Social Attribute Annotation for Personal Photo Collection

Social attributes for photos, which simply refer to a set of labels {Who, When, Where, What}, are intrinsic attributes of an image. For instance, given a scenery photo without human bodies or faces, we cannot say the photo has no relation with social individuals. In fact, it could have been taken when we went travelling with other friends. To effectively annotate social attributes, we obtain training images from friends' SNS albums. Moreover, to cope with limited training data and organize photos in a feature-effective way, we introduce a batch-based framework, which pre-clusters photos by events. After graph learning based annotation, a post processing step is proposed to refine the annotation result. Experimental results show the effectiveness of the proposed batch-based social attribute annotation framework.