Researchers at DarwinAI and the University of Waterloo have developed a new neural network called AttendSeg, which can perform semantic segmentation on devices with limited memory and compute resources.
Semantic segmentation is a key component of mobile robots, self-driving cars and other AI applications that must detect objects in live image feeds. Being able to perform these functions on device is very important for these applications, where the AI must often work in offline environments or can’t afford the lag caused by sending image data to the cloud.
AttendSeg manages to provide semantic segmentation with accuracy that is on par with large-scale neural networks while cutting down the size of the neural network by two orders of magnitude.
AttendSeg will be presented at this year’s Conference on Computer Vision and Pattern Recognition (CVPR).
For more on AI research: