The VisEmoCom dataset contains annotations on the emotions expressed by characters in comic books.
Despite the fact that text represents a significant part of the content, the transcriptions of the dialogues are not included.
The main goal of this dataset is to focus on visual elements such as facial expressions or symbols, intentionally drawn by the artists to communicate a piece of information.
For each targeted character, multiple annotators were asked their interpretation on the expressed emotion.
This dataset was built on images that come from different public datasets: Manga109 dataset, EmoRecCom challenge and IMCDB dataset.
Images gathered from each subset were split into train and test sets depending on the book they were extracted from.
The following table displays the size of each subset and split.
We provided samples and questionnaires to a set of annotators.
For each sample, annotators were asked to select which emotions they interpret on a panel with the character of interest surrounded with a bounding box.
Due to the subjective nature of the topic, each image was processed by multiple annotators in order to gather various points of view.
Annotators had to select one or multiple emotions among this set:
{Anger, Disgust, Fear, Joy, Neutral, Sadness, Surprise}.
The resulting annotations are stored in CSV files, available on this link.
For Manga109, users are required to go through the official application on the Manga109 website to download the images and the bounding boxes coordinates.
Instead, the dataset only provides for each sample:
EmoRecCom images are panels, not pages, so each sample provides face and body boxes.
As they belong to the public domain, we can provide a zip archive with the panel images, downloadable here.
IMCDB data is available on this repository.
If you use this dataset for your own research, please consider citing our paper:
@InProceedings{theodose2024visemocomic,
author={Th{\'e}odose, Ruddy and Burie, Jean-Christophe},
title={VisEmoComic: Visual Emotion Recognition in Comics Image},
booktitle={Pattern Recognition},
year={2025},
}
This work is supported by the Research National Agency (ANR) in the framework of the 2017 LabCom program (ANR 17-LCV2-0006-01) and the Region Nouvelle Aquitaine in the framework of the EmoRecCom project. This work benefited from access to the computing resources of the L3i laboratory, operated and hosted by the University of La Rochelle and the computing resources of the “CALI 3” cluster, operated and hosted by the University of Limoges, part of the HPC network in the Nouvelle-Aquitaine Region. Both are financed by the State and the Region.