Temporal Smoothing for 3D Human Pose Estimation and Localization for Occluded People

This repository contains the code for the paper "Temporal Smoothing for 3D Human Pose Estimation and Localization for Occluded People" link.

Requirements

The code was tested with the following libraries:

• PyTorch 1.5.1
• numpy 1.16.4
• scipy 0.18.1

Usage

Preparations

First you have to download the preprocessed dataset from here and unzip it in the root folder. For evaluation, you also need to download the MuPoTS dataset. For training, MPII-3DHP is also needed. After extraction, the folder structure should look like this:

root
 +-- datasets
 |    +-- Mpi3DHP
 |    |    +-- mpi_inf_3dhp_test_set
 |    |    +-- S1
 |    |    ...
 |    +-- MucoTemp
 |    |    +-- frames
 |    |    +-- hrnet_keypoints
 |    +-- MuPoTS
 |         +-- hrnet_pose2d
 |         +-- MultiPersonTestSet
 +-- src

You can also download the pretrained models from here.

Running

To evaluate the pretrained models, simply run the following command:

root_dir/src$ python3 scripts/eval.py -r normal

The -r switch controls whether to use pose refinement or not. In the paper, mm based metrics were calculated on a model using the 'normal' MPII coordinates, while 3DPCK was trained with 'universal' coordinates. If you want to evaluate the latter, use universal instead of normal in the command line above.

To train the model, use train.py. The parameters can be set in the script:

root_dir/src$ python3 scripts/train.py

After, to evaluate the model, use:

root_dir/src$ python3 scripts/eval.py -r ../output

Muco-Temp

To save space, the frames of the MuCo-Temp dataset are not included, only the pregenerated 2D pose estimations. If you need the frames, you can use generate_muco_temp.py to recreate them. You'll need the MPII-3DHP dataset downloaded.

Inference on videos

To run the model on videos, you have to get Detectron2 and HR-Net. The following script downloads them and creates a Conda environment with all the necessary dependencies:

(base) root_dir$ ./install_dependencies

You also need to download the pretrained models and unzip them in the root directory. Once done, you can create predictions using the following command:

(base) root_dir$ conda activate pose-refinement
(pose-refinement) root_dir$ cd src
(pose-refinement) root_dir/src$ python scripts/predict.py -r video.mp4 output.pkl

The -r switch controls whether to use pose refinement or not. If you have the camera calibration parameters you can also provide them using the -f, -cx/-cy arguments.

Citatation

@InProceedings{veges2020temporal,
  author="V{\'e}ges, M. and L{\H{o}}rincz, A.",
  title="Temporal Smoothing for 3D Human Pose Estimation and Localization for Occluded  People",
  booktitle="Neural Information Processing",
  year="2020",
  pages="557--568",
}