LaSOT
收藏资源简介:
大规模单目标跟踪 (LaSOT) 旨在为训练需要大量数据的深度跟踪器以及评估长期跟踪性能提供专用平台。 LaSOT 具有以下特点: 大规模:1,550 个序列,超过 387 万帧 高质量:手动注释,每帧仔细检查 类别平衡:85 个类别,每个类别包含二十(70 个类别)或十个(15 个类别)序列 长期跟踪:平均视频长度约为 2,500 帧(即 83 秒) 全面的标注:为每个序列提供视觉和语言标注 灵活的评估协议:在三种不同协议下进行评估:无约束、完全重叠和一次性
The Large-scale Single-object Tracking (LaSOT) dataset is designed as a dedicated platform for training data-intensive deep trackers and evaluating long-term tracking performance. LaSOT has the following characteristics: - Large-scale: 1,550 sequences with over 3.87 million frames in total - High-quality: Manually annotated and carefully inspected frame by frame - Class-balanced: A total of 85 categories, where 70 categories contain 20 sequences each, and the remaining 15 categories contain 10 sequences each - Long-term tracking support: The average duration of the video sequences is approximately 2,500 frames (equivalent to 83 seconds) - Comprehensive annotations: Both visual and linguistic annotations are provided for each sequence - Flexible evaluation protocols: Evaluations are conducted under three distinct protocols: unconstrained, full overlap, and one-shot

- LaSOT数据集首次发表,由Heng Fan等人提出,旨在为视觉目标跟踪领域提供一个大规模、高质量的基准测试数据集。
- LaSOT数据集在CVPR 2019会议上正式发布,并迅速成为视觉目标跟踪研究中的重要基准之一。
- 随着LaSOT数据集的广泛应用,多个基于该数据集的跟踪算法被提出,显著推动了视觉目标跟踪技术的发展。
- LaSOT数据集的影响力进一步扩大,被多个国际顶级会议和期刊引用,成为评估跟踪算法性能的标准数据集之一。



