CoVA (CoVA dataset for Webpage Object Detection / Information Extraction)

我们标记了跨越 408 个域（亚马逊、沃尔玛、Target 等）的 7,740 个网页截图。这些网页中的每一个都只包含一个标有价格、标题和图像的标签。所有其他 Web 元素都标记为背景。平均而言，一个网页中有 90 个 Web 元素。 网页截图和边界框可以在这里获取 训练验证测试拆分 我们创建了一个跨域拆分，以确保每个训练集、验证集和测试集都包含来自不同域的网页。具体来说，我们根据不同域的数量构建了一个 3 : 1 : 1 的分割。我们观察到排名前 5 的域名（基于样本数量）是 Amazon、EBay、Walmart、Etsy 和 Target。因此，我们为 5 折交叉验证创建了 5 个不同的拆分，以便每个主要域都存在于测试数据的 5 个拆分之一中。We annotated 7,740 web page screenshots spanning 408 domains (Amazon, Walmart, Target, etc.). Each web page contains exactly one annotated target with price, title, and image attributes. All other web elements are labeled as background. On average, each web page contains 90 web elements. The web page screenshots and corresponding bounding boxes are accessible here. Train-Validation-Test Split We implemented a cross-domain split strategy to ensure that the training, validation, and test sets each include web pages from distinct domains. Specifically, we formulated a 3:1:1 partition based on the count of unique domains. We found that the top 5 domains by sample volume are Amazon, eBay, Walmart, Etsy, and Target. Consequently, we developed 5 distinct splits for 5-fold cross-validation, ensuring that each major domain appears in exactly one of the five test splits.

搜集汇总

背景与挑战

背景概述

CoVA数据集包含7740个网页截图，覆盖408个域名，专门用于网页目标检测和信息提取任务。每个网页标注了价格、标题和图像等关键元素，其余元素视为背景，并采用跨域拆分方式划分为训练、验证和测试集，以支持模型评估。