Nombre: ANTONIO RICARDO ALEXANDRE BRASIL
Fecha de publicación: 24/11/2025
Junta de examinadores:
Nombre |
Papel |
|---|---|
| CLARIMAR JOSÉ COELHO | Examinador Externo |
| DANILO SALES BOCALINI | Examinador Interno |
| FILIPE WALL MUTZ | Examinador Interno |
| KARIN SATIE KOMATI | Examinador Externo |
| PATRICK MARQUES CIARELLI | Presidente |
Sumario: Somatotype is a method that classifies the human body as the combination of three
components: endomorph, ectomorph, and mesomorph, with applications in both health
and sports domains. Currently, the most common approach to estimating somatotype
relies on anthropometric measurements collected directly from the human body, which
requires trained specialists, is time-consuming, and prone to human error. In view of these
limitations, this work proposes the use of convolutional neural networks as an approach for
estimating and classifying somatotype from digital images, exploring three datasets: two
used in previous studies, one containing images of 46 individuals and another with 339
body images obtained through 3D scanning, and a dataset collected for the present study,
comprising images and measurements of 122 individuals. Experiments were conducted
focused on tasks of classification and estimation of the numerical values of the somatotype
components. Preliminary results indicate that it is possible to achieve up to 97% accuracy
in predicting an individual’s predominant somatotype and to estimate the somatotype
components with a mean squared error (MSE) as low as 0.02, revealing promising findings
and opening new perspectives for estimating somatotype using computational methods.
