Adversarial VQA (AVQA)

Adversarial VQA (AVQA) 是一个通过对抗性人机循环迭代收集的大型视觉问答（VQA）基准数据集，旨在评估VQA模型在面对自然环境中的问题时的鲁棒性。该数据集包含243,000个问题，覆盖了多种图像来源，如网络图像、用户生成图像和电影图像，以确保问题的多样性和挑战性。AVQA的创建过程涉及让标注者设计能够揭示模型弱点的对抗性问题，并将这些问题用于训练更强大的模型。该数据集的应用领域包括提升模型在复杂和多变环境下的性能，以及解决VQA模型在实际应用中的鲁棒性问题。

Adversarial VQA (AVQA) is a large-scale visual question answering (VQA) benchmark dataset collected via an adversarial human-machine iterative loop, which aims to assess the robustness of VQA models when handling natural real-world questions. This dataset contains 243,000 questions spanning diverse image sources including web images, user-generated images, and movie images, ensuring the questions are both diverse and challenging. The construction of AVQA involves having annotators develop adversarial questions that can uncover the weaknesses of current VQA models, and utilize these questions to train more robust VQA systems. The applications of this dataset cover improving model performance in complex and variable environments, as well as addressing the robustness issues of VQA models in practical deployment scenarios.

搜集汇总

背景与挑战

背景概述

Adversarial VQA (AVQA) 是一个用于评估视觉问答模型鲁棒性的基准数据集，采用人机协同方法收集。它包含两个子集：域内的AdVQA和域外的AVQA，总计约81,253张图像，每个图像平均有1.9个人工验证的对抗性问题，每个验证问题提供10个人工编写的答案，以测试模型在对抗性场景下的表现。