EgoTracks
收藏资源简介:
EgoTracks是首个大规模长期第一人称视觉对象跟踪数据集,由Meta AI和UT Austin联合创建。数据集包含22,028个跟踪轨迹,源自5708个平均时长6分钟的视频,这些视频均来自Ego4D数据集,涵盖了日常生活中的多种场景。EgoTracks的创建过程涉及从Ego4D中筛选视频,并对其进行密集标注,每个视频平均包含约1800帧,标注工作耗时约1到2小时每轨迹。该数据集主要用于训练和评估长期跟踪模型,特别是在处理频繁的对象出现与消失、视角变化、手部与对象互动等挑战性场景中的应用。
EgoTracks is the first large-scale long-term first-person visual object tracking dataset, jointly created by Meta AI and The University of Texas at Austin. The dataset contains 22,028 tracking trajectories, sourced from 5,708 videos with an average duration of 6 minutes, all extracted from the Ego4D dataset and covering various daily life scenarios. The construction of EgoTracks involves screening videos from Ego4D and performing dense annotations on them. Each video contains approximately 1,800 frames on average, and the annotation work takes about 1 to 2 hours per trajectory. This dataset is primarily used for training and evaluating long-term tracking models, especially for applications in challenging scenarios such as frequent object appearance and disappearance, viewpoint changes, and hand-object interactions.

- 1EgoTracks: A Long-term Egocentric Visual Object Tracking DatasetMeta AI · 2023年



