ARAN: Age-Restricted Anonymized Dataset of Children Images and Body Measurements

Precisely estimating a child’s body measurements and weight from a single image is useful in pediatrics for monitoring growth and detecting early signs of malnutrition. The development of estimation models for this task is hampered by the unavailability of a labeled image dataset to support supervised learning. This paper introduces the “Age-Restricted Anonymized” (ARAN) dataset, the first labeled image dataset of children with body measurements approved by an ethics committee under the European General Data Protection Regulation guidelines. The ARAN dataset consists of images of 512 children aged 16 to 98 months, each captured from four different viewpoints, i.e., 2048 images in total. The dataset is anonymized manually on the spot through a face mask and includes each child’s height, weight, age, waist circumference, and head circumference measurements. The dataset is a solid foundation for developing prediction models for various tasks related to these measurements; it addresses the gap in computer vision tasks related to body measurements as it is significantly larger than any other comparable dataset of children, along with diverse viewpoints. To create a suitable reference, we trained state-of-the-art deep learning algorithms on the ARAN dataset to predict body measurements from the images. The best results are obtained by a DenseNet121 model achieving competitive estimates for the body measurements, outperforming state-of-the-art results on similar tasks. The ARAN dataset is developed as part of a collaboration to create a mobile app to measure children’s growth and detect early signs of malnutrition, contributing to the United Nations Sustainable Development Goals.