GTAutoAct

GTAutoAct是由东京大学开发的一个创新数据集生成框架，利用游戏引擎技术推动动作识别的发展。该框架能够自动创建大规模、高质量、多视角的视频数据集，涵盖广泛的动作类别。GTAutoAct通过将现有的基于坐标的3D人体运动转换为旋转导向的表示，提高了多视角的适应性，并通过动态骨骼插值算法生成平滑且真实的动画。此外，该框架还实现了自动视频捕捉和处理流程，包括随机导航摄像头和自动修剪及标注功能，适用于多种定制场景，旨在解决传统动作识别数据集在类别范围、视角多样性、环境多样性、视频质量和人工收集成本等方面的问题。

GTAutoAct is an innovative dataset generation framework developed by The University of Tokyo that advances action recognition research through game engine technologies. This framework can automatically generate large-scale, high-quality multi-view video datasets covering a broad spectrum of action categories. By converting existing coordinate-based 3D human motion into rotation-guided representations, GTAutoAct enhances multi-view adaptability, and generates smooth and realistic animations via dynamic bone interpolation algorithms. Additionally, the framework implements an automated video capture and processing pipeline, including randomly navigated cameras, automatic trimming and annotation functionalities, and supports diverse customized scenarios. It is designed to resolve the long-standing issues of traditional action recognition datasets, including limited category coverage, insufficient viewpoint and environmental diversity, subpar video quality, and high manual collection costs.