VidOR

VidOR（视频对象关系）数据集包含来自 YFCC100M 集合的 10,000 个视频（98.6 小时）以及大量用于关系理解的细粒度注释。特别是，80 类对象用边界框轨迹进行注释，以指示它们在视频中的时空位置；并且在具有起始和结束帧索引的所有注释对象对中注释了50类关系谓词。这导致大约 50,000 个对象和 380,000 个关系实例被注释。为了使用该数据集进行模型开发，该数据集分为 7,000 个用于训练的视频、835 个用于验证的视频和 2,165 个用于测试的视频。

VidOR (Video Object Relation) dataset contains 10,000 videos (98.6 hours) sourced from the YFCC100M collection, along with extensive fine-grained annotations for relation understanding. Specifically, 80 object categories are annotated with bounding box trajectories to indicate their spatiotemporal locations within the videos; additionally, 50 categories of relational predicates are annotated for all pairs of annotated objects, with the start and end frame indices of each relational instance specified. This results in approximately 50,000 object instances and 380,000 relational instances being annotated. For model development using this dataset, it is split into three subsets: 7,000 videos for training, 835 videos for validation, and 2,165 videos for testing.

搜集汇总

背景与挑战

背景概述

VidOR是一个视频对象关系数据集，包含10,000个视频，专注于细粒度的视觉关系检测和标记，涵盖80类对象和50类关系谓词，总计约50,000个对象和380,000个关系实例。该数据集主要用于视频中的时空关系理解研究，并已划分为训练、验证和测试集，以支持模型开发。