Counterfactual Mix-Up for Visual Question Answering

Counterfactuals have been shown to be a powerful method in Visual Question Answering the alleviation of Answering’s unimodal bias. However, existing counterfactual methods tend generate samples that are not diverse or require auxiliary models synthesize additional data. In this regard, we propose more and simple sample synthesis called Counterfactual Mix-Up (CoMiU), which generates image features questions through batch-wise swapping local object- word-level. This efficiently facilitates generation abundant samples, help improve robustness models. Moreover, with creation introduce two robust stable contrastive loss functions, namely Batch-Contrastive Answer-Contrastive loss. We test our on various challenging testing setups show advantages proposed compared current state-of-the-art methods.