VideoPipe
收藏资源简介:
视频理解是计算机视觉中的一个重要问题。目前,该问题中经过广泛研究的任务是人类行为识别,其中从长视频中手动剪辑出片段,并且每个片段都假定为一个人类行为类别。然而,在工业应用中,我们可能面临更加复杂的情境。例如,在实际城市管道系统中,异常缺陷是细粒度、多标记、相关领域的。要正确识别它们,我们需要理解细节视频内容。 为此,我们提出了将视频理解的研究重点从传统的动作识别转向工业异常分析。特别是,我们引入了两个高质量的视频基准,即 QV-Pipe 和 CCTV-Pipe,用于实际城市管道系统中的异常检测。基于这些新数据集,我们将举办两个比赛,为智慧城市等领域的视频理解带来新机遇和挑战。
Video understanding is a critical problem in computer vision. Currently, the most widely studied task under this domain is human action recognition, where clips are manually cropped from long videos, and each clip is assumed to belong to a single human action category. However, in industrial applications, we may face more complex scenarios. For instance, in real-world urban pipeline systems, abnormal defects are fine-grained, multi-label and domain-specific. To correctly identify them, we need to comprehend the detailed content of the videos. To this end, we propose to shift the research focus of video understanding from traditional action recognition to industrial anomaly analysis. Specifically, we introduce two high-quality video benchmarks, namely QV-Pipe and CCTV-Pipe, for anomaly detection in real-world urban pipeline systems. Based on these new datasets, we will host two competitions, bringing new opportunities and challenges for video understanding in fields such as smart cities.




