ImageNet-100

ImageNet-100是广泛使用的大规模视觉识别数据库ImageNet (ILSVRC-2012) 的一个特定子集，常被用作评估连续学习和少样本连续学习（FSCIL）算法在处理高分辨率图像时的核心基准。ImageNet-100源自包含1000个类别的原始ImageNet数据集 。由于完整版ImageNet在进行连续学习实验时计算开销巨大，研究人员通常人工划分出一个包含100个类别的子集来进行算法验证 。与其他数据集相比，ImageNet-100的图像分辨率更高、包含更多细节，能够更真实地模拟现实世界中的视觉任务，有效评估模型在面对复杂、高维数据流时克服“灾难性遗忘”的能力 。该数据集基于ImageNet (ILSVRC-2012) 筛选产生 。主要内容为彩色RGB物体图像。与CIFAR-100的低分辨率（32×32）不同，ImageNet-100中的图像通常被调整为224×224像素的标准分辨率输入神经网络 。 在连续学习实验设置中，这100个类别通常被按照随机顺序或语义相关性划分为多个连续的任务步骤（例如：一个包含60类的基础任务，随后是多个包含5类或10类的增量任务），以测试模型的增量适应能力 。ImageNet-100包含100个类别。数据量通常包含约130,000张训练图像（每类约1300张）和5,000张验证图像（每类50张）。 在少样本连续学习（FSCIL）的标准协议中，通常将前60类作为基础任务提供充足数据，而剩余的40类作为新任务，以N-way K-shot（如5-way 5-shot）的形式极少量地提供，要求模型在不遗忘旧知识的前提下快速学习新概念。

ImageNet-100 is a specific subset of the widely used large-scale visual recognition database ImageNet (ILSVRC-2012), and it is commonly employed as a core benchmark for evaluating continual learning and few-shot continual learning (FSCIL) algorithms when processing high-resolution images. ImageNet-100 is derived from the original ImageNet dataset that includes 1000 categories. Since the full ImageNet dataset incurs exorbitant computational overhead in continual learning experiments, researchers typically manually split a 100-category subset for algorithm validation. Compared with other datasets, ImageNet-100 boasts higher image resolution and more detailed content, which can more realistically simulate real-world visual tasks and effectively assess the ability of models to overcome "catastrophic forgetting" when confronted with complex, high-dimensional data streams. This dataset is screened based on ImageNet (ILSVRC-2012), and its main content consists of color RGB object images. Unlike the low-resolution (32×32) CIFAR-100, images in ImageNet-100 are usually resized to the standard 224×224 pixel resolution for input into neural networks. In the experimental setup of continual learning, these 100 categories are often divided into multiple sequential task steps according to random order or semantic relevance—for example, a base task containing 60 categories, followed by multiple incremental tasks each with 5 or 10 categories—to test the incremental adaptation capability of models. ImageNet-100 encompasses 100 categories. The dataset generally contains approximately 130,000 training images (around 1300 images per category) and 5,000 validation images (50 images per category). In the standard protocol of few-shot continual learning (FSCIL), the first 60 categories are usually used as the base task with sufficient data, while the remaining 40 categories serve as new tasks, which are provided in extremely limited quantities in the form of N-way K-shot (e.g., 5-way 5-shot), requiring models to quickly learn new concepts without forgetting old knowledge.

搜集汇总

背景与挑战

背景概述

ImageNet-100是ImageNet的一个子集，包含100个类别和约130,000张训练图像，主要用于评估连续学习和少样本连续学习算法。其高分辨率图像（224×224像素）能有效模拟现实世界视觉任务，测试模型在增量学习中的适应能力。