Human motion synthesis and compression

The synthesis of human motion plays a pivotal role in applications ranging from character animation to autonomous driving. Recent advances in human motion synthesis are driven by powerful denoising diffusion models and transformer architectures. This thesis explores two fundamental challenges in human motion synthesis: designing effective architectural frameworks and developing motion compression components with strong reconstruction capabilities and well-conditioned latent spaces. While the current transformer architectures are predominantly temporal-focused, the spatial structure is inherent in human body. we introduce Positional Mask-Guided Spatial-Temporal Fusion (\ours) -- a novel approach to modeling human motion in a bi-dimensional manner, thus enabling a more nuanced generation of human behavior. Specifically, we design a spatial-temporal transformer architecture with homogeneous and symmetric dual branches for learning representations from human motion sequences. To facilitate the refined interplay between spatial and temporal features, we propose positional masks to guide the fusion process. Extensive experiments demonstrate the state-of-the-art performance of \ours across tasks and datasets. Efficiently compressing human motion sequences allows for a significant reduction in computational overhead and facilitates more complex analyses and synthesis in constrained environments. In order to build an effective two-person motion compression model, researchers should identify the crucial loss terms, adapt adequate network architecture, and control the variance in the latent space. Through exhaustive experiments, the thesis offers deep insights into the optimal design of motion compression systems for future applications. From those two aspects, the work paves the way for the future research in human motion synthesis.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Li, Zhengyuan

Contributors dc:contributor

• Gui, Liangyan

• Human Motion

Statement dc:rights

• Copyright 2024 Zhengyuan Li

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/124589

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F124589