ICE: A Benchmark for Human-like Image Commenting
Qiuyuan Huang, Yandong Guo, Oussama Elachqar, Kedhar Nath Narahari, Ito Kazushige, Donald Brinkman, Xiaodong He, Lei Zhang, Yu‐Ting Kuo · 2018
With the growing usage of images in conversations with Artificial intelligent(AI) agents, it has become imperative for AI agents to respond to images in more human-like manner. To achieve this, agents should not only describe the facts of an image but also express emotions and opinions about it. Despite phenomenal progress in the past several years, most existing image captioning systems mainly describe the concepts, objects or actions present in an image. Hence, they have an inherent drawback of neither being human-like nor facilitating deeper engagement. To overcome this limitation, we introduce a novel task of Image CommEnting (ICE) with the objective of generating a human-like comment given an image. To further facilitate research in this domain, we release a dataset, ICE v1, which includes one million images and their corresponding real human comments. We also explore how different existing image commenting techniques perform with this dataset and establish strong baselines for this tasks.