A Somatotype Classification Approach Based on Generative AI and Frontal Images of Individuals

Somatotype is a definition that classifies human bodies according to their shape, and its determination is crucial for evaluating and improving athletic performance. However, the traditional method used for its estimation has some limitations: measurements are directly collected from the human body, specialists are required, and it is time-consuming. In this paper, two approaches are evaluated: GPT-4’s performance in classifying the predominant somatotype using only human body frontal images, and a proposed multimodal architecture that combines the textual descriptions provided by GPT-4 with these images, using pre-trained MiniLM models for the textual information and pre-trained Vision Transformer (ViT) with transfer learning for images. Frontal images were collected from 50 volunteers aged between 18 and 65 years, with 10 anthropometric measurements following Heath and Carter protocol. The results show that GPT-4 achieved an average accuracy of 78% in somatotype classification for the first approach. By combining the textual descriptions with the frontal images in the multimodal model, the average accuracy was of 97%, with a precision of 94% and recall of 89%. These results indicate that combining multimodal language models with deep learning improved accuracy in automatic somatotype classification compared to GPT-4 classification. The results are promising and open up new possibilities in the classification of somatotypes using only digital images.