Commit 9cb48a85 authored by Konstantinos Papadopoulos's avatar Konstantinos Papadopoulos
Browse files


parent a2aae2de
# Spatial Temporal Graph Convolutional Networks (ST-GCN)
A graph convolutional network for skeleton based action recognition.
# DeepVI
This repository contains the implementation of DeepVI framework that is based on the original ST-GCN network [1]. We used the VNect pose estimator [2] to generate 3D skeletons from the provided RGB videos.
<div align="center">
<img src="resource/info/pipeline.png">
In this repo, we show the example of our model on NTU-RGB+D dataset.
This repository holds the codebase, dataset and models for the paper>
Please, refer to for more information on how to do the training and testing on the dataset.
**Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition** Sijie Yan, Yuanjun Xiong and Dahua Lin, AAAI 2018.
[1] Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition, Sijie Yan, Yuanjun Xiong and Dahua Lin, AAAI 2018.
[[Arxiv Preprint]](
## News & Updates
- Feb. 21, 2019 - We provide pretrained models and training scripts on **NTU-RGB+D** and **kinetics-skeleton** datasets. So that you can achieve the performance we mentioned in the paper.
- June. 5, 2018 - A demo for feature visualization and skeleton based action recognition is released.
- June. 1, 2018 - We update our code base and complete the PyTorch 0.4.0 migration.
## Visulization of ST-GCN in Action
Our demo for skeleton based action recognition:
<p align="center">
<img src="resource/info/demo_video.gif", width="1200">
ST-GCN is able to exploit local pattern and correlation from human skeletons.
Below figures show the neural response magnitude of each node in the last layer of our ST-GCN.
<table style="width:100%; table-layout:fixed;">
<td><img width="150px" src="resource/info/S001C001P001R001A044_w.gif"></td>
<td><img width="150px" src="resource/info/S003C001P008R001A008_w.gif"></td>
<td><img width="150px" src="resource/info/S002C001P010R001A017_w.gif"></td>
<td><img width="150px" src="resource/info/S003C001P008R001A002_w.gif"></td>
<td><img width="150px" src="resource/info/S001C001P001R001A051_w.gif"></td>
<td><font size="1">Touch head<font></td>
<td><font size="1">Sitting down<font></td>
<td><font size="1">Take off a shoe<font></td>
<td><font size="1">Eat meal/snack<font></td>
<td><font size="1">Kick other person<font></td>
<td><img width="150px" src="resource/info/hammer_throw_w.gif"></td>
<td><img width="150px" src="resource/info/clean_and_jerk_w.gif"></td>
<td><img width="150px" src="resource/info/pull_ups_w.gif"></td>
<td><img width="150px" src="resource/info/tai_chi_w.gif"></td>
<td><img width="150px" src="resource/info/juggling_balls_w.gif"></td>
<td><font size="1">Hammer throw<font></td>
<td><font size="1">Clean and jerk<font></td>
<td><font size="1">Pull ups<font></td>
<td><font size="1">Tai chi<font></td>
<td><font size="1">Juggling ball<font></td>
The first row of above results is from **NTU-RGB+D** dataset, and the second row is from **Kinetics-skeleton**.
## Prerequisites
Our codebase is based on **Python3** (>=3.5). There are a few dependencies to run the code. The major libraries we depend are
- [PyTorch]( (Release version 0.4.0)
- [Openpose@92cdcad]( (Optional: for demo only)
- FFmpeg (Optional: for demo only), which can be installed by `sudo apt-get install ffmpeg`
- Other Python libraries can be installed by `pip install -r requirements.txt`
### Installation
cd torchlight; python install; cd ..
### Get pretrained models
We provided the pretrained model weithts of our **ST-GCN**. The model weights can be downloaded by running the script
bash tools/
<!-- The downloaded models will be stored under ```./models```. -->
You can also obtain models from [GoogleDrive]( or [BaiduYun](, and manually put them into ```./models```.
## Demo
To visualize how ST-GCN exploit local correlation and local pattern, we compute the feature vector magnitude of each node in the final spatial temporal graph, and overlay them on the original video. **Openpose** should be ready for extracting human skeletons from videos. The skeleton based action recognition results is also shwon thereon.
Run the demo by this command:
python demo --openpose <path to openpose build directory> [--video <path to your video> --device <gpu0> <gpu1>]
A video as above will be generated and saved under ```data/demo_result/```.
