SYSU-30k

SYSU-30k是由中山大学数据科学与计算机科学学院创建的大规模人物重识别基准数据集。该数据集包含30,000个个体，总计29,606,918张图像，是现有数据集如CUHK03和Market-1501的20倍以上。SYSU-30k通过使用包级别的标注大大减少了标注工作量，适用于弱监督人物重识别问题的研究。数据集创建过程中，从互联网下载了大量短节目视频，并使用YOLO-v2进行行人边界框检测，最终由三名标注者进行了20天的标注工作。SYSU-30k不仅为弱监督人物重识别问题提供了一个大型平台，还提供了一个更贴近实际设置的更具挑战性的测试集，旨在解决大规模视频监控中的人物识别问题。SYSU-30k is a large-scale person re-identification benchmark dataset developed by the School of Data and Computer Science, Sun Yat-sen University. This dataset encompasses 30,000 distinct individuals and a total of 29,606,918 images, which is more than 20 times the size of existing benchmark datasets such as CUHK03 and Market-1501. By adopting package-level annotation, SYSU-30k greatly reduces annotation workload, making it highly applicable to research on weakly-supervised person re-identification. During the dataset construction process, a large volume of short video clips were downloaded from the Internet, and pedestrian bounding box detection was conducted using YOLO-v2. Ultimately, three annotators completed the annotation task over a period of 20 days. SYSU-30k not only provides a large-scale platform for weakly-supervised person re-identification research, but also offers a more challenging test set that better aligns with real-world deployment settings, with the goal of addressing person recognition challenges in large-scale video surveillance scenarios.

搜集汇总

背景与挑战

背景概述

SYSU-30k是一个大规模弱监督行人重识别数据集，包含30,508个不同人物类别和约2,960万张图像，规模远超现有主流Re-ID数据集。该数据集提供bag-level弱监督标注，测试集设计极具挑战性，包含大量干扰项，旨在评估模型在复杂场景下的可扩展性和鲁棒性。