Exploring ubiquitous relations for boosting classification and localization

Abstract Although the weakly supervised learning can effectively avoid the tedious data annotating process of deep learning approaches, the performance is still in urgent need of enhancement. In this paper, we endeavor to mine a ubiquitous and fundamental knowledge—Relation, to boost several existing classification and localization models without changing the original structure. We first propose a universal relation exploring scheme to mine the relations among entities. This scheme can be specialized into different instantiations including object, superpixel and pixel relations to stimulating different learning models. We adopt the object relations on a few-shot classification model to concentrate on the dominant object, and to boost its discriminative capacity. The superpixel relation is utilized to improve the performance of the saliency object detection models. The sensitivity of the pixel relations to the uncertain regions makes it suitable for distinguishing the disputed area in saliency detection results. Our experiments demonstrate that all the three relation instantiations can significantly boost the performance of the state-of-the-art learning models and optimize the visual result.