LitePose is an efficient, scale invariant, multi-person pose estimation model. Its light architecture allows to perform real-time inference with low computational power devices. This is a (non official) implementation of the original LitePose.
| Keypoint detection | Pose estimation |
|---|---|
 |
 |
LitePose follows a bottom-up pose estimation approach. The single-branch architecture ensures high efficiency, whereas the Fusion Deconv Head implements the scale invariance by using high resolution features. MobileNet structure with large kernels convolution is used as backbone. The whole network is scalable according to the number of joints and the maximum number of people that the image may contain.
Clone the repository:
git clone https://github.com/DaniAffCH/litepose-pose-estimation
Install python requirements:
pip install -r requirements.txt
Download both annotations and images of CrowdPose Dataset from the official repository and install CrowdPose APIs.
Then recreate a directory structure as:
crowdpose
├─── images
│ ├── 112934.jpg
│ ...
└── json
├── crowdpose_test.json
├── crowdpose_train.json
├── crowdpose_trainval.json
└── crowdpose_val.json
Finally edit src/lp_config/lp_common_config.py and modify the variable dataset_root with your installation path (default is ~/dataset/crowdpose).
In order to check if the installation was successful you can run python main.py --test, if it passes all the test cases then the set up is working correctly.
Every setting can be modified in src/lp_config:
lp_model_config.pycontains the settings about the network architecture.lp_common_config.pycontains the general configurations about training and inference.
If you modify the network structure you have to train the custom network, otherwise pretrained models are available in src/lp_trained_models.
For training run:
python main.py --train
For inference run:
python main.py --inference lp_trained_models/bigarch
For model evaluation run:
python main.py --score lp_trained_models/bigarch
You can replace lp_trained_models/bigarch with any trained model
The file demo.ipynb is a notebook that shows an example of code usage and provides further details about the project.
Contributions are welcome! If you find any issues or have suggestions for improvements, please open an issue or submit a pull request.
This project is licensed under the MIT License.
This work is based on LitePose paper and HigherHRNet for the network architecture. Moreover, it uses the Associative Embedding to cluster the keypoints.