Parallel Implementation of a Visual Saliency Model

International audienceThe GPIG team from GIPSA-Lab works on visual attention models that mimic the capacity of a primate's visual system to concentrate on a single object in the visual environment. These models tend to reduce the spotlight of focus to a single object or a portion of the visual scene; called the salient regions that guide the attention by locating the spatial discontinuities using different channels like intensity, color, orientation, depth, motion, and many others. The bottom-up visual attention model discussed here is inspired from the primate's visual system, and is modeled all the way from the retina to visual cortex cells. This visual attention model is used to determine explicitly or implicitly where the source of attention lies and the amount of concentration used to contribute or initiate other tasks. This saliency map can be used to predict such areas, finding its applications in video content analysis like surveillance using video analysis, video reframing process to deliver comforting viewing experience on mobile devices, in gaze tracking systems to find ROI (Region of Interest) maps, in video compression. The implementation is demonstrated using a typical webcam interfaced using OpenVIDIA library on an ordinary computer system. Real-time processing is achieved for an image of size 512*512 running at 25fps, that is 12 minutes to process a 17000 frames video