Skip to content

Latest commit

 

History

History
127 lines (91 loc) · 3.19 KB

README.md

File metadata and controls

127 lines (91 loc) · 3.19 KB
Raw

GNet-pose

Project Page: http://guanghan.info/projects/guided-fractal/

Overview

Knowledge-Guided Deep Fractal Neural Networks for Human Pose Estimation.

Source code release of the paper for reproduction of experimental results, and to aid researchers in future research.

Prerequisites

Getting Started

1. Download Data and Pre-trained Models

  • Datasets (MPII [1], LSP [2])

    bash ./get_dataset.sh

  • Models

    bash ./get_models.sh

  • Predictions (optional)

    bash ./get_preds.sh

2. Testing

  • Generate cropped patches from the dataset for testing:

    cd testing/
matlab gen_cropped_LSP_test_images.m
matlab gen_cropped_MPII_test_images.m
cd -

    This will generate images with 368-by-368 resolution.

  • Reproduce the results with the pre-trained model:

    cd testing/
python .test.py
cd -

    You can choose different dataset to test on, with different models. You can also choose different settings in test.py, e.g., with or without flipping, scaling, cross-heatmap regression, etc.

3. Training

  • Generate Annotations

    cd training/Annotations/
matlab MPI.m LEEDS.m
cd -

    This will generate annotations in json files.

  • Generate LMDB

    python ./training/Data/genLMDB.py

    This will load images from dataset and annotations from json files, and generate lmdb files for caffe training.

  • Generate Prototxt files (optional)

    python ./training/GNet/scripts/gen_GNet.py
python ./training/GNet/scripts/gen_fractal.py
python ./training/GNet/scripts/gen_hourglass.py

  • Training:

     bash ./training/train.sh

4. Performance Evaluation

cd testing/eval_LSP/; matlab test_evaluation_lsp.m; cd../

cd testing/eval_MPII/; matlab test_evaluation_mpii_test.m

5. Results

More Qualitative results can be found in the project page. Quantitative results please refer to the arxiv paper.

License

GNet-pose is released under the Apache License Version 2.0 (refer to the LICENSE file for details).

Citation

The details are published as a technical report on arXiv. If you use the code and models, please cite the following paper: arXiv:1705.02407.

@article{ning2017knowledge,
  title={Knowledge-Guided Deep Fractal Neural Networks for Human Pose Estimation},
  author={Ning, Guanghan and Zhang, Zhi and He, Zhihai},
  journal={arXiv preprint arXiv:1705.02407},
  year={2017}
}

Reference

[1] Andriluka M, Pishchulin L, Gehler P, et al. "2d human pose estimation: New benchmark and state of the art analysis." CVPR (2014).

[2] Sam Johnson and Mark Everingham. "Clustered Pose and Nonlinear Appearance Models for Human Pose Estimation." BMVC (2010).