Enriching Image Datasets with Unrestrained Emotional Data: A Study with Users

Elicitation of emotions is typically done through the presentation of emotionally salient material, like images or videos, thus requiring reliably annotated datasets. Although there are datasets with emotional information, these only describe either emotional polarities or discrete emotions. The only available dataset with both types of information restrained the participants during the study by separating a priori the images according to their polarity (positive or negative). In this paper, we describe an unrestrained study with 60 participants, where we asked them to rate the polarities and discrete emotions elicited by a set of images. The analysis of the emotional ratings made by the users revealed the most frequent correlations between the basic emotions. Furthermore, the analysis of the ratings’ agreement among participants and existing datasets shows that our results are aligned with the existing ones. As a result of our study, we make available to researchers a more informative picture dataset annotated with emotional polarities and multiple emotions, as a complement to existing datasets.