TAO
收藏资源简介:
TAO 是用于跟踪任何对象的联合数据集,包含 2,907 个高分辨率视频,在不同环境中捕获,平均时长为半分钟。我们采用自下而上的方法来发现 833 个类别的大量词汇,比之前的跟踪基准高出一个数量级。为此,我们要求注释者为在视频中任意点移动的对象标记轨迹,并在事后为它们命名。我们的词汇量比现有的跟踪数据集要大得多,而且在质量上也有所不同。为了确保注释的可扩展性,我们采用了一种联合方法,该方法将手动工作集中在为视频中的相关对象(例如,那些移动的对象)标记轨道上。我们对最先进的跟踪方法进行了广泛的评估,并就开放世界中的大词汇表跟踪做出了许多重要发现。
TAO is a joint dataset for tracking arbitrary objects, which contains 2,907 high-resolution videos captured across diverse environments, with an average duration of 30 seconds. We adopt a bottom-up approach to discover a large vocabulary of 833 categories, which is an order of magnitude larger than prior tracking benchmarks. To this end, we ask annotators to mark trajectories for objects moving at any point in the videos, and assign names to them afterwards. Our vocabulary is not only substantially larger than existing tracking datasets, but also qualitatively distinct. To ensure annotation scalability, we employ a joint approach that focuses manual labeling efforts on marking trajectories of relevant objects in the videos, such as those in motion. We conduct extensive evaluations on state-of-the-art tracking methods, and yield numerous critical findings regarding large-vocabulary tracking in the open world.

- TAO数据集首次发表,由Facebook AI Research(FAIR)团队发布,旨在推动视频目标检测和跟踪的研究。
- TAO数据集首次应用于CVPR 2020的挑战赛,吸引了全球多个研究团队参与,推动了视频目标检测技术的进步。
- TAO数据集在ICCV 2021上再次成为焦点,多个基于TAO数据集的研究成果被展示,进一步验证了其在视频分析领域的价值。
- 1TAO: A Large-Scale Benchmark for Tracking Any ObjectFacebook AI Research · 2020年
- 2Towards Real-Time Multi-Object TrackingZhejiang University · 2019年
- 3Tracking Any Object with a Generalizable ApproachUniversity of California, Berkeley · 2021年
- 4A Comprehensive Study on Object Tracking with Deep LearningStanford University · 2022年
- 5Exploring the Limits of Object Tracking with TAO DatasetMassachusetts Institute of Technology · 2022年



