ScribbleSup (PASCAL-Scribble Dataset)

PASCAL-Scribble 数据集是 PASCAL 数据集的扩展，带有用于语义分割的涂鸦注释。注释遵循两种不同的协议。在第一个协议中，PASCAL VOC 2012 集被注释，具有 20 个对象类别（飞机、自行车等）和一个背景类别。标注了 12031 张图像，其中训练集中的 10582 张图像和验证集中的 1449 张图像。在第二个协议中，使用了 PASCAL-CONTEXT 数据集中涉及的 59 个对象/东西类别和一个背景类别。除了第一个协议中的 20 个对象类别外，还包括 39 个额外的类别（雪、树……）。遵循此协议来注释 PASCAL-CONTEXT 数据集。训练集中的 4,998 张图像已被标注。The PASCAL-Scribble dataset is an extension of the PASCAL dataset, featuring scribble annotations for semantic segmentation. The annotations follow two distinct protocols. For the first protocol, the PASCAL VOC 2012 dataset is annotated, which covers 20 object categories (e.g., airplane, bicycle) and one background category. A total of 12,031 images are annotated, including 10,582 images in the training split and 1,449 images in the validation split. For the second protocol, 59 object/stuff categories from the PASCAL-CONTEXT dataset plus one background category are adopted. In addition to the 20 object categories included in the first protocol, 39 additional categories (e.g., snow, tree, etc.) are added. The PASCAL-CONTEXT dataset is annotated following this protocol, with 4,998 images in its training split being annotated.

搜集汇总

背景与挑战

背景概述

ScribbleSup (PASCAL-Scribble Dataset)是基于PASCAL数据集扩展的语义分割数据集，提供两种注释协议（20类和59类对象），包含大量标注图像（总计约16,000张），适用于弱监督语义分割研究。