Contrastive learning-driven spatiotemporal dynamically adaptive framework for stylized 3D human motion generation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Computer Graphics

Background:

• Existing 3D human motion generation methods often overlook local stylistic variations, leading to generated sequences lacking expressive detail.
• Global temporal style statistics are insufficient for capturing the nuances of dynamic human movements.

Purpose of the Study:

• To propose a contrastive learning-driven framework for spatiotemporal dynamically adaptive stylized 3D human motion generation.
• To enhance the ability to capture local stylistic variations and improve the expressive detail in generated 3D human motions.

Main Methods:

• Introduced spatial attention instance normalization (SAIN) and temporal attention instance normalization (TAIN) to extract local and global motion style features.
• Employed a dual-path structure to isolate motion content and a style injector (SADA, TADA) for fine-grained style integration.
• Utilized style and content contrastive losses during training to improve feature clustering and separation.

Main Results:

• The proposed method achieved superior performance on the Xia dataset, with FID of 0.06, accuracy of 96.70%, diversity of 5.67, and multimodality of 0.97, closely matching real data.
• In motion style transfer tasks, the model attained 94.11 CRA and 89.41 SRA, outperforming existing state-of-the-art methods.

Conclusions:

• The developed framework effectively disentangles motion style and content, enabling fine-grained, dynamically adaptive stylized 3D human motion generation.
• The contrastive learning approach enhances stylistic diversity and content fidelity, producing more expressive and realistic human motions.