ImageNet-22K

ImageNet-21K数据集包含更多的图片和类，用于预训练的频率较低，这主要是由于其复杂性以及与标准ImageNet-1K预训练相比低估了其附加值。本文旨在缩小这一差距，并为每个人提供高质量的有效ImageNet-21K预培训。通过专用的预处理阶段，利用WordNet层次结构和称为语义softmax的新颖训练方案，我们展示了不同的模型，包括面向移动的小型模型，显著受益于对众多数据集和任务的ImageNet-21K预训练。我们还表明，对于ViT等著名的新模型，我们的表现优于以前的ImageNet-21K预训练方案。注意，ImageNet-21K数据集和ImageNet-22K是同一个数据集，由于理解上的差异，名称发生了变化

ImageNet-21K contains significantly more images and classes, yet has been less frequently adopted for pre-training, primarily due to its complexity and the undervalued added value compared to standard ImageNet-1K pre-training. This work aims to bridge this gap and deliver high-quality, effective ImageNet-21K pre-training for all users. Through a dedicated preprocessing stage leveraging the WordNet hierarchy and a novel training scheme termed semantic softmax, we demonstrate that various models—including mobile-oriented small-scale models—can substantially benefit from ImageNet-21K pre-training across a wide range of datasets and tasks. We also show that our approach outperforms prior ImageNet-21K pre-training schemes for prominent modern models such as Vision Transformer (ViT). Note that the ImageNet-21K dataset and ImageNet-22K refer to the same dataset, with the name changed due to differences in understanding.

搜集汇总

背景与挑战

背景概述

ImageNet-22K（即ImageNet-21K）是一个包含大量图片和类别的图像分类数据集，主要用于模型预训练。虽然因复杂性使用频率较低，但通过专用预处理和语义softmax方案可显著提升各类模型的性能表现。