Object instance search in videos
Jingjing Meng, Junsong Yuan, Gang Wang, Jianbo Xu · 2013
In this paper, we propose a novel approach for object instance search in videos. Employing discriminative mutual information score and inferring the location of target object centers from matched local feature descriptors using Hough voting, we achieve robust matching and per-frame localization despite orientation and scale variations. We then leverage Max-Path search [1] to efficiently find the globally optimal spatio-temporal trajectory of the object center in each video sequence. Experimental results on a collection of mobile-captured videos in real-world environments demonstrate the effectiveness and accuracy of our method.