LABIC - Bioinformatics and Computational Intelligence
Laboratory
Local Repository of Research Datasets
UTFPR-SBD3: Soft Biometrics Dataset for clothing segmentation
Introduction
Soft biometrics is an emerging area of research, mainly due to its extensive applicability in people surveillance. It is related to human characteristics that can be used for people tracking and identification based on appearance, including physical, behavioral or adhered (such as clothing) features. Semantic segmentation of clothes is still a challenge for researchers because of the wide variety of clothing styles, layering, and shapes.
Datasets available for the clothing segmentation task, such as CFPD and Fashionista, are quite small,
they have several annotation errors (at the pixel level) and a high unbalanced class distribution. To overcome these problems, we propose a new benchmark,
named UTFPR-SBD3, containing 4,500 images manually annotated within 18 classes, plus
the background.
All images in the dataset are standardized at 400x600 pixels in RGB, and they were manually annotated using the JS Segment Annotator, a free web-based image annotation tool.
Clothes were grouped into classes, as follows:
Considering that all images in the dataset were annotated in a per-pixel basis,
the plot below shows the proportion of classes and annotated pixels in the
dataset.
Inácio, A.S., Brilhador, A., Lopes, H.S., Semantic segmentation of
clothes in the contex of soft biometrics using deep learning methods.
In: Proceedings of XIV Brazilian Congress on Computational Intelligence,
Belém (PA), 2019.
