Mini-ImageNet

Mini-ImageNet最初由Google DeepMind（Vinyals等人）在2016年为了评估匹配网络（Matching Networks）而提出，现已成为少样本学习和少样本连续学习（FSCIL）领域最权威、最通用的基准数据集之一。Mini-ImageNet是从原始的大规模ImageNet (ILSVRC-2012) 数据集中抽样构建的子集。它的提出是为了解决CIFAR-100图像过于简单（分辨率低），而完整ImageNet计算过于庞大且不便于快速迭代少样本算法的问题。Mini-ImageNet保留了ImageNet中图像的复杂性和多样性，但大幅缩减了规模，因此被广泛用于评估模型在数据受限场景下的泛化能力，特别是在类增量学习设置下，测试模型能否通过极少量样本（如每类仅5张）迅速掌握新类别而不遗忘旧知识。该数据集基于ImageNet (ILSVRC-2012) 数据库人工筛选产生。 主要内容为彩色RGB物体图像。 为了适应少样本学习算法的训练效率，图像通常被调整为84×84像素。 在少样本连续学习（FSCIL）的标准实验协议中，通常采用“60+40”的划分模式：将数据集划分为一个包含60个类别的基础任务（Base Task），以及随后分8-9个阶段到达的40个新类别（Incremental Tasks）。Mini-ImageNet共包含100个类别（如鸟类、犬类、容器、车辆等）。 数据量包含60,000张图像，每个类别包含600张图像（通常按500张训练、100张测试划分）。 在典型的N-way K-shot（如5-way 5-shot）连续学习实验中，模型首先在数据充足的60个基础类上训练，随后的每个增量阶段仅提供5个新类，且每个新类仅提供5张样本，要求模型实现持续的知识累积。

Mini-ImageNet was originally proposed by Google DeepMind (Vinyals et al.) in 2016 for the evaluation of Matching Networks, and has now emerged as one of the most authoritative and widely adopted benchmark datasets in the fields of few-shot learning and few-shot continual learning (FSCIL). It is a curated subset sampled from the original large-scale ImageNet (ILSVRC-2012) dataset. The dataset was developed to address the shortcomings of existing benchmarks: CIFAR-100 images are too simple with low resolution, while the full ImageNet dataset is computationally too intensive and not conducive to rapid iterative optimization of few-shot learning algorithms. Mini-ImageNet retains the complexity and diversity of ImageNet images while drastically downsizing the dataset scale, making it widely used to evaluate the generalization performance of models in data-scarce scenarios, particularly in class-incremental learning settings, where models are required to quickly acquire new categories with extremely limited samples (e.g., only 5 images per class) without forgetting previously learned knowledge. This dataset is manually curated based on the ImageNet (ILSVRC-2012) database. Its core content consists of color RGB object images. To accommodate the training efficiency requirements of few-shot learning algorithms, images are typically resized to 84×84 pixels. In the standard experimental protocol for few-shot continual learning (FSCIL), a "60+40" division scheme is commonly adopted: the dataset is split into a base task containing 60 classes, followed by 40 new classes (incremental tasks) introduced across 8 to 9 subsequent stages. Mini-ImageNet comprises a total of 100 categories, including birds, canines, containers, vehicles, and others. The dataset contains 60,000 images in total, with 600 images per category, which are typically split into 500 training samples and 100 test samples per class. In typical N-way K-shot (e.g., 5-way 5-shot) continual learning experiments, models are first trained on the sufficiently sampled 60 base classes. Then, each subsequent incremental stage only provides 5 new classes, with only 5 samples available per new class, requiring the model to achieve continuous knowledge accumulation.

搜集汇总

背景与挑战

背景概述

Mini-ImageNet是从ImageNet (ILSVRC-2012)抽样构建的子集，包含100个类别、60,000张图像，图像调整为84×84像素，旨在作为少样本学习和少样本连续学习（FSCIL）的权威基准数据集，用于评估模型在数据受限场景下的泛化能力。数据集采用“60+40”划分模式，支持类增量学习实验，每个新类仅提供极少量样本（如5张），以测试模型持续学习能力。