MAMe (Museum Art Medium dataset)

我们介绍了 MAMe 数据集，这是一个具有显着高分辨率和可变形状属性的图像分类数据集。 MAMe 的目标是提供一种工具来研究这些属性在图像分类中的影响，同时促进该领域的研究。 MAMe 数据集包含来自三个不同博物馆的数千件艺术品，并提出了一项分类任务，该任务涉及区分由艺术专家监督的 29 种媒介（即材料和技术）。在回顾了当前图像分类任务上下文中 MAMe 的奇异性之后，提供了任务和数据集统计的全面描述。进行了实验以评估同时使用高分辨率图像、可变形状输入以及这两种属性的影响。结果说明了使用高分辨率图像时对性能的积极影响，同时强调了缺乏利用可变形状的解决方案。另一个实验揭示了 MAMe 数据集和原型 ImageNet 数据集之间的独特性。最后，使用可解释的方法和专业知识检查基线，以深入了解剩余的挑战。在回顾了当前图像分类任务上下文中 MAMe 的奇异性之后，提供了任务和数据集统计的全面描述。

We present the MAMe Dataset, an image classification dataset characterized by significantly high-resolution and variable-shaped attributes. The MAMe Dataset aims to provide a tool for investigating the influence of these attributes on image classification, while advancing research in this domain. The MAMe Dataset comprises thousands of artworks sourced from three distinct museums, and defines a classification task that requires distinguishing 29 supervised mediums (i.e., materials and techniques) annotated by art experts. After reviewing the unique characteristics of MAMe within the context of contemporary image classification tasks, a comprehensive overview of the task and dataset statistics is presented. Experiments were conducted to evaluate the effects of using high-resolution images, variable-shaped inputs, and both attributes in combination. The results demonstrate a positive impact on model performance when utilizing high-resolution images, while highlighting the scarcity of solutions that leverage variable-shaped inputs. Another experiment reveals the uniqueness of the MAMe Dataset compared to the prototypical ImageNet Dataset. Finally, baseline models are examined using explainable AI methods and domain expertise to gain in-depth insights into the remaining challenges. After reviewing the unique characteristics of MAMe within the context of contemporary image classification tasks, a comprehensive overview of the task and dataset statistics is presented.

搜集汇总

背景与挑战

背景概述

MAMe数据集是一个高分辨率、可变形状的艺术品图像分类数据集，包含来自三个博物馆的数千件作品，任务涉及区分29种艺术媒介。该数据集旨在研究高分辨率和可变形状属性对图像分类的影响，实验表明高分辨率提升性能，但可变形状处理仍是当前挑战，同时突显了其与ImageNet数据集的独特性。