FindingEmo-Light — Dataset Only

This repository contains only the annotations and URL list of the FindingEmo dataset, reported in the paper "FindingEmo: A Picture Dataset for Emotion Recognition in the Wild" by Laurent Mertens et al., accepted at the NeurIPS 2024 Datasets and Benchmarks Track.

If you also require the code, please visit https://gitlab.com/EAVISE/lme/findingemo instead.

Obtaining the annotated images

Solution 1: Using the PyPi package

Install the FindingEmo-Light package:

pip install findingemo-light

Then, in a Python script, do

from findingemo_light.paper.download_multi import download_data

download_data(target_dir='./Path/To/Where/You/Want/To/Download/The/Images')

To get the annotations, use

from findingemo_light.data.read_annotations import read_annotations

ann_data = read_annotations()
print(ann_data)

Solution 2: Using the codebase

Clone the repository locally, execute paper/download_multi.py and save images to Config.DIR_IMAGES (which is the default save location).

The annotations are located under data/annotations_single.ann.

Logo

The ./data/Logo folder contains image files with the dataset logo in various formats. If you find this data useful, feel free to use the logo on your poster.

Dataset

The annotations are stored in the data/annotations_single.ann file, which is a simple text file in CSV format. The data/dataset_urls_exploded.json file contains the URLs for the annotated images, with multiple URLs provided for a large number of images. We intend to update this file as we obtain backup URLs for more images.

A Croissant metadata file is also included (./croissant-findingemo.json to allow loading the dataset through the Croissant framework.

Dataset documentation can be found at datasheet/datasheet.md.

Legal Compliance and Privacy

This dataset contains URLs to potentially copyrighted material. If you are a member of a research institution located within the European Union, you are allowed to use this material for non-commercial research purposes by virtue of Title II, Article 3 of the InfSoc directive). If you are not located within the European Union, it is your responsibility to verify with local legislation whether you are allowed to use this material or not.

In case you are the legal copyright holder of any of the images we provide a link to, or are depicted in any of these images, and you do not wish that your material and/or likeness be used for Machine Learning purposes, you can contact either laurent.mertens@kuleuven.be or joost.vennekens@kuleuven.be with the details of the image, and we will immediately remove it from the dataset.

IN NO CASE SHALL THIS DATASET BE USED FOR ANY COMMERCIAL PURPOSE.

Licensing

All data (annotations + list of URLs) is shared under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (see LICENSE_data.md).

Data overview

The data folder contains the CSV/TXT files containing the results per experiment used to generate the results reported in the paper.

The data/Logo folder contains image files with the dataset logo.

The annotations_single.ann file, which is a simple text file in CSV format, contains the image annotations.

The dataset_urls_exploded.json file is the extended version of dataset_urls.txt, which list multiple URLs for a number of images, and is used by the paper/download_multi.py script.

Acknowledgment

This work was funded by KU Leuven grant IDN/21/010.

Author: Laurent Mertens
Mail: laurent.mertens@kuleuven.be