Enhancing robust VQA via contrastive and self-supervised learning

Runlin Cao, Zhixin Li, Zhenjun Tang, Canlong Zhang, Huifang Ma

Abstract: Highlights•CSL-VQA combines self-supervised and contrastive learning for unbiased VQA training.•BSD distinguishes biased and unbiased samples, boosting generalization in VQA.•PNSG module balances dataset bias without manual annotation, enhancing VQA.•Achieves state-of-the-art 62.30% on VQA-CP v2 while preserving VQA v2 accuracy.