## Data Preparation
We experimented on two skeleton-based action recognition datasts: **Kinetics-skeleton** and **NTU RGB+D**.
### Kinetics-skeleton
[Kinetics]( is a video-based dataset for action recognition which only provide raw video clips without skeleton data. Kinetics dataset include To obatin the joint locations, we first resized all videos to the resolution of 340x256 and converted the frame rate to 30 fps. Then, we extracted skeletons from each frame in Kinetics by [Openpose]( The extracted skeleton data we called **Kinetics-skeleton**(7.5GB) can be directly downloaded from [GoogleDrive]( or [BaiduYun](
After uncompressing, rebuild the database by this command:
python tools/ --data_path <path to kinetics-skeleton>
NTU RGB+D can be downloaded from [their website](
Only the **3D skeletons**(5.8GB) modality is required in our experiments. After that, this command should be used to build the database for training or evaluation:
python tools/ --data_path <path to nturgbd+d_skeletons>
where the ```<path to nturgbd+d_skeletons>``` points to the 3D skeletons modality of NTU RGB+D dataset you download.
## Testing Pretrained Models
<!-- ### Evaluation
Once datasets ready, we can start the evaluation. -->
To evaluate ST-GCN model pretrained on **Kinetcis-skeleton**, run
python recognition -c config/st_gcn/kinetics-skeleton/test.yaml
For **cross-view** evaluation in **NTU RGB+D**, run
python recognition -c config/st_gcn/ntu-xview/test.yaml
For **cross-subject** evaluation in **NTU RGB+D**, run
python recognition -c config/st_gcn/ntu-xsub/test.yaml
<!-- Similary, the configuration file for testing baseline models can be found under the ```./config/baseline```. -->
To speed up evaluation by multi-gpu inference or modify batch size for reducing the memory cost, set ```--test_batch_size``` and ```--device``` like:
python recognition -c <config file> --test_batch_size <batch size> --device <gpu0> <gpu1> ...
### Results
The expected **Top-1** **accuracy** of provided models are shown here:
| Model| Kinetics-<br>skeleton (%)|NTU RGB+D <br> Cross View (%) |NTU RGB+D <br> Cross Subject (%) |
| :------| :------: | :------: | :------: |
|Baseline[1]| 20.3 | 83.1 | 74.3 |
|**ST-GCN** (Ours)| **31.6**| **88.8** | **81.6** |
[1] Kim, T. S., and Reiter, A. 2017. Interpretable 3d human action analysis with temporal convolutional networks. In BNMW CVPRW.
## Training
To train a new ST-GCN model, run
python recognition -c config/st_gcn/<dataset>/train.yaml [--work_dir <work folder>]
where the ```<dataset>``` must be ```ntu-xsub```, ```ntu-xview``` or ```kinetics-skeleton```, depending on the dataset you want to use.
The training results, including **model weights**, configurations and logging files, will be saved under the ```./work_dir``` by default or ```<work folder>``` if you appoint it.
You can modify the training parameters such as ```work_dir```, ```batch_size```, ```step```, ```base_lr``` and ```device``` in the command line or configuration files. The order of priority is: command line > config file > default parameter. For more information, use ``` -h```.
Finally, custom model evaluation can be achieved by this command as we mentioned above:
python recognition -c config/st_gcn/<dataset>/test.yaml --weights <path to model weights>
[2] VNect: Real-time 3D Human Pose Estimation with a Single RGB Camera, Mehta, Dushyant and Sridhar, Srinath and Sotnychenko, Oleksandr and Rhodin, Helge and Shafiei, Mohammad and Seidel, Hans-Peter and Xu, Weipeng and Casas, Dan and Theobalt, Christian, ACM ToG 2017
## Citation
Please cite the following paper if you use this repository in your reseach.
title = {Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition},
author = {Sijie Yan and Yuanjun Xiong and Dahua Lin},
booktitle = {AAAI},
year = {2018},
title={Deepvi: A novel framework for learning deep view-invariant human action representations using a single rgb camera},
author={Papadopoulos, Konstantinos and Ghorbel, Enjie and Oyedotun, Oyebade and Aouada, Djamila and Ottersten, Bj{\"o}rn},
booktitle={IEEE International Conference on Automatic Face and Gesture Recognition, Buenos Aires},
## Contact
For any question, feel free to contact
Sijie Yan :
Yuanjun Xiong :
Konstantinos Papadopoulos:
Enjie Ghorbel :
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment