A Survey on Causal Inference in Image Captioning

Although numerous research efforts in the field of Image Captioning (IC) have been conducted, the problem of dataset bias, which causes spurious correlations during training, remains. Recent studies in vision-language (VL) tasks have proposed causal inference as a debiasing method, showing significantly advanced performances. In this paper, we present a comprehensive survey of state-of-the-art IC models based on causal inference and categorize them according to their adjustment. We aim to provide an understanding of how causal inference can be utilized in IC. Furthermore, since IC is one of the foundations in VL tasks, we believe that this review will also help to get insight when applying causal inference in other VL tasks.