SentiImgBank: A Large Scale Visual Repository for Image Sentiment Analysis

In contrast to existing methods which detect sentiment directly from low/high-level features, we construct a large-scale visual repository, namely SentiImgBank, with the aim of providing a bench-marking visual sentiment lexicon. More specially, the SentiImgBank consists of 24 categories, 5,487 adjective-noun pairs (ANPs), in total of 648,946 images that are collected from social media such as Twitter. In view that ANPs might express different sentiments in different contexts, i.e., contextuality, SentiImgBank annotates the discrete sentiment and emotion scores instead of directly defining the golden label. Hence, each image is associated with ten numerical scores, where three of them are sentiment scores, the remaining seven scores denote different emotions. To alleviate the manually annotation cost, a committee of 15 pre-trained language models based classifiers is proposed to automatically produce the sentiment and emotion scores. Finally, the strong baselines are proposed to evaluate the potential of SentiImgBank. We hope this study provides a publicly available resource for visual sentiment analysis. The full dataset will be publicly available for research (https://github.com/ anonymity2024/SentiImgBank).