2026-08-182026-08-182026-07-10ARAUJO, Márcio Gabriel dos Santos de. Visão computacional e IA para medidas corporais: uma análise comparativa entre MNASNet e SnapMeasurementAI. Orientador: Daniel da Conceição Pinheiro; Coorientador: Raphael Barros Teixeira. 2026. 64 f. Trabalho de Curso (Bacharelado em Engenharia Elétrica) – Faculdade de Engenharia Elétrica, Campus Universitário de Tucuruí, Universidade Federal do Pará, Tucuruí, 2026. Disponível em: https://bdm.ufpa.br/handle/prefix/9962. Acesso em:.https://bdm.ufpa.br/handle/prefix/9962This study investigates the application of artificial intelligence and computer vision techniques for estimating human body measurements from images, focusing on their use in clothing size recommendation systems for e-commerce. The rapid growth of online fashion retail has highlighted significant challenges related to incorrect size selection, which directly impacts customer experience and operational costs. In this context, a comparative analysis was conducted between two approaches: MNASNet, a model trained on real-world data, and SnapMeasurementAI, a commercially available solution. The methodology involved the use of datasets containing frontal and lateral images combined with anthropometric information, as well as model evaluation using the Mean Absolute Error (MAE) metric and clothing size classification based on commercial sizing charts. The results revealed significant differences between the evaluated approaches. MNASNet achieved an MAE of 3.7 cm for female subjects and 6.9 cm for male subjects, indicating better performance for women and evidencing the presence of bias in the model. In practical evaluations based on commercial sizing charts, the model achieved a 50% accuracy rate for chest and waist size classification using Insider® clothing tables, but only 17% accuracy for hip measurements, with average errors reaching 10.1 cm. In contrast, SnapMeasurementAI demonstrated more balanced performance across genders and lower overall variability, indicating greater robustness and generalization capability. The results also showed that converting estimated body measurements into clothing size categories remains one of the main practical challenges, since variations of only a few centimeters can lead to different size recommendations. It is concluded that, although both approaches demonstrate potential for size recommendation systems, limitations related to accuracy, robustness, and generalization still affect their practical applicability. The findings highlight the need for further improvements in data balancing, bias reduction, and the development of more reliable strategies for converting body measurements into commercial clothing sizes.Acesso AbertoInteligência artificialVisão computacionalMedidas corporaisEstimativa antropométricaRecomendação de tamanhosArtificial intelligenceComputer visionBody measurementsAnthropometric estimationSize recommendation systemsCNPQ::ENGENHARIASCNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAOTrabalho de Curso - Graduação - MonografiaAttribution-NonCommercial-NoDerivs 3.0 Brazil