Application of the MUTAN Model to the VQA Task

This research compares Simple-VQA-V2 and Complex-VQA-V2 models in the context of Visual Question Answering (VQA). Simple-VQA-V2 focuses on basic elements and counting, while Complex-VQA-V2 incorporates MLB fusion and Mutan fusion. Both models exhibit similar patterns of accuracy improvement during training, starting with a gradual increase and then reaching a plateau. However, Complex-VQA-V2 demonstrates more variability and lower accuracy compared to Simple-VQA-V2. Surprisingly, the advanced fusion methods do not significantly boost the performance of Complex-VQA-V2. The study underscores the importance of understanding dataset characteristics and making informed choices regarding fusion techniques for multimodal VQA models. It also underscores the potential of Mutan fusion for object counting in images.