creative-graphic-design/PittImageVideoAdsDataset
收藏资源简介:
PittImageVideoAdsDataset 是一个图像和视频广告数据集,随论文《Automatic Understanding of Image and Video Advertisements》发布。该论文报告了64,832张图像广告和3,477个YouTube广告视频,并包含人工标注的主题、情感、口号、说服策略、符号参考以及行动/原因问答。这个Hugging Face版本公开了可以从官方发布文件中重建的标注行。该加载器提供两种配置:1) image_ads:标注的图像广告,其中image列是从官方Google Storage图像ZIP文件加载的datasets.Image特征,image_url保存了相应的匹兹堡源URL作为文本;2) video_ads:包含YouTube视频ID和URL,以及原始标注者响应和清理后的多数投票标签,但不下载视频文件。数据集支持广告理解任务,如主题识别、情感识别、说服策略分析、符号检测和行动/原因响应建模。标注文本和问答响应均为英文。数据集中包含一个train分割,因为上游发布是一个标注语料库,未定义训练、验证或测试分区。
PittImageVideoAdsDataset is an image and video advertising dataset released alongside the paper *Automatic Understanding of Image and Video Advertisements*. The paper presents 64,832 image advertisements and 3,477 YouTube advertising videos, with manually annotated topics, sentiments, slogans, persuasion strategies, symbolic references, and action-reason question-answering pairs. This Hugging Face version makes accessible the annotation lines that can be reconstructed from the officially released files. The included dataset loader provides two configuration options: 1) `image_ads`: annotated image advertisements, where the `image` column is a `datasets.Image` feature loaded from the official Google Storage image ZIP archive, and `image_url` stores the corresponding Pittsburgh source URL as plain text; 2) `video_ads`: containing YouTube video IDs and URLs, alongside raw annotator responses and cleaned majority-vote labels, without downloading the actual video files. The dataset supports a range of advertising understanding tasks including topic recognition, sentiment recognition, persuasion strategy analysis, symbolic detection, and action-reason response modeling. All annotation texts and question-answering responses are in English. The dataset includes a single `train` split, as the upstream released resource is an annotated corpus with no predefined training, validation, or test partitions.




