Mental Visual Browsing

Jun He, Xindi Shang, Hanwang Zhang, Tat-Seng Chua

Abstract: We present a surprisingly easy-to-use video browser for helping users to pinpoint a specific video shot in mind, within a long video. At each interactive iteration, the only user effort required is to click 1 shot, which most visually relates to the user’s mental target, out of 8 displayed shots. Then, the system updates the browsing model and display another 8 shots for the next iteration. The proposed system is underpinned by a theoretically-sound Bayesian framework that maintains the probabilities of all the video shots segmented from the long video. This framework guarantees that we can find the target shot out of around 1-h video within 3–5 iterations. We believe that our system will perform well in the Video Broswer Showdown game of MMM 2016.