Comparative Evaluation of Visual Question Answering with Scene Graph Utilization

Integrating scene graphs enhances Visual Question Answering (VQA) systems' performance in dynamic and diverse real-world applications. This study examines the impact of scene graphs and external knowledge on VQA systems. Using the LLM-based model, LLaVA-v1.5, we trained and evaluated models with and without scene graphs. Additionally, we compared models trained solely with scene graphs to those trained with both scene graphs and external knowledge.