-resolution computer vision tasks, such as medical image segmentation.
