Leveraging social media for scalable object detection

摘要

In this manuscript we present a method that leverages social media for the effortless learning of object detectors. We are motivated by the fact that the increased training cost of methods demanding manual annotation, limits their ability to easily scale in different types of objects and domains. At the same time, the rapidly growing social media applications have made available a tremendous volume of tagged images, which could serve as a solution for this problem. However, the nature of annotations (i.e. global level) and the noise existing in the associated information (due to lack of structure, ambiguity, redundancy, and emotional tagging), prevents them from being readily compatible (i.e. accurate region level annotations) with the existing methods for training object detectors. We present a novel approach to overcome this deficiency using the collective knowledge aggregated in social sites to automatically determine a set of image regions that can be associated with a certain object. We study theoretically and experimentally when the prevailing trends (in terms of appearance frequency) in visual and tag information space converge into the same object, and how this convergence is influenced by the number of utilized images and the accuracy of the visual analysis algorithms. Evaluation results show that although the models trained using leveraged social media are inferior to the ones trained manually, there are cases where the user contributed content can be successfully used to facilitate scalable and effortless learning of object detectors.