Large Scale Multimodal Classification Using an Ensemble of Transformer Models and Co-Attention

A drawback of these methods is that they consider only global image context, which may contain information Accurate and e cient product classi cation is signi cant for E-irrelevant to the question. To overcome this, some methods commerce applications, as it enables various downstream tasks have proposed visual attention models that attend to local spatial such as recommendation, retrieval, and pricing. Items often contain regions pertaining to a given question, and then perform multimodal textual and visual information, and utilizing both modalities usually fusion to classify answers accurately [4, 19, 21, 22]. More outperforms classi cation utilizing either mode alone. In this recently, dual attention models have been proposed.