R2-Bench

R2-Bench是一个用于评估指称感知模型在扰动下鲁棒性的基准数据集，由卡内基梅隆大学开发。该数据集专注于通过多模态指称指导实现视觉对象的定位，特别是在现实世界场景中可能遇到的干扰，如环境噪声、人为错误和传感器限制。R2-Bench包括一个全面的扰动分类和一个可定制的扰动合成工具箱，用于创建噪声增强的数据集，以评估感知模型的鲁棒性。该数据集适用于五个关键任务，包括指称图像分割、视频对象分割等，旨在解决当前模型在面对现实世界扰动时的脆弱性问题。

R2-Bench is a benchmark dataset developed by Carnegie Mellon University for evaluating the robustness of referential perception models under perturbations. This dataset focuses on visual object localization guided by multimodal referential expressions, particularly addressing disturbances that may be encountered in real-world scenarios, such as ambient noise, human errors, and sensor limitations. R2-Bench includes a comprehensive perturbation taxonomy and a customizable perturbation synthesis toolbox for creating noise-augmented datasets to evaluate the robustness of perception models. This dataset supports five key tasks, including referential image segmentation, video object segmentation, and others, aiming to resolve the vulnerability issue of current models when facing real-world perturbations.