敦煌舞三维动作数据集

Dunhuang dance is an artistic treasure of Chinese traditional culture, which originates from Dunhuang mural paintings and is an important component of Dunhuang culture, and its digital preservation, display and research are of great significance. In order to promote the digitization process and development of Dunhuang Dance, this study proposes to combine Dunhuang Dance with three-dimensional human posture estimation technology to construct a three-dimensional movement database of Dunhuang Dance. For Dunhuang dance, the Dunhuang dance is divided, choreographed and captured according to character images, and a video action database of Dunhuang dance is constructed; for posture estimation, AMASS (Archive of Motion Capture as Surface Shapes) large-scale 3D human body motion capture dataset is used as the basis for generative adversarial training, and ResNet is used to extract the image features, and a gated recurrent network is used as the time encoder. The gated recurrent network is used as a time encoder, the parameter regressor regresses the SMPL parameter model, and the attention mechanism is used to amplify the degree of contribution of different frames of the sequence to give the probability of estimating the pose to be true or false on the basis of the human body's real motion AMASS, and then construct the Dunhuang dance 3D action database. This database divides Dunhuang dance into 7 themes, 83 basic movements and 16 long movements. Better results were obtained in quantitative, qualitative and manual evaluation, which laid the foundation for the preservation, application and development of Dunhuang dance; and provided new ideas for the research, promotion and inheritance of Dunhuang culture. Subsequent utilization of this database can be applied to generative artificial intelligence, digital exhibition and performance of Dunhuang dance culture, education and research of Dunhuang dance, and digital media and entertainment of Dunhuang dance.

敦煌舞是中国传统文化的艺术瑰宝，源自敦煌壁画，是敦煌文化的重要组成部分，其数字化保存、展示与研究均具有重要意义。为推动敦煌舞的数字化进程与发展，本研究提出将敦煌舞与三维人体姿态估计技术相结合，构建敦煌舞三维动作数据库。针对敦煌舞环节，本研究依据人物形象对敦煌舞进行拆分、编创与动作采集，构建敦煌舞视频动作数据库；针对姿态估计任务，本研究以AMASS（Archive of Motion Capture as Surface Shapes）大型三维人体运动捕捉数据集作为生成对抗训练的基础，采用ResNet（残差网络）提取图像特征，将门控循环网络作为时序编码器，通过参数回归器对SMPL参数模型完成参数回归，并利用注意力机制放大序列中不同帧的贡献权重，基于人体真实运动的AMASS数据集生成姿态估计结果的真伪概率，进而构建敦煌舞三维动作数据库。该数据库将敦煌舞划分为7大主题、83个基础动作与16个长时长动作。在定量、定性与人工评估中均取得了优异效果，为敦煌舞的保存、应用与发展奠定了坚实基础，并为敦煌文化的研究、推广与传承提供了全新思路。后续可依托该数据库开展多场景应用，包括敦煌舞文化的生成式人工智能（Generative AI）创作、敦煌舞文化的数字化展示与演出、敦煌舞的教育与研究，以及敦煌舞数字媒体与娱乐等领域。

搜集汇总

背景与挑战

背景概述

敦煌舞三维动作数据集是一个结合三维人体姿态估计技术的数据库，包含7个主题、83个基本动作和16个长动作，总数据量为1.05GB。该数据集旨在促进敦煌舞的数字化保存、应用和发展，适用于生成式人工智能、数字展示、教育研究等多个领域。