The San Diego Stingray is a special purpose architecture which targets the evolving computer vision and machine learning application domains. Since the 1980's, computer vision research has become increasingly decoupled from considerations of the underlying capabilities that modern VLSI provide. As the end of Moore's law looms on the horizon (e.g., at 2 nm), computer vision will increasingly need to incorporate the computation and power constraints of technology in order to retain its applicability. Stingray attempts to bridge this gap by creating a path from MATLAB (the language of choice for vision researchers) to VLSI that will make exploration of real-time computer vision more practical. Ideally, Stingray will be used both to accelerate current-day computer vision algorithms and to influence the course of the field in the future.

 

An example application that Stingray will be able to perform is real-time visual scene text recognition. As shown in the above figure, this system will detect and interpret text in the input image. This system will help individuals with sight disabilities, travelers in foreign countries, and public service workers in diverse environments (e.g., firemen in Chinatown). Stingray will not only boost response time of the current text recognition systems, but also enable to use more advanced detection or recognition algorithm to increase accuracy.

 

Another example application is visual SLAM (simultaneous localization and mapping). Although current robotics technology implements realtime SLAM with GPS (global positioning system) signals, and IMU (Inertia Measurement Unit) and laser sensors, these systems have their limits. For example, GPS signals are often lost in big cities or indoors, and are not available on the Moon or on Mars. Localization devices such as IMU and odometers can fail when the vehicle sticks in a sand dune or slips on ice. Visual SLAM can help resolve these problems, but is often bottlenecked by computational performance. The Stingray system will enable realtime VSLAM within a fixed power budget.