Picture News Collection: A Dataset for Automatic Picture News Thumbnail Selection

摘要

Picture news has become more and more popular among online news in recent years. As the first impression to viewers, thumbnail plays a very important role in picture news. However, it is time consuming to manually select thumbnails for a huge amount of picture news. In this paper, we introduce a new task of automatic picture news thumbnail selection. Given a piece of picture news containing a set of images, this task is to select several appropriate images from the picture news as candidate thumbnails. To this end, we present a large publicly available image dataset for this task, called Picture News Collection (The Picture News Collection 0.1 version can be publicly available online at https://github.com/anonymity01/Picture-News-Collection.). The Picture News Collection contains more than 4 million images of 347,731 picture news from two famous news websites, Sina News and NetEase News. Selecting good enough thumbnails is complicated and needs to consider many aspects, such as attraction, hot topics, content integrity, etc. In order to select appropriate candidate thumbnails, we propose an attention-based thumbnail selection model, and the experimental results comparing with three image classification based baselines show that our proposed methods outperform the baselines. We introduce the automatic picture news thumbnail selection task and the dataset to encourage further studies of this challenge.