LAION-C
收藏资源简介:
LAION-C是一个为评估大规模视觉模型在非分布(OOD)数据上的鲁棒性而设计的基准数据集。它包含六种新颖的失真类型,这些失真即使在网络规模数据集上也是OOD。该数据集旨在测试模型是否能够超越其训练分布,并在具有挑战性的环境中表现良好。数据集包括273张图片,每张图片都有五种强度级别的六种失真类型。此外,该数据集还包含了通过心理物理实验获得的19位参与者的11000次试验结果,以提供一个稳健的人类OOD泛化基线。
LAION-C is a benchmark dataset developed to evaluate the robustness of large-scale visual models on out-of-distribution (OOD) data. It includes six novel distortion types that are classified as OOD even on web-scale datasets. This dataset aims to test whether models can exceed their training distributions and deliver strong performance in challenging environments. The dataset comprises 273 images, each of which is processed with six distortion types across five intensity levels. Furthermore, the dataset contains 11,000 trial results from 19 participants obtained via psychophysical experiments, providing a robust human baseline for OOD generalization.

- 1通过德国马克斯·普朗克智能系统研究所 · 2025年



