HiEve(Human-in-Events)

现代智慧城市的发展高度依赖于以人为本的分析技术的进步。智能多媒体理解是视觉分析的基本技术之一，它需要许多以人为中心和事件驱动的视觉理解任务，例如人体姿态估计、行人跟踪和动作识别。
在这个巨大的挑战中，我们专注于在各种人群和复杂事件中非常具有挑战性和现实性的以人为中心的分析任务，包括地铁上下车、碰撞、战斗和地震逃生（参见图 1）。据我们所知，现有的人类分析方法很少能报告它们在如此复杂的事件下的表现。考虑到这一点，我们进一步提出了一个数据集（称为 Human-in-Events 或 HiEve），该数据集具有大规模和密集注释的标签，涵盖了以人为中心的分析中的广泛任务。

我们的 HiEve 数据集包括当前最大数量的姿势 (>1M)、最大数量的复杂事件动作标签 (>56k)，以及最大数量的长期轨迹之一（平均轨迹长度 >480）。
在我们的数据集上建立了四项具有挑战性的任务，旨在将多媒体和计算机视觉社区的研究人员聚集在一起，从三个方面提高人体运动、姿势和动作分析方法的性能：
• 通过以人为中心的综合分析任务来组织我们大规模数据集上的挑战，并促进以人为中心的理解中的多媒体和人工智能研究和应用。
• 鼓励和加速开发以人为本的分析和理解复杂事件领域的新技术。
• 培育“复杂事件中以人为中心的大规模视觉分析”的新思路和新方向。

The development of modern smart cities is highly dependent on the advancement of human-centric analytical technologies. Intelligent multimedia understanding is one of the fundamental technologies for visual analysis, which encompasses numerous human-centric and event-driven visual understanding tasks, such as human pose estimation, pedestrian tracking, and action recognition.
Amidst this formidable challenge, we focus on highly challenging and realistic human-centric analytical tasks across diverse crowds and complex events, including subway boarding and alighting, collisions, fights, and earthquake evacuations (see Figure 1). To the best of our knowledge, existing human analysis methods rarely report their performance under such complex event scenarios. In view of this, we further propose a dataset (named Human-in-Events or HiEve) that features large-scale and densely annotated labels, covering a wide range of tasks in human-centric analysis.
Our HiEve dataset boasts the largest number of annotated human poses (>1M), the largest number of complex event action labels (>56k), and one of the largest collections of long-term trajectories, with an average trajectory length exceeding 480.
Four challenging tasks are established on our dataset, aiming to bring together researchers from the multimedia and computer vision communities to advance the performance of human motion, pose, and action analysis methods from three perspectives:
• Organize challenges on our large-scale dataset via human-centric comprehensive analytical tasks, and promote research and applications of multimedia and artificial intelligence in human-centric understanding.
• Encourage and accelerate the development of novel technologies in the field of human-centric analysis and complex event understanding.
• Foster new ideas and research directions for "large-scale visual analysis of human-centric scenarios in complex events".

搜集汇总

背景与挑战

背景概述

HiEve数据集是一个大规模、密集标注的以人为中心的视频分析数据集，包含超过100万个人体姿态、5.6万个复杂事件动作标签以及平均长度超过480的长期轨迹。该数据集旨在促进复杂事件中人体运动、姿态和动作分析技术的研究与发展。