Practical overview
When it fits
Use MS-G3D when working on action classification from skeletal pose sequences and you need a research implementation with documented data preparation, training and evaluation workflows.
Who it is for
- Computer-vision researchers experimenting with skeleton-based action recognition.
- Python and PyTorch practitioners who want to train or evaluate graph-convolution models on NTU RGB+D or Kinetics skeleton datasets.
Documented uses
- Prepare supported skeleton datasets and generate bone data for model workflows.
- Train and test models using dataset-specific configuration files.
- Evaluate the project’s pretrained models.
- Combine joint- and bone-stream results with the documented fusion workflow.
Requirements & limits
The repository specifies the MIT license. The README lists Python 3.6 or later, PyTorch 1.2.0 or later, NVIDIA Apex for automatic mixed-precision training, PyYAML, tqdm and tensorboardX.
- The documented input is skeletal data; the repository does not describe a workflow that takes raw video directly.
- Dataset preparation can require substantial storage, with README estimates of 38GB, 77GB and 63GB after preprocessing for NTU RGB+D 60, NTU RGB+D 120 and Kinetics 400, respectively.
- The README provides command templates and configuration-based workflows, so users should be comfortable preparing data and working with Python model code.
- For NTU datasets, the README directs users to request access from the dataset provider.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare skeleton data for MS-G3D
Generate processed NTU RGB+D skeleton and bone data for the repository’s training and evaluation workflows.
Before you start
- Python 3.6 or later and the repository dependencies, including PyTorch, PyYAML, tqdm and tensorboardX.
- NTU RGB+D skeleton data. The README directs users to request the datasets from their provider and download the skeleton-only archives.
- The corresponding missing-skeleton lookup files for NTU RGB+D 60 or 120.
- Substantial disk space: the README estimates about 38GB for processed NTU RGB+D 60 and 77GB for NTU RGB+D 120; raw and intermediate files may require more.
Steps
Place the downloaded NTU skeleton files and missing-skeleton lookup files in the directory structure shown in the README.
From the data_gen directory, run python3 ntu_gendata.py for NTU RGB+D 60 or python3 ntu120_gendata.py for NTU RGB+D 120.
Generate bone data with python gen_bone_data.py --dataset ntu or python gen_bone_data.py --dataset ntu120, as appropriate.
Expected outcome
Processed NTU skeleton data and corresponding bone data for the repository’s model workflows.
Guide limitations
- The recipe covers data preparation, not model training or evaluation.
- The README estimates about three hours to generate NTU 120 data on a single core; processing time and storage needs can vary.
- NTU data access requires a request to the dataset provider.
Evaluate a pretrained MS-G3D model
Run the repository’s documented pretrained-model evaluation workflow.
Before you start
- A prepared repository environment with Python and the project dependencies.
- Pretrained model files for one of the documented datasets: NTU RGB+D 60, NTU RGB+D 120 or Kinetics Skeleton 400.
- The corresponding dataset and configuration files required by the chosen evaluation workflow.
Steps
Download the pretrained models using one of the links listed in the repository README.
Place the pretrained-models folder at the repository root, alongside main.py.
Run bash eval_pretrained.sh.
Expected outcome
The repository’s pretrained evaluation workflow runs against the supplied model files and available dataset configuration.
Guide limitations
- The README does not specify a single configuration-independent result; dataset preparation and matching files are needed for the chosen evaluation.
- The README notes that PyTorch, CUDA and GPU differences can affect results.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- kenziyuliu/MS-G3D · Official homepage title · retained in its source language