Title Page

Contents

Abstract 10

Chapter 1. Introduction 12

Chapter 2. Literature Review 17

2.1. Visual Sentiment Analysis 17

2.2. Employing Dimensional Emotion 19

2.3. Art Emotion Dataset 21

Chapter 3. D-ViSA 24

3.1. Data Collection and Annotation 24

3.2. Data Validation 26

3.3. Descriptive Statistics of Datasets 29

Chapter 4. Experiments 32

4.1. Detecting Emotions Evoked by Art 32

4.1.1. Single-feature Dimensional Emotion Detection 33

4.1.2. Multi-feature Dimensional Emotion Detection 34

4.2. Experimental Settings 34

4.3. Evaluation Metrics 36

4.4. Baseline Models 36

Chapter 5. Results 39

Chapter 6. Scalability Evaluation 42

Chapter 7. Discussion & Conclusion 44

References 47

초록 59

Table 3.1. Overall statistics of each emotion category 29

Table 5.1. The overall performance of single-feature regression and multi-feature regression models on D-ViSA. Pearson's correlation coefficients were derived to evaluate the performance. 39

Figure 1.1. Examples with images and their corresponding categorical emotions and VAD levels 13

Figure 3.1. Summary of the annotation procedures (VAD levels) 28

Figure 3.2. Pearson's correlation coefficients among VAD levels 30

Figure 3.3. Pearson's correlation coefficients of mean VAD levels between eight emotion categories 31

Figure 4.1. Single-feature regression model framework 33

Figure 4.2. Multi-feature regression model framework 35

Figure 5.1. Sample images and predicted VAD levels of multi-feature PDANet model. 40

Figure 6.1. Performance comparison of prior single-feature baseline models on D-ViSA and those pre-trained models tested on MART dataset 43