Aplicação do Modelo de Linguagem BLIP-2 na Geração Automática de Descrições em Vídeos Esportivos

Abstract

The growing demand for audio description in sports videos and the need to make audiovisual content more accessible have driven the development of automatic technologies that can generate accurate and coherent descriptions of sports events. This study investigates the application of the BLIP-2 model in generating automatic descriptions of videos from various sports, such as goal ball, volleyball, and soccer, aiming to capture essential visual details and pro vide consistent and precise descriptions. To assess the quality of these descrip tions, evaluation metrics like METEOR, with scores up to 0,60, and ROUGE-L, reaching 0,63, were used, indicating semantic and structural alignment with manual references. While the model demonstrated effectiveness in captioning popular sports such as soccer, with consistent results and positive metrics, a limitation was observed in less widely covered sports, such as goalball, where the model showed difficulty in achieving accuracy. This is one of the study’s key f indings, highlighting the need for more training data for less popular sports.