VisE-D: Visual Event Classification Dataset

# VisE-D: Visual Event Classification Dataset This repository contains the *Visual Event Classification Dataset (VisE-D)* introduced in the paper: > Eric Müller-Budack, Matthias Springstein, Sherzod Hakimov, Kevin Mrutzek, and Ralph Ewerth: "Ontology-driven Event Type Classification in Images". In: *IEEE Winter Conference on Applications of Computer Vision (WACV)*, IEEE, 2021. ## Content - VisE-D.tar.gz: - List of image URLs with corresponding meta information for the: - *VisE-D* train dataset - *VisE-D* validation dataset - *VisE-Bing* test dataset - *VisE-Bing* test dataset - Different versions of the *Visual Event Ontology (VisE-O)*: - Initial Ontology - Disambiguated Ontology - Disambigauted Ontology for *Event Nodes* related to a minimum of ten *Events* - Refined Ontology - VisE-C.tar.gz: Classification baseline (denoted as C) based on the ResNet-50 architecture trained on *VisE-D*. - VisE-CO_cel.tar.gz: Ontology-driven approach with redundancy removal (denoted as CO^{cel}_{6\omega}-RR) based on the ResNet-50 architecture trained on *VisE-D*. - VisE-CO_cos.tar.gz: Ontology-driven approach (denoted as CO^{cos}_{\delta}) based on the ResNet-50 architecture trained on *VisE-D*. - RED.tar.gz: Ontologies and splits for the *RED* benchmark datasets (Ahsan et al. 2017) - SocEID.tar.gz: Ontologies and splits for the *SocEID* benchmark datasets (Ahsan et al. 2017) - WIDER.tar.gz: Ontologies and splits for the *WIDER* benchmark datasets (Xiong et al. 2015) ## Source Code The source code to reproduce our results can be found on our GitHub page: https://github.com/TIBHannover/VisE