ImageNet-Sketch

ImageNet-Sketch 数据集由 50000 张图像组成，1000 个 ImageNet 类中的每一个都有 50 张图像。数据集是用谷歌图片查询“sketch of”构建的，其中是标准类名。仅在“黑白”配色方案内进行搜索。每个类最初查询 100 张图像，通过删除不相关的图像和相似但不同类的图像来清理提取的图像。对于某些类，手动清理后的图像少于 50 张，然后通过翻转和旋转图像来扩充数据集。

The ImageNet-Sketch dataset consists of 50,000 images, with 50 images per each of the 1000 ImageNet classes. The dataset is constructed by querying Google Images with the phrase "sketch of" followed by the standard class name, with searches restricted exclusively to the black-and-white color scheme. Initially, 100 images were retrieved for each class via the queries, and the extracted images were cleaned by removing irrelevant ones and those that were similar but belonged to different classes. For some classes, fewer than 50 images remained after manual cleaning, so the dataset was augmented via image flipping and rotation.

搜集汇总

背景与挑战

背景概述

ImageNet-Sketch是一个用于图像分类的素描风格数据集，包含50,000张黑白图像，覆盖1,000个ImageNet类别，每个类别有50张图像。该数据集通过谷歌图片查询'sketch of'构建，经过清理和扩充处理，旨在评估模型在素描图像上的泛化能力，由卡内基梅隆大学于2019年发布。