Learning deep feature fusion for group images classification

With the rapid development of social media, people tend to post multiple images under the same message. These images, we call it group images, may have very different contents, however are highly correlated in semantic space, which refers to the same theme that can be understood by a reader, easily. Understanding images present in one group has potential applications such as recommendation, user analysis, etc. In this paper, we propose a new research topic beyond the traditional image classification that aims at classifying a group of images in social media into corresponding classes. To this end, we design an end-to-end network which accepts variable number of images as input and fuses features extracted from them for classification. The method are tested on two newly collected datasets from Microblog and compared with a baseline method. The experiment demonstrates the effectiveness of our method.