… biased samples and unbiased samples, we design a data classifier module that is pretrained
on VQA v2 train split and validate split and fine-tuned on VQA-… samples classifier (cf. Fig. 3). …

… To show the efficacy and robustness of our method, we … robustness testing VQA datasets
and various different VQA ar… score on VQA v2 validation set, the subset of samples where the …

… Dataset bias in VQA. Prior work has found it is easy to introduce undesirable artifacts during
… where k represents the RBF kernel and ns, nt represent sample size in the source and target …

… Thus, how to train a robust model from the biased dataset remains a major challenge. To
alleviate the challenge in VQA, most existing methods mainly focus on weakening the …

… (ID) data (which is dominated by the biased samples). Therefore, we propose a … robust
VQA models by Making the Most of Biased Samples. Specifically, we construct positive samples …

… the bias factors and clearly monitor the progress of VQA research, … VQA models with all
original and synthesized samples. After training with numerous complementary samples, the VQA …

… robust VQA methods, we stress the language bias in VQA that comes from two aspects, ie,
distribution bias and shortcut bias. We … Counterfactual samples synthesizing for robust visual …

… 1] have been studied, we explore how robust VQA models are to semantic changes in the
images. … On CV-VQA, if the answer on the edited samples is not one less than the prediction on …

… that existing VQA models are not robust enough. Meanwhile, … adversarial examples – data
samples collected using one set … bias, VQA-Rephrasings [39] exposes the brittleness of VQA …

… , we aimed to ensure the robustness and reliability of our proposed VQA model, adequately
… out discrimination information from generated samples for robust visual question answering,” …