A high-speed multi-GPU implementation of bottom-up attention using CUDA

In this paper a novel implementation of the saliency map model on a multi-GPU platform using CUDA technology is presented. The saliency map model is a well-known computational model for bottom-up attention selection and serves as a basis of many attention control strategies of cognitive vision systems. A real-time implementation is the prerequisite of an application of bottom-up attention on mobile robots and vehicles. Parallel computation on graphics processing unit (GPU) provides an excellent solution for this kind of compute-intensive image processing. Running on 1 to 4 NVIDIA GeForce 8800 (GTX) graphics cards a frame rate of 313 fps at resolution of 640 times 480 is achieved, which is approximately 8.5 times faster than the standard implementations on CPUs. The implementation is also evaluated using a high-speed camera at 200 Hz. Using two GPUs only 2 ms extra computational time for the saliency map generation in addition to the camera capture time is required for images of 640 times 480 pixels.