|
Folk dance is an artistic treasure of Chinese traditional culture with a long history, and it is of great significance to explore how to digitize folk dance for preservation, display and research. This paper proposes to combine folk dance with 3D human posture estimation technology to explore the feasibility of storing folk dance as 3D skeletal movement files BVH (Biovision Hierarchy, BVH), in order to advance the development of the digitization process of folk dance. The datasets of three dance types, Daizu dance, Zangzu dance and Dunhuang dance, which are publicly available and field-collected, are utilized in the ethnic dance aspect; the main objective of the 3D human pose estimation aspect is to predict the pose and shape of the 3D human body from an RGB image or a sequence of multiple RGB images, and at the same time, to learn the attentional weights of each body part, so as to solve the occlusion problem that exists in the complex environments or complex movements, and to improve the method's Robustness. The method contains two main branches, the 2D part branch and the 3D body branch, where the 2D part branch is responsible for generating the attention weights while the 3D body branch is used to predict the pose and shape parameters. The model introduces a part-based attention mechanism in training which uses partially segmented labels to guide the attention branch in the early stages of training and unsupervised in the later stages in such a way that the attention mechanism is able to obtain effective information from the image and surrounding pixels unsupervised, and to focus more flexibly on the useful regions in the pose estimation. The whole model is supervised by a series of regression losses, and achieves better results in quantitative, qualitative and manual evaluations, exploring a new way for the digital preservation and presentation of folk dances.
|
