GenBench

GenBench是一个包含22个数据集、2548个类别的广泛基准，用于评估生成数据在各种视觉识别任务中的效果。该数据集分为三个主要组：常见概念、细粒度概念和罕见概念。常见概念主要涵盖日常生活中的通用对象，如ImageNet-1K和CIFAR-10；细粒度概念涉及元类别内的子类别，如Oxford Pets和Stanford Cars；罕见概念包括现实世界中较少见的对象，如PatchCamelyon和EuroSAT。GenBench的创建旨在通过全面的基准和分析，突出生成数据在视觉识别中的潜力，并识别未来研究的关键挑战。GenBench is a comprehensive benchmark encompassing 22 datasets and 2,548 classes, designed to evaluate the performance of generated data across a wide range of visual recognition tasks. This benchmark is divided into three main groups: common concepts, fine-grained concepts, and rare concepts. Common concepts primarily cover general objects in daily life, such as ImageNet-1K and CIFAR-10; fine-grained concepts involve subcategories within their respective meta-categories, exemplified by Oxford Pets and Stanford Cars; rare concepts include relatively scarce real-world objects like PatchCamelyon and EuroSAT. GenBench was developed to highlight the potential of generated data in visual recognition and pinpoint key challenges for future research through comprehensive benchmarking and analysis.

搜集汇总

背景与挑战

背景概述

GenBench是一个包含22个数据集和2548个类别的广泛基准，用于评估生成数据在视觉识别任务中的效果。它分为三个主要组：常见概念涵盖日常通用对象，细粒度概念涉及元类别的子类别，罕见概念包括现实世界中较少的对象。该数据集旨在通过全面的基准分析，突出生成数据的潜力并识别研究挑战。