This repository contains pseudo-GT 3D human pose data produced by Exemplar Fine-Tuning (EFT) method, published in 3DV 2021. The 3D pose data is in the form of SMPL parameters, and this can be used as a supervision to train a 3D pose estimation algiritm (e.g., SPIN or HMR). We found that our EFT dataset is sufficient to build a model that is comparable to the previous SOTA algorithms without using any other indoor 3D pose dataset. See our paper for more details.

This repository also contains the pre-trained 3D pose estimation model trained with our EFT dataset and monocular motion capture demo tools. See README_bodymocap.

• We have released the EFT fitting codes. See the README_run_eft file:

• We have released FrankMocap by which you can obtain both 3D body+hand outputs. The body module is the same as this repository's model. We encourage to use FrankMocap for body pose estimation.

It is convenient and safe to use conda environment

conda create -n venv_eft python=3.6
conda activate venv_eft
pip install -r requirements.txt

This repository only provides corresponding SMPL parameters for public 2D keypoint datasets (such as COCO, MPII). You need to download images from the original dataset website.

Run the following script to download our EFT fitting data:

sh scripts/download_eft.sh 

• The EFT data will be saved in ./eft_fit/(DB_name).json. Each json file contains a version EFT fitting for a public dataset.
• See Data Format for details
• Currently available EFT fitting outputs:

• COCO2014-All-ver01.json: COCO 2014 training set by selecting the samples 6 keypoints or more keypoints are annotated.
• COCO2014-Part-ver01.json: COCO 2014 training set by selecting the sample that 12 limb keypoints or more are annotated.
• COCO2014-Val-ver10.json: COCO 2014 val set.
• MPII_ver01.json : MPII Keypoint Dataset
• PoseTrack_ver01.json : PoseTrack Dataset by selecting the sample that 12 limb keypoints or more are annotated.
• LSPet: LSPet Dataset
• OCHuman : OCHuman Dataset
• Note that the number of samples are fewer than the original sample numbers in each DB, since we automatically (or manually) filtered out bad samples
• Manual Filtering: Manual quality check and filtering is done to keep high quality results only. See paper for details

• SMPL Model (Neutral model: basicModel_neutral_lbs_10_207_0_v1.0.0.pkl):

  • Download in the original website. You need to register to download the SMPL data.
  • Put the file in: ./extradata/smpl/basicModel_neutral_lbs_10_207_0_v1.0.0.pkl
  • Copy the smpl pkl file to a different name (SMPL_NEUTRAL.pkl). You need both files:

  cp ./extradata/smpl/basicModel_neutral_lbs_10_207_0_v1.0.0.pkl ./extradata/smpl/SMPL_NEUTRAL.pkl
  

• Densepose (optional, for Densepose rendering):

  • Run the following script

      sh scriptsdownload_dp_uv.sh    
  

  • Files are saved in ./extradata/densepose_uv_data/

• COCO: 2014 Training set
• MPII: Download Link
• LSPet: Download Link

• See README_eft_vis

• See README_run_eft

• We also share pre-trained models trained with diverse dataset.
• Plaese see "scripts/download_model_zoo.sh"

• See README_bodymocap

@inproceedings{joo2020eft,
  title={Exemplar Fine-Tuning for 3D Human Pose Fitting Towards In-the-Wild 3D Human Pose Estimation},
  author={Joo, Hanbyul and Neverova, Natalia and Vedaldi, Andrea},
  booktitle={3DV},
  year={2020}
}

CC-BY-NC 4.0. See the LICENSE file.

The body mocap code is a modified version of SPIN, and the majority of this code is borrowed from it.