nlphuji/vasr
收藏资源简介:
--- annotations_creators: - crowdsourced language: - en language_creators: - found license: - cc-by-4.0 multilinguality: - monolingual paperswithcode_id: vasr pretty_name: VASR size_categories: - 1K<n<10K source_datasets: - original tags: - commonsense-reasoning - visual-reasoning task_ids: [] extra_gated_prompt: "By clicking on “Access repository” below, you also agree that you are using it solely for research purposes. The full license agreement is available in the dataset files." --- # Dataset Card for VASR - [Dataset Description](#dataset-description) - [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards) - [How to Submit Predictions?](#how-to-submit-predictions?) - [Colab notebook code for VASR evaluation with ViT](#colab-notebook-code-for-vasr-evaluation-with-clip) - [Languages](#languages) - [Dataset Structure](#dataset-structure) - [Data Fields](#data-fields) - [Data Splits](#data-splits) - [Dataset Creation](#dataset-creation) - [Considerations for Using the Data](#considerations-for-using-the-data) - [Licensing Information](#licensing-information) - [Citation Information](#citation-information) ## Dataset Description VASR is a challenging dataset for evaluating computer vision commonsense reasoning abilities. Given a triplet of images, the task is to select an image candidate B' that completes the analogy (A to A' is like B to what?). Unlike previous work on visual analogy that focused on simple image transformations, we tackle complex analogies requiring understanding of scenes. Our experiments demonstrate that state-of-the-art models struggle with carefully chosen distractors (±53%, compared to 90% human accuracy). - **Homepage:** https://vasr-dataset.github.io/ - **Colab** https://colab.research.google.com/drive/1HUg0aHonFDK3hVFrIRYdSEfpUJeY-4dI - **Repository:** https://github.com/vasr-dataset/vasr/tree/main/experiments - **Paper:** https://arxiv.org/abs/2212.04542 - **Leaderboard:** https://vasr-dataset.github.io/ - **Point of Contact:** yonatan.bitton@mail.huji.ac.il ## Supported Tasks and Leaderboards https://vasr.github.io/leaderboard. https://paperswithcode.com/dataset/vasr. ## How to Submit Predictions? To submit predictions, please send a prediction CSV file to vasr.benchmark@gmail.com / yonatan.bitton@mail.huji.ac.il. The prediction file should include a "B'" column with the predicted candidate name that best solves the analogy, and an index from 1 to 4 indicating the location of the predicted candidate in the given candidate list. An example prediction file is available [HERE](https://drive.google.com/file/d/1NvBNdvlWmEOYjIVi2xdmQ_tUm-TXo42u/view?usp=share_link). A submission is allowed once a week, and you will receive a response within a week. ## Colab notebook code for VASR evaluation with ViT https://colab.research.google.com/drive/1HUg0aHonFDK3hVFrIRYdSEfpUJeY-4dI ### Languages English. ## Dataset Structure ### Data Fields A: datasets.Image() - the first input image, **A**:A'. A': datasets.Image() - the second input image, different from A in a single key, A:**A'**. B: datasets.Image() - the third input image, has the same different item as A, **B**:B'. B': datasets.Image() - the forth image, which is the analogy solution. Different from B in a single key (the same different one as in A:A'), B:**B'**. Hidden in the test set. candidates_images: [datasets.Image()] - a list of candidate images solutions to the analogy. label: datasets.Value("int64") - the index of the ground-truth solution. Hidden in the test set. candidates: [datasets.Value("string")] - a list of candidate string solutions to the analogy. ### Data Splits There are three splits, TRAIN, VALIDATION, and TEST. Since there are four candidates and one solution, random chance is 25%. ## Dataset Creation We leverage situation recognition annotations and the CLIP model to generate a large set of 500k candidate analogies. There are two types of labels: - Silver labels, obtained from the automatic generation. - Gold labels, obtained from human annotations over the silver annotations. In the huggingface version we provide only the gold labeled dataset. Please refer to the project website download page if you want to download the silver labels version. ### Annotations #### Annotation process We paid Amazon Mechanical Turk Workers to solve analogies, five annotators for each analogy. Workers were asked to select the image that best solves the analogy. The resulting dataset is composed of the 3,820 instances agreed upon with a majority vote of at least 3 annotators, which was obtained in 93% of the cases. ## Considerations for Using the Data All associations were obtained with human annotators. All used images are from the imSitu dataset (http://imsitu.org/) Using this data is allowed for academic research alone. ### Licensing Information CC-By 4.0 ### Citation Information NA
