IMAGENET-C, ICONS-50
收藏资源简介:
本文介绍了两个新的数据集:IMAGENET-C和ICONS-50,用于评估图像分类器的鲁棒性。IMAGENET-C包含75种常见的视觉损坏,应用于ImageNet对象识别挑战,旨在作为评估图像损坏鲁棒性的通用数据集。ICONS-50则专注于表面变化鲁棒性,包含10,000张来自不同技术公司的图标图像,用于研究对新样式和已知对象的意外实例的鲁棒性。这两个数据集的创建和应用,有助于推动网络学习基本类别结构并稳健地泛化到意外输入的研究。
This paper presents two novel datasets, IMAGENET-C and ICONS-50, for evaluating the robustness of image classifiers. IMAGENET-C includes 75 common visual corruptions applied to the ImageNet Object Recognition Challenge, aiming to serve as a general-purpose dataset for assessing robustness against image corruptions. ICONS-50, by contrast, focuses on robustness to surface variations. It contains 10,000 icon images sourced from various technology companies, and is designed for researching robustness against unseen instances of novel styles and known objects. The creation and application of these two datasets facilitate research on enabling neural networks to learn fundamental category structures and generalize robustly to unexpected inputs.

- 1Benchmarking Neural Network Robustness to Common Corruptions and Surface Variations加州大学伯克利分校 · 2019年



