Performance Analysis of Traditional Methods for Visual Question Answering under Computational Resource Constraints

We systematically evaluate traditional Visual Question Answering (VQA) models, focusing on practical aspects such as hyperparameter choices (RNN type, attention size) and feature extraction methods. Our experiments demonstrate that while hyperparameter variations have limited effects on overall accuracy, the use of Bottom-Up features significantly enhances performance compared to conventional CNN-based methods. Crucially, we highlight the practical strengths of traditional VQA methods, particularly their minimal computational requirements, interpretability, and modular architectures, making them ideal for deployment in resource-constrained environments such as embedded and edge devices. This study thus serves as a practical guide and reference for effectively leveraging traditional VQA models in scenarios where computational resources are limited, while providing a meaningful comparison with state-of-the-art multimodal frameworks.