NICO

NICO 数据集专为非 I.I.D. 设计。或 OOD（分布外）图像分类。它模拟了一个真实世界的设置，即测试分布可能会导致训练分布的任意偏移，这违反了传统的 I.I.D.大多数机器学习方法的假设。数据集能够很好地支持的典型研究方向包括但不限于迁移学习或领域适应（当测试分布已知时）和稳定学习或领域泛化（当测试分布未知时）。
构建数据集的基本思想是用主要概念（例如狗）和视觉概念出现的上下文（例如在草地上）标记图像。通过调整训练和测试数据中不同上下文的比例，可以控制分布程度灵活转变，对不同类型的非 I.I.D. 进行研究。设置。

The NICO dataset is specifically designed for non-I.I.D. or out-of-distribution (OOD) image classification. It simulates a real-world scenario where the test distribution may exhibit arbitrary shifts relative to the training distribution, violating the I.I.D. assumption that underpins most traditional machine learning methods. Typical research directions well supported by this dataset include, but are not limited to, transfer learning or domain adaptation when the test distribution is known, as well as stable learning or domain generalization when the test distribution is unknown. The core idea behind constructing this dataset is to annotate images with both the primary concept (e.g., a dog) and the contextual environment in which the visual concept appears (e.g., on a grassy field). By adjusting the proportions of different contexts in the training and test datasets, one can flexibly control the magnitude of distribution shifts to investigate various types of non-I.I.D. scenarios.

搜集汇总

背景与挑战

背景概述

NICO是一个专为非独立同分布（非I.I.D.）或分布外（OOD）图像分类设计的数据集，模拟真实世界测试分布偏移的场景，支持迁移学习、领域适应和领域泛化等研究。数据集包含10个动物类和9个交通工具类，每个类细分为10个上下文子类（如'在草地上'、'在雪中'），通过调整上下文比例灵活控制分布偏移，提供细粒度标注以促进模型鲁棒性研究。该数据集由清华大学于2020年发布，规模为16.6GB，包含19.2k张图像。