A Unified Framework for Real Time Motion Completion
Authors: Yinglin Duan, Yue Lin, Zhengxia Zou, Yi Yuan, Zhehui Qian, Bohan Zhang4459-4467
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our experiment, we evaluate our method across three different motion completion tasks: 1. In-betweening on La FAN1 (Harvey et al. 2020)... 2. In-filling on Anidance (Tang, Mao, and Jia 2018)... 3. Blending on our newly proposed dance dataset... |
| Researcher Affiliation | Collaboration | Yinglin Duan1, Yue Lin1, Zhengxia Zou2 , Yi Yuan3, Zhehui Qian3, Bohan Zhang3 1 Net Ease Games AI Lab, Guangzhou, China, 2 Beihang University, 3 Netease, Inc., Hangzhou, China {duanyinglin, gzlinyue, yuanyi, qianzhehui, hzzhangbohan}@corp.netease.com, zhengxiazou@buaa.edu.cn |
| Pseudocode | No | The paper does not include pseudocode or clearly labeled algorithm blocks; the method is described in text and equations. |
| Open Source Code | Yes | 1Our project is available at: https://github.com/SilvanDuan/MotionCompletion |
| Open Datasets | Yes | 1. In-betweening on La FAN1 (Harvey et al. 2020): La FAN1 is a public high-quality general motion dataset introduced by Harvey et al. from Ubisoft. ... 2. In-filling on Anidance (Tang, Mao, and Jia 2018): Anidance is a public music-dance dataset proposed by Tang et al. (Tang, Mao, and Jia 2018). |
| Dataset Splits | No | The paper explicitly states 'training set' and 'test set' splits with specific counts and percentages for La FAN1 and Anidance datasets, but it does not explicitly mention a separate 'validation' split or its details (e.g., percentages or counts) as part of the dataset partitioning. |
| Hardware Specification | Yes | On a single CPU desktop (I7-8700K @ 3.70GHz), our method can run in real time (40 motion sequences per second, each with 50 frames long), 4x faster than previous methods. |
| Software Dependencies | No | The paper mentions that 'The whole framework is implemented using Py Torch (Paszke et al. 2019)', but it does not specify a version number for PyTorch or any other software libraries. |
| Experiment Setup | Yes | We adopt BERT (Devlin et al. 2018) as the backbone of our transformer with 8 encoder layers. In each encoder layer, we set the number of attention heads to M = 8. For our input and output Conv1D layers, the kernel size is set to 3 and the padding is set to 1. We set the dimension of the feature embedding in the MHSA layers to 256, and set those in the FFN layers to 512. In our training loss, we set αrec = αper = 1.0 and αK = 0.01. ... We train our network by using Adam optimizer (Kingma and Ba 2014). We set the maximum learning rate to 10 3. |