Expressive Forecasting of 3D Whole-Body Human Motions
Authors: Pengxiang Ding, Qiongjie Cui, Haofan Wang, Min Zhang, Mengyuan Liu, Donglin Wang
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on a newly-introduced large-scale benchmark and achieve state-of-the-art performance. |
| Researcher Affiliation | Collaboration | Pengxiang Ding1,2, Qiongjie Cui3,5*, Haofan Wang5, Min Zhang1,2, Mengyuan Liu4, Donglin Wang1 1Mi LAB, Westlake University 2Zhejiang University 3Nanjing University of Science and Technology 4Shenzhen Graduate School, Peking University 5Xiaohongshu Inc. |
| Pseudocode | No | The paper describes the proposed method textually and with mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is public for research purposes at https://github.com/Dingpx/EAI. |
| Open Datasets | Yes | To be compatible with our proposed novel task, here we select the GRAB (Taheri et al. 2020). It is a recently released dataset with over 1.6 million frames of 10 different actors performing a total of 29 actions. GRAB provides SMPL-X (Pavlakos et al. 2019) parameters from which we extract 25 joints (3D position) defined as the body (Nm = 25), and each hand is represented as 15-joints (Nl = Nr = 15). Also, previous widely-used datasets, e.g., H3.6M (Ionescu et al. 2013), 3DPW (von Marcard et al. 2018), only record the major body motions (without human hands). |
| Dataset Splits | Yes | We apply two training strategies to investigate this new task. (1) For the divided (D) training, we separately train the baselines for each human components. This independent strategy lacks the interaction of components and thus can be used to illustrate the effectiveness of XCI. (2) For the united (D) training, we extend the node number of GCNs to 55 (Nm = 25, Nl = Nr = 15), as in our experimental setup. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory) used for running its experiments. It only implies standard computational resources for training. |
| Software Dependencies | No | The paper mentions employing 'Adam W (Loshchilov and Hutter 2017) optimizer' but does not specify software versions for libraries (e.g., PyTorch, TensorFlow) or programming languages (e.g., Python version) used for implementation. |
| Experiment Setup | Yes | We employ Adam W (Loshchilov and Hutter 2017) optimizer with an initial learning rate of 0.001 and batch size of 64 to train our model (50 epochs). The learning rate is decayed by 0.96 for every two epochs. The trade-off parameters {λ1, λ2, λ3} are set as {1, 0.1, 0.001}. |