2FPS. Automatic Human Visual Attention Modeling for Video. Automatic video eyetracking and advertisement optimization using computer vision only. Heatmap generation using computer vision only. Example of an eyetracking "heatmap" that shows how much users are likely to look at different parts of a video. Areas where users looked the most are colored red-black; the yellow-green areas indicate fewer fixations, followed by the least-viewed blue areas. Gray areas didn't attract any fixations. No expensive eye tracking studies needed. Both low level (color, edge, texture, ...) and high level (text recognition and understanding, object recognition and understanding, ) analysis is done. Direction of human attention is computed using 80000 training (human-based) samples. Very fast processing time - 2 fps. Accurate and robust (87% compared to eye tracking results on the same database). For more information please contact tsishkoudv at gmail.com.
very interesting. where can one get 80000 training (human-based) samples?
rafaelmizrahi 3 years ago
Thanks. > 3/4 of the database = youtube-like videos (keywords: eyetracking, heatmaps, ...)with groundtruth info (blue dot, red cross). All you need - is to remove groundtruth using image processing to have both. Next do learning/testing.
Each 3min video of 15fps gives you 2700 static images + attention direction dynamics as a bonus :) 80000 images = 90min of video.
tsishkou 3 years ago