CUB-200-2011

CUB-200-2011（Caltech-UCSD Birds-200-2011），是由加州理工学院和加州大学圣地亚哥分校的研究人员（Wah等人）构建发布的经典数据集，是评估细粒度视觉分类、细粒度连续学习及少样本连续学习（FSCIL）算法性能的核心基准。CUB-200专注于细粒度领域，即区分同一大类下的不同子类（如区分“加州海鸥”与“北极海鸥”）。在连续学习背景下，该数据集极具挑战性，因为不同类别间的视觉特征差异极其微小（高类间相似性），且同一类别受姿态、背景影响差异较大（高类内差异性）。这要求模型在持续学习新鸟类的过程中，不仅要克服遗忘，还要保持对微细特征（如喙的形状、羽毛纹理）的敏锐辨别力。产生方法与主要内容： 该数据集基于Flickr图像搜索筛选产生，并进行了人工标注。主要内容为自然场景下的鸟类彩色RGB图像。 除了类别标签外，该数据集还提供了丰富的额外观测值，包括15个解剖部件的位置（关键点）、312个二进制视觉属性（如“翅膀颜色：蓝色”、“喙形状：钩状”）以及物体边界框（Bounding Box）。 在连续学习与少样本连续学习的标准实验协议中，通常不使用这些额外标注，仅使用图像和类别标签。主要内容与体量： CUB-200-2011包含200个鸟类物种。 数据量包含11,788张图像，其中5,994张用于训练，5,794张用于测试（每类约60张）。 在少样本连续学习（FSCIL）的标准设置中，通常采用“100+100”的划分模式：将前100个类别作为基础任务（Base Task）提供充足数据训练，剩余的100个类别被划分为10个增量阶段（Sessions），每个阶段包含10个新类，通常以N-way K-shot（如10-way 5-shot）的形式提供，测试模型在细粒度领域的增量适应能力。

CUB-200-2011 (Caltech-UCSD Birds-200-2011) is a classic dataset constructed and released by researchers from the California Institute of Technology and the University of California, San Diego (Wah et al.). It serves as a core benchmark for evaluating the performance of fine-grained visual classification, fine-grained continual learning, and few-shot continual learning (FSCIL) algorithms.

CUB-200 focuses on the fine-grained domain, which involves distinguishing different subcategories under the same broad category (e.g., differentiating "California Gull" from "Arctic Tern"). In the context of continual learning, this dataset is highly challenging due to two prominent challenges: extremely subtle visual differences between distinct categories (high inter-class similarity) and significant intra-class variations caused by different postures and backgrounds (high intra-class variability). This demands that models not only mitigate catastrophic forgetting when continuously learning new bird species but also retain sharp discriminatory capabilities for micro-features such as beak shape and feather texture.

Dataset Generation and Core Contents: This dataset was screened via Flickr image searches and underwent manual annotation. Its primary content comprises color RGB images of birds captured in natural scenes. Apart from category labels, the dataset also provides rich supplementary annotations, including the coordinates of 15 anatomical key points, 312 binary visual attributes (e.g., "wing color: blue", "beak shape: hooked"), and object bounding boxes. In standard experimental protocols for continual learning and few-shot continual learning, these additional annotations are typically not utilized, and only images and category labels are adopted.

Main Content and Scale: CUB-200-2011 encompasses 200 bird species, with a total of 11,788 images. Of these, 5,994 are allocated for training and 5,794 for testing, with approximately 60 images per species. In the standard setup for few-shot continual learning (FSCIL), a "100+100" splitting pattern is commonly employed: the first 100 categories are used as the base task with sufficient data for training, while the remaining 100 categories are divided into 10 incremental sessions, each containing 10 new classes. The setup usually follows the N-way K-shot format (e.g., 10-way 5-shot) to evaluate the model's incremental adaptation ability in the fine-grained domain.

搜集汇总

背景与挑战

背景概述

CUB-200-2011是一个经典的细粒度鸟类图像数据集，包含200个物种、11,788张图像，用于评估细粒度视觉分类和连续学习算法。其特点在于高类间相似性和高类内差异性，挑战模型对微细特征的辨别力，并提供关键点、属性等额外标注。在少样本连续学习标准协议中，通常将前100类作为基础任务，后100类分为10个增量阶段进行测试。