Building a Large Scale Dataset for Image Emotion Recognition: The Fine Print and The Benchmark

• DOI: 10.1609/aaai.v30i1.9987
• Publication Date: 2022-06-24T06:12:22Z
• Authors: Quanzeng You; Jiebo Luo; Hailin Jin; Jianchao Yang
• Keywords: Benchmarking; Unavailability; Data set; Benchmark (surveying)

Psychological research results have confirmed that people can different emotional reactions to visual stimuli. Several papers been published on the problem of emotion analysis. In particular, attempts made analyze and predict people's reaction towards images. To this end, kinds hand-tuned features are proposed. The reported several carefully selected labeled small image data sets promise such features. While recent successes many computer vision related tasks due adoption Convolutional Neural Networks (CNNs), analysis has not achieved same level success. This may be primarily unavailability confidently relatively large for work, we introduce a new set, which started from 3+ million weakly images emotions ended up 30 times as current largest publicly available set. We hope set encourages further also perform extensive benchmarking analyses using state art methods including CNNs.