Rethinking Data Augmentation for Robust Visual Question Answering

AbstractData Augmentation (DA) — generating extra training samples beyond the original set has been widely-used in today’s unbiased VQA models to mitigate language biases. Current mainstream DA strategies are synthetic-based methods, which synthesize new by either editing some visual regions/words, or re-generating them from scratch. However, these synthetic always unnatural and error-prone. To avoid this issue, a recent work composes augmented randomly pairing pristine images other human-written questions. Unfortunately, guarantee have reasonable ground-truth answers, they manually design of heuristic rules for several question types, extremely limits its generalization abilities. end, we propose Knowledge Distillation based Data VQA, dubbed KDDAug. Specifically, first relax requirements image-question pairs, can be easily applied any type. Then, knowledge distillation (KD) answer assignment generate pseudo answers all composed robust both in-domain out-of-distribution settings. Since KDDAug is model-agnostic strategy, it seamlessly incorporated into architecture. Extensive ablation studies on multiple backbones benchmarks demonstrated effectiveness abilities KDDAug.KeywordsVQAData augmentationKnowledge