Un-Attention Grabbing Special Effects (2+1D): the Watchmen TrailerNote: It's interesting that where the special effects are impressive but still fall flat, there is a lack of saliency by this measure. Just my opinion, there. This is a test/demonstration of an extremely fast O(nlogn) and, if I may say so, kinda clever algorithm for determining an information based measure of low level visual saliency. Low level visual saliency attempts to answer the question, "What is important in this scene?" based solely on the scene itself. By contrast, high level saliency might attempt to highlight key-like objects when looking for your keys. Such an algorithm not only helps us understand what may be taking place in our biological visual system but also could act as a preprocessing stage for computer vision/analysis of images. Outside the visual world, this algorithm could be used to prioritize quickly streaming input of any quantitative sort. The subject here was "Movie Trailers" for their availability, lack of copyright enforcement, and uniform size and frame rates. The downside of this set of movies is that they are absolutely full of visual hullabaloo. Still the algorithm did pretty good. Top frame: pixel by pixel statistical saliency. Middle frame: original film clip. Bottom frame: The masked original film clip. Mask generated by thresholding a Gaussian blurred (sd=5pixels) version of the saliency map to reveal the most salient 20% (blurred) of each frame. code pour embarquer la vidéo : >>> http://www.youtube.com/embed/kjsGE5uqRtU <<< |