Detecting visual sentiment evoked by art images

다국어 초록 (Multilingual Abstract)

Detecting emotions evoked by art has been receiving great attention recently. Although previous works provide a variety of datasets consisting of art images and corresponding emotion labels, little attention has been paid to the continuous and dimensional characteristics of human emotions, especially in the domain of art. We propose a dataset for detecting visual sentiment from art images, D-ViSA, whose labels consist of both categorical and dimensional emotions which can be implemented in a wide range of visual sentiment analysis research regarding art. We compare several deep learning baselines in two specific tasks, single-feature, and multi-feature dimensional emotion detection. Furthermore, through scalability evaluation by comparing D-ViSA with another art image dataset, we reveal that the final constructed dataset is labeled appropriately compared to the previous bag of knowledge. Our experiments lead to the conclusion that our dataset is plausible for both dimensional emotion detection tasks with deep learning baselines and show improved performance on multi-feature task. We assume that our dataset contributes to the field of artwork analysis and provides insights into human emotions evoked by art. The dataset is publicly available at https://github.com/dxlabskku/D-ViSA

목차 (Table of Contents)

• 1 Introduction 1
• 2 Literature Review 6
• 2.1 Visual Sentiment Analysis 6
• 2.2 Employing Dimensional Emotion 8
• 2.3 Art Emotion Dataset 10
• 3 D-ViSA 13
• 3.1 Data Collection and Annotation 13
• 3.2 Data Validation 15
• 3.3 Descriptive Statistics of Datasets 18
• 4 Experiments 21
• 4.1 Detecting Emotions Evoked by Art 21
• 4.2 Experimental Settings 23
• 4.3 Evaluation Metrics 25
• 4.4 Baseline Models 25
• 5 Results 28
• 6 Scalability Evaluation 31
• 7 Discussion & Conclusion 33
• References 36
• 초 록 48