Latent SVM for Object Localization in Weakly Labeled Videos

We consider the problem of object localization in Internet videos. An Internet video (e.g. YouTube videos) is often associated with a semantic label (also known as a tag) describing the main object present within it. However, the tag does not provide any spatial or temporal information about the main object in the video. Such videos are weakly labelled. Given weakly labelled video with video-level object class tags, our goal is to learn a model that can be used to localize the objects in other videos with such tags. We define a latent SVM based learning framework to tackle this problem. We demonstrate the effectiveness of our method on a dataset composed of videos collected from YouTube.