GNet-pose

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

UPDATE 9/27/2018:

Prototxts and model that achieved 93.9Pck on LSP dataset. http://guanghan.info/download/Data/GNet_update.zip

When I was replying e-mails, it occurred to me that the models that I had uploaded was around May/June 2017 (performance in old arxiv version), and in August 2017 the performance was improved to 93.9 on LSP with a newer caffe version which fixed the downsampling and/or upsampling deprecation problem (Yeah, it "magically" improved the performance). The best model was 94.0071 on LSP dataset, but it was not uploaded nor published on the benchmark.

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

• Python 2.7 or Python 3.3+
• Modified Caffe

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

If you use the code and models, please cite the following paper: TMM 2017.

@article{ning2017knowledge, 
 author={G. Ning and Z. Zhang and Z. He}, 
     journal={IEEE Transactions on Multimedia}, 
     title={Knowledge-Guided Deep Fractal Neural Networks for Human Pose Estimation}, 
     year={2017}, 
     doi={10.1109/TMM.2017.2762010}, 
     ISSN={1520-9210}, }

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).