Generative Human Motion Stylization in Latent Space

Authors: chuan guo, Yuxuan Mu, Xinxin Zuo, Peng Dai, Youliang Yan, Juwei Lu, Li Cheng

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that our proposed stylization models, despite their lightweight design, outperform the state-of-the-arts in style reeanactment, content preservation, and generalization across various applications and settings.
Researcher Affiliation Collaboration 1University of Alberta, 2Noah s Ark Lab, Huawei Canada
Pseudocode No No pseudocode or algorithm block was found in the paper.
Open Source Code Yes Code and Model. The code of our approach and implemented baselines are also submitted for reference. Code and trained model will be publicly available upon acceptance.
Open Datasets Yes We adopt three datasets for comprehensive evaluation. (Aberman et al., 2020) is a widely used motion style dataset, which contains 16 distinct style labels including angry, happy, Old, etc, with total duration of 193 minute. (Xia et al., 2015) is much smaller motion style collection (25 mins) that is captured in 8 styles, with accurate action type annotation (8 actions). The other one is CMU Mocap (CMU), an unlabeled dataset with high diversity and quantity of motion data. All motion data is retargeted to the same 21-joint skeleton structure, with a 10% held-out subset for evaluation.
Dataset Splits No All motion data is retargeted to the same 21-joint skeleton structure, with a 10% held-out subset for evaluation.
Hardware Specification Yes Table 5 presents the comparisons of average time cost for a single forward pass with 160-frame motion inputs, evaluated on a single Tesla P100 16G GPU.
Software Dependencies No Our models are implemented by Pytorch.
Experiment Setup Yes The values of "lambda"l kld, "lambda"l1 and "lambda"sms are all set to 0.001, and dimension Dz of z is 512. During training our latent stylization network, the value of "lambda"hsa, "lambda"cyc and "lambda"kl are (1, 0.1, 0.1) and (0.1, 1, 0.01) in supervised setting and unsupervised setting, respectively.