PixelWeb
收藏资源简介:
PixelWeb是一个包含超过10万个带注释网页的大规模图形用户界面(GUI)数据集。该数据集通过新颖的自动标注方法构建,该方法结合了视觉特征提取和文档对象模型(DOM)结构分析,通过两个核心模块:通道提取和层分析。通道提取确保在遮挡和重叠元素的情况下准确定位GUI元素,通过提取BGRA四通道位图注释实现。层分析使用DOM来确定元素的可视性和堆叠顺序,提供精确的BBox注释。此外,PixelWeb还包括全面的元数据,如元素图像、轮廓和掩码注释。通过三个独立标注者的手动验证,证实了PixelWeb注释的高质量和准确性。在GUI元素检测任务上的实验结果表明,PixelWeb在mAP95指标上的性能比现有数据集高出3-7倍。
PixelWeb is a large-scale graphical user interface (GUI) dataset encompassing over 100,000 annotated web pages. It is built using a novel automatic annotation framework that integrates visual feature extraction and document object model (DOM) structure analysis, with two core modules: channel extraction and layer analysis. Channel extraction ensures accurate localization of GUI elements even in scenarios with occluded and overlapping elements, which is accomplished by extracting BGRA four-channel bitmap annotations. Layer analysis leverages the DOM to identify element visibility and stacking order, delivering precise bounding box (BBox) annotations. Furthermore, PixelWeb provides comprehensive metadata including element images, contour annotations and mask annotations. The high quality and annotation accuracy of PixelWeb are validated via manual reviews conducted by three independent annotators. Experimental results for GUI element detection tasks show that PixelWeb outperforms existing datasets by 3 to 7 times when evaluated using the mAP95 metric.
PixelWeb数据集概述
基本信息
- 许可证: Apache-2.0
- 任务类别: 目标检测
- 语言: 英语
- 规模: 10K<n<100K
数据集版本
- PixelWeb-1K: 包含1,000个GUI截图,带有掩码、轮廓和边界框标注
- PixelWeb-10K: 包含10,000个GUI截图,带有掩码、轮廓和边界框标注
- PixelWeb-100K: 即将发布
文件结构
{id}-screenshot.png: 网页截图{id}-bbox.json: 网页的边界框标注,格式为[[left,top,width,height],...]{id}-contour.json: 网页的轮廓标注,格式为[[[x1,y1,x2,y2,...],...],...]{id}-mask.json: 网页的掩码标注,格式为[[element_id,...],...]{id}-class.json: 网页的类别标注,格式为[axtree_label,...]
数据提取
使用以下命令解压tar.gz文件:
tar -xzvf pixelweb_1k.tar.gz




