MovieEmotion-IMG: An Emotion Distribution Dataset of Movie Scene Images

Abstract

One of the biggest challenges in sentiment analysis is to predict the complex feelings that humans evoke when viewing images. Due to the limitation of the quality and quantity of the dataset, the current learning-based algorithm mainly analyzes a single dominant emotion. Based on the above, this paper introduces an emotion distribution dataset of movie scene images with rich emotional semantics and professional aesthetic design, called MovieEmotion-IMG. Firstly nine labels, including eight emotional words and a “neutrality” label, were identified, and a set of emotional evaluation rules were designed. Then, the emotion distribution annotation experiment was carried out on 17140 movie images, and an emotion annotation system was developed to assist the annotation work. Finally, two reliability test methods were designed based on the external reliability test method to verify the reliability of the results and ensure the availability of the dataset. This is the first large-scale emotion dataset of movie scene images and annotated distributed emotions to our knowledge.