EgoHumans
收藏资源简介:
EgoHumans是由Meta Reality Labs创建的全新多视角多人视频基准数据集,专注于推动第一人称视角下的3D人体姿态估计和跟踪技术。该数据集包含超过125,000张第一人称视角的RGB图像,涵盖多种自然环境下的挑战性多人活动,如网球、击剑、排球等。数据集通过使用消费级可穿戴相机眼镜捕捉动态活动,并结合多视角设置生成精确的3D地面实况,即使在严重或完全遮挡的情况下也能保持准确性。EgoHumans不仅提供了高质量的3D地面实况,还通过严格的评估揭示了现有技术的局限性,并提出了新的方法EgoFormer,该方法通过多流变换器架构和明确的3D空间推理,显著提高了跟踪性能。
EgoHumans is a novel multi-view multi-person video benchmark dataset developed by Meta Reality Labs, dedicated to advancing 3D human pose estimation and tracking technologies in first-person perspectives. This dataset includes over 125,000 first-person RGB images, covering challenging multi-person activities in various natural environments such as tennis, fencing, volleyball, and others. It captures dynamic activities using consumer-grade wearable camera glasses, and generates accurate 3D ground truth through multi-view configurations, maintaining accuracy even under severe or total occlusions. EgoHumans not only provides high-quality 3D ground truth, but also reveals the limitations of existing state-of-the-art techniques via rigorous evaluations, and proposes a novel method named EgoFormer. This method utilizes multi-stream Transformer architectures and explicit 3D spatial reasoning to substantially improve tracking performance.




