野外运动视频（SVW）：用于运动分析的视频数据集

正在创建的数字视频数量呈指数级增长，例如，YouTube已达到每分钟100小时视频的上传速度。这种增长很大程度上是由于智能手机的普及和无处不在的互联网接入。这意味着业余用户生成的视频形成了内容生成的新趋势。因此，迫切需要鲁棒算法来自动分析和检索这些视频。另一方面，许多计算机视觉问题都是由数据驱动的，而具有代表性和真实性的数据集的存在对于开发鲁棒算法是必要的。因此，我们提出了一个高度无约束的体育视频数据集，称为野外体育视频（SVW）。 SVW由4200个视频组成，这些视频由Coach's Eye智能手机应用程序的用户单独使用智能手机拍摄，该应用程序是TechSmith公司开发的一款领先的体育训练应用程序。SVW包括30类运动和44种不同的动作。由于业余玩家的不完善练习和业余用户的不专业捕捉，SVW对于自动化分析非常具有挑战性。 SVW的潜在应用包括：类型分类、动作识别、动作检测和时空对齐。 1、每个视频都附有运动类型的注释。此外，对于40%的视频，还指定了每个动作的时间跨度以及在动作的开始和结束帧显示动作的空间范围的边界框。 2、在SVW中，与现有数据集不同，存在来自同一运动类型的多个动作，使得基于外观的识别不可行。 排球流派类别视频中的注释动作类别（[343，359，前臂]，[380，400，布景]，[438，454，扣球]）。 Comparison with existing datasets Statistics evaluation protocol For questions regarding this dataset please contact Morteza Safdarnejad (safdarne [at] egr.msu.edu). Citation If you use SVW dataset, please refer to this paper in your publications:

The number of digital videos being created is growing exponentially; for example, YouTube now sees 100 hours of video uploaded every minute. This growth is largely driven by the widespread adoption of smartphones and ubiquitous internet access, which has fostered a new trend in content generation with amateur user-generated videos forming the mainstream. Thus, there is an urgent demand for robust algorithms to automatically analyze and retrieve these videos. On the other hand, many computer vision tasks are data-driven, and the availability of representative and authentic datasets is critical for developing robust algorithms. Therefore, we propose a highly unconstrained sports video dataset named Sports in the Wild (SVW). SVW consists of 4200 videos captured solely by users using smartphones with the Coach's Eye mobile application—a leading sports training app developed by TechSmith Corporation. SVW covers 30 sports categories and 44 distinct action classes. Due to the imperfect practice of amateur athletes and unprofessional filming by casual users, SVW poses substantial challenges for automated video analysis. Potential applications of SVW include sports category classification, action recognition, action detection, and spatio-temporal alignment.
1. Each video is annotated with its corresponding sports category. Additionally, for 40% of the videos, the temporal interval of each action and the bounding boxes delineating the spatial extent of the action in its start and end frames are specified.
2. Unlike existing datasets, SVW contains multiple action classes within the same sports category, rendering appearance-based recognition infeasible.
Annotated action categories in volleyball genre videos: ([343, 359, Forearm Pass], [380, 400, Set], [438, 454, Spike]).
Comparison with existing datasets
Statistics
Evaluation protocol
For questions regarding this dataset, please contact Morteza Safdarnejad (safdarne [at] egr.msu.edu).
Citation
If you use the SVW dataset, please cite this paper in your publications:

搜集汇总

背景与挑战

背景概述

野外运动视频（SVW）数据集包含4200个由智能手机拍摄的运动视频，涵盖30类运动和44种不同动作，适用于运动类型分类、动作识别等分析任务。该数据集的特点是高度无约束，视频由业余玩家拍摄，具有挑战性，适合开发鲁棒的自动化分析算法。