RobustCLEVR

RobustCLEVR数据集由约翰霍普金斯大学创建，旨在通过模拟各种图像损坏情况来评估对象中心学习方法的鲁棒性。数据集包含多种图像损坏类型，如模糊、云雾、失焦等，每种损坏都基于专家知识和因果模型生成，确保损坏的多样性和真实性。该数据集不仅用于测试模型在标准损坏下的表现，还特别关注模型对因果依赖性损坏的响应，从而更全面地评估模型的鲁棒性。应用领域主要集中在提高计算机视觉系统在复杂和不可预测环境下的稳定性和可靠性。

The RobustCLEVR dataset was created by Johns Hopkins University to evaluate the robustness of object-centric learning methods by simulating various image corruptions. The dataset includes multiple types of image corruptions, such as blurring, fogging, defocusing, etc. Each corruption is generated based on expert knowledge and causal models, ensuring the diversity and authenticity of the corruptions. This dataset is not only used to test the performance of models under standard corruptions, but also specifically focuses on the response of models to causally dependent corruptions, thereby enabling a more comprehensive evaluation of model robustness. Its application fields mainly concentrate on enhancing the stability and reliability of computer vision systems in complex and unpredictable environments.