Skip to content

Code and dataset of human retargeting approach to transfer motion and appearance between monocular videos. Check the examples from our dataset with several paired actors and motions!

Notifications You must be signed in to change notification settings

verlab/ShapeAwareHumanRetargeting_IJCV_2021

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

13 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Shape-Aware Human Retargeting in Monocular Videos
[Project Page] [Paper]

Python and TensorFlow code implementation of the shape-aware human motion and appearance transference between monocular videos. The project webpage of the paper is https://verlab.github.io/ShapeAwareHumanRetargeting_IJCV_2021/

Introduction

This repository contains the original implementation of the human retargeting approach and dataset of the paper: "A shape-aware retargeting approach to transfer human motion and appearance in monocular videos". The method is composed of four main components and synthesizes new videos of people in a different context where they were initially recorded. Differently from recent human neural rendering methods, our approach takes into account jointly body shape, appearance and motion constraints in the transfer.

Please cite this code and paper in your publications if it helps your research:

@Article{Gomes2021,
author={Gomes, Thiago L. and Martins, Renato and Ferreira, Jo{\~a}o and Azevedo, Rafael and Torres, Guilherme and Nascimento, Erickson R.},
title={A Shape-Aware Retargeting Approach to Transfer Human Motion and Appearance in Monocular Videos},
journal={International Journal of Computer Vision},
year={2021},
month={Apr},
day={29},
issn={1573-1405},
doi={10.1007/s11263-021-01471-x},
url={https://doi.org/10.1007/s11263-021-01471-x}
}

Setup and Installation

We provide install and running examples for the main components of the approach. For the motion reconstruction and retargeting, please follow the instructions given in README_Retargeting.md.

For the texture extraction, body geometry refinement and the image-based rendering into the background of the source video, please follow the instructions provided in README_DeformationRender.md.

Human Retargeting Dataset

We created a dataset with several paired human motion sequences from different actors and annotated motion retargeting constraints. The dataset contains 8 actors (S0-S7) performing eight different actions. Link to download

dataset actors

The dataset sequences are divided into two sets: training and test.

Training Set

For each actor we provide 4 types of data: 4-min videos; 8-views; A-pose and dance.

  • 4-min: 4 minutes of random movements performed by each actor. The directory contains a folder called images with the .jpg images, a text file (person.txt) with the heigth and smpl parameters, and a consensus of the smpl shape (.pkl).

  • 8-views: This folder contains 8-views from each actor with 45 degree variation between each pose. It has three additional folders: segmentations with binary masks of the actors; semantic_label containing the human body parts labeled semantically; and smpl_pose with the SPIN pose estimations.

  • A-pose: Original 360 degree movements used to build the 8-views folder.

  • dance: Random dance movements performed by each actor.

train
└───S0
└───S1
|   ...
└───S7    
│   └─── 4-min
|       | person.txt
|       | smpl_consensus_shape.pkl
|       └─── images
|   |        
|   └─── 8-views
|       | person.txt
|       | smpl_consensus_shape.pkl
|       └─── images
|       └─── segmentations
|       └─── semantic_label
|       └─── smpl_pose   
|   |
|   └─── A-pose
|       | person.txt
|       | smpl_consensus_shape.pkl
|       └─── images
|   |
|   └─── dance
|       | person.txt
|       | smpl_consensus_shape.pkl
|       └─── images

Test Set

For the test set, we provide 8 movements performed by each actor:

  • box
  • cone
  • fusion dance
  • jump
  • pull down
  • shake hands
  • spinning
  • walk

test movements

Each movement folder contains a text file (person.txt) with the height and SPML parameters, as well as the annotated 2D and 3D motion restrictions. There is also three sub-folders containing: i) images from the scene; ii) openpose with the actors 2D poses from OpenPose/SPIN; and iii) smpl_pose with the corresponding SMPL model estimation.

test
└───S0
└───S1
|   ...
└───S7 
    └─── box
    └─── cone
    | ...
    └─── walk
          | person.txt
          | restrictions-3D.json
          └─── images
          └─── openpose
          └─── smpl_pose

License & Disclaimer

This is research code, expect that it can change and any fitness for a particular purpose is disclaimed. This software is under GNU General Public License Version 3 (GPLv3).


About

Code and dataset of human retargeting approach to transfer motion and appearance between monocular videos. Check the examples from our dataset with several paired actors and motions!

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published