annotations_creators: - 众包(crowdsourced) language: - 英语(en) language_creators: - 现有采集(found) license: - 知识共享署名4.0国际许可协议(cc-by-4.0) multilinguality: - 单语言(monolingual) paperswithcode_id: vasr pretty_name: VASR size_categories: - 1K<n<10K source_datasets: - 原始数据集(original) tags: - 常识推理(commonsense-reasoning) - 视觉推理(visual-reasoning) task_ids: [] extra_gated_prompt: "点击下方“访问仓库”即表示您同意仅将其用于研究目的。完整许可协议可在数据集文件中查看。" --- # VASR 数据集卡片 - [数据集描述](#dataset-description) - [支持任务与排行榜](#supported-tasks-and-leaderboards) - [如何提交预测结果?](#how-to-submit-predictions?) - [基于视觉Transformer(ViT)的VASR评估Colab笔记本代码](#colab-notebook-code-for-vasr-evaluation-with-vit) - [语言](#languages) - [数据集结构](#dataset-structure) - [数据字段](#data-fields) - [数据划分](#data-splits) - [数据集构建](#dataset-creation) - [数据集使用注意事项](#considerations-for-using-the-data) - [许可信息](#licensing-information) - [引用信息](#citation-information) ## 数据集描述 VASR 是一款用于评估计算机视觉常识推理能力的挑战性数据集。给定一组三张图像的三元组,任务需选出能够完成类比关系的候选图像B',即类比关系为“A对应A',如同B对应何者?”。与此前聚焦于简单图像变换的视觉类比研究不同,本数据集针对需要理解场景的复杂类比任务。实验结果表明,当前最先进的模型在精心设计的干扰项面前表现欠佳,准确率仅约53%,而人类准确率可达90%。 - **主页(Homepage):** https://vasr-dataset.github.io/ - **Colab 笔记本** https://colab.research.google.com/drive/1HUg0aHonFDK3hVFrIRYdSEfpUJeY-4dI - **代码仓库(Repository):** https://github.com/vasr-dataset/vasr/tree/main/experiments - **论文(Paper):** https://arxiv.org/abs/2212.04542 - **排行榜(Leaderboard):** https://vasr-dataset.github.io/ - **联系方式(Point of Contact):** yonatan.bitton@mail.huji.ac.il ## 支持任务与排行榜 https://vasr.github.io/leaderboard. https://paperswithcode.com/dataset/vasr. ## 如何提交预测结果? 若需提交预测结果,请将预测CSV文件发送至vasr.benchmark@gmail.com 或 yonatan.bitton@mail.huji.ac.il。 预测文件需包含名为“B'”的列,填入可最佳解决该类比任务的候选图像名称,以及一个1至4的索引,用于标注该预测候选在给定候选列表中的位置。 示例预测文件可通过[此处](https://drive.google.com/file/d/1NvBNdvlWmEOYjIVi2xdmQ_tUm-TXo42u/view?usp=share_link)获取。 提交请求每周仅可发起一次,您将在一周内收到回复。 ## 基于视觉Transformer(ViT)的VASR评估Colab笔记本代码 https://colab.research.google.com/drive/1HUg0aHonFDK3hVFrIRYdSEfpUJeY-4dI ### 语言 英语。 ## 数据集结构 ### 数据字段 - **A**: datasets.Image() — 第一张输入图像,对应类比关系 **A**:A'。 - **A'**: datasets.Image() — 第二张输入图像,仅在单个特征维度上与A存在差异,对应类比关系 A:**A'**。 - **B**: datasets.Image() — 第三张输入图像,与A共享相同的差异项,对应类比关系 **B**:B'。 - **B'**: datasets.Image() — 第四张图像,为该类比任务的解。仅在单个特征维度上与B存在差异(与A:A'中的差异项一致),对应类比关系 B:**B'**。该字段在测试集内被隐藏。 - **candidates_images**: [datasets.Image()] — 该类比任务的候选解图像列表。 - **label**: datasets.Value("int64") — 真实解的索引。该字段在测试集内被隐藏。 - **candidates**: [datasets.Value("string")] — 该类比任务的候选解字符串列表。 ### 数据划分 数据集包含训练(TRAIN)、验证(VALIDATION)与测试(TEST)三个划分。由于每个样本对应4个候选与1个正确解,随机猜测的准确率为25%。 ## 数据集构建 我们借助场景识别标注与CLIP(CLIP)模型生成了总计50万组候选类比样本。 该数据集包含两类标签: - **银标(Silver labels)**:通过自动生成流程获得的标签。 - **金标(Gold labels)**:基于银标样本经人工标注得到的标签。 在Hugging Face版本的数据集仅提供金标样本。若需下载银标版本,请前往项目官网的下载页面获取。 ### 标注流程 我们聘请亚马逊机械 Turk(Amazon Mechanical Turk)标注人员完成类比任务求解,每个类比样本由5名标注人员共同完成。 标注人员需选出可最佳解决该类比任务的图像。 最终数据集共包含3820个样本,所有样本均通过至少3名标注人员的多数投票达成一致,该筛选规则覆盖了93%的原始生成样本。 ## 数据集使用注意事项 所有关联关系均由人工标注人员获取。 所有使用的图像均来自imSitu数据集(http://imsitu.org/)。 本数据集仅可用于学术研究用途。 ### 许可信息 知识共享署名4.0国际许可协议(cc-by-4.0) ### 引用信息 无(NA)
数据集概述
基本信息
- 名称: VASR
- 语言: 英语
- 许可证: CC-BY-4.0
- 多语言性: 单语种
- 规模: 1K<n<10K
- 来源: 原始数据
- 标签: 常识推理, 视觉推理
数据集描述
- 任务: 给定一组三张图片,任务是选择一张图片候选B,完成类比推理(A到A类似于B到什么?)。
- 特点: 与以往关注简单图像变换的视觉类比不同,VASR处理复杂的类比,需要理解场景。
- 性能: 当前最先进模型在精心选择的干扰项上表现不佳(±53%,相比人类准确率90%)。
数据集结构
- 数据字段:
- A, A: 输入图像,表示初始状态和变化后的状态。
- B, B: 输入图像,B为类比解决方案,测试集中隐藏。
- candidates_images: 候选图像列表。
- label: 真实解决方案的索引,测试集中隐藏。
- candidates: 候选字符串列表。
- 数据分割: TRAIN, VALIDATION, TEST三个分割。
数据集创建
- 注释过程: 通过Amazon Mechanical Turk Workers进行,每个类比由五名注释者完成,采用多数投票机制。
- 标签类型: 银标(自动生成)和金标(人工注释)。
使用考虑
- 使用限制: 仅允许用于学术研究。
- 图像来源: imSitu数据集。
许可证信息
- 许可证: CC-BY-4.0
引用信息
- 引用信息: 未提供。




