Page under construction…


Virtual World dataset

The virtual world training set was generated using a plugin for GTA V. It is composed of around 140,000 images, automatically annotated. There are 7 object classes: Bare Head, Helmet, Ear Protection, Welding Mask, Bare Chest, High Visibility Vest, Person. The images have been generated in 10 different locations of the game map, with three weather and time variations for each location. Each image contains an average of 12 people with or without personal protection equipment.

The archive is~131GB, each image is 1088×612. The directory structure is kept as it was generated by the plugin. Inside there are two txt files which divide the dataset in training and validation.

OneDrive Download: (131 GB)

Real World dataset

The real world dataset was manually created by the authors. It is composed of around 220 images, all copyright free and manually annotated. The same object classes are used.

The archive is ~220MB, each image has different size and resolution. Inside there are two directories that separate the images used also for domain adaptation from those used only for testing. You can find also two files, images.txt and pages.txt which provide due attribution to all images.

OneDrive Download: (220 MB)


The above datasets were used for our work on protection equipment detection published in Multimedia Tools and Application.

To cite the article, use the following bibtex:

title={Learning accurate personal protective equipment detection from virtual worlds},
author={Di Benedetto, Marco and Carrara, Fabio and Meloni, Enrico and Amato, Giuseppe and Falchi, Fabrizio and Gennaro, Claudio},
journal={Multimedia Tools and Applications},

Download Authors’ Version