README.md

Residual Pose: A Decoupled Approach for Depth-based 3D Human Pose Estimation

Introduction

This is the code release of our paper Residual Pose: A Decoupled Approach for Depth-based 3D Human Pose Estimation

If you happen to use the models and code for your work, please cite the following paper

@inproceedings{Martinez_IROS_2020,
author={Angel {Martínez-González} and Michael {Villamizar} and Olivier {Canévet} and Jean-Marc {Odobez}},
journal={IEEE/RSJ International Conference on Intelligent Robots and Systems},
title={Residual Pose: A Decoupled Approach for Depth-Based 3D Human Pose Estimation},
year={2020}
}

We include in this repo

• Hourglass model for multi-person 2D pose estimation from depth images.
• Our regressor NN architecture for 3D human pose estimation.
• 3D pose prior for recovering from 2D missed detections.
• Tranined models for 2D and 3D pose estimation.
• Code for obtaining 2D and 3D pose from a depth image.

Requirements

• Pytorch >= 1.3
• OpenCV >= 3.4
• Python >= 3.5

Datasets

We use the following datasets with single and multi-person scenarios. Please refer to our paper for more information

1. ITOP [link]
2. CMU-Panoptic [link]

Some images from these datasets have been included in the repo for demonstration purposes. Please refer to their sites and authors for complete access.

Testing

To test our pretrained models on data from the ITOP dataset run the following

python main.py --config_file config/itop_config_file.json \
               --image_sample img_samples/itop/ITOP_train_000000000053.npy \
               --output_path output_dir

and to test our pretrained models on data from the CMU-Panoptic dataset run the following

python main.py --config_file config/panoptic_config_file.json \
               --image_sample img_samples/panoptic/depth_02_000000005014.mat \
               --output_path output_dir

which will save visualizations and results on output_dir.