October 15, 2019

801 words 4 mins read



Official implementation of “PifPaf: Composite Fields for Human Pose Estimation” in PyTorch.

repo name vita-epfl/openpifpaf
repo link https://github.com/vita-epfl/openpifpaf
homepage https://arxiv.org/abs/1903.06593
language Python
size (curr.) 19516 kB
stars (curr.) 458
created 2019-02-21
license Other


Continuously tested on Linux, MacOS and Windows: Build Status CVPR 2019 paper, arxiv.org/abs/1903.06593

We propose a new bottom-up method for multi-person 2D human pose estimation that is particularly well suited for urban mobility such as self-driving cars and delivery robots. The new method, PifPaf, uses a Part Intensity Field (PIF) to localize body parts and a Part Association Field (PAF) to associate body parts with each other to form full human poses. Our method outperforms previous methods at low resolution and in crowded, cluttered and occluded scenes thanks to (i) our new composite field PAF encoding fine-grained information and (ii) the choice of Laplace loss for regressions which incorporates a notion of uncertainty. Our architecture is based on a fully convolutional, single-shot, box-free design. We perform on par with the existing state-of-the-art bottom-up method on the standard COCO keypoint task and produce state-of-the-art results on a modified COCO keypoint task for the transportation domain.


example image with overlaid pose skeleton

Image credit: “Learning to surf” by fotologic which is licensed under CC-BY-2.0. Created with: python3 -m openpifpaf.predict --show docs/coco/000000081988.jpg

More demos:


Python 3 is required. Python 2 is not supported. Do not clone this repository and make sure there is no folder named openpifpaf in your current directory.

pip3 install openpifpaf

For a live demo, we recommend to try the openpifpafwebdemo project. Alternatively, openpifpaf.webcam provides a live demo as well. It requires OpenCV.

For development of the openpifpaf source code itself, you need to clone this repository and then:

pip3 install numpy cython
pip3 install --editable '.[train,test]'

The last command installs the Python package in the current directory (signified by the dot) with the optional dependencies needed for training and testing.


Tools to work with models:

Pre-trained Models

Performance metrics with version 0.10.1 on the COCO val set obtained with a GTX1080Ti:

Backbone AP APᴹ APᴸ t_{total} [ms] t_{dec} [ms]
shufflenetv2x2 60.4 55.5 67.8 56 33
resnet50 64.4 61.1 69.9 76 32
(v0.8) resnext50 63.8 61.1 68.1 93 33
resnet101 67.8 63.6 74.3 97 28
(v0.8) resnet152 67.8 64.4 73.3 122 30

Pretrained model files are shared in the releases of the openpifpaf-torchhub repository. The pretrained models are downloaded automatically when using the command line option --checkpoint backbonenameasintableabove.

To visualize logs:

python3 -m openpifpaf.logs \
  outputs/resnet50block5-pif-paf-edge401-190424-122009.pkl.log \
  outputs/resnet101block5-pif-paf-edge401-190412-151013.pkl.log \


See datasets for setup instructions. See studies.ipynb for previous studies.

The exact training command that was used for a model is in the first line of the training log file.

Train a ResNet model:

time CUDA_VISIBLE_DEVICES=0,1 python3 -m openpifpaf.train \
  --lr=1e-3 \
  --momentum=0.95 \
  --epochs=150 \
  --lr-decay 120 140 \
  --batch-size=16 \
  --basenet=resnet101 \
  --head-quad=1 \
  --headnets pif paf paf25 \
  --square-edge=401 \
  --lambdas 10 1 1 15 1 1 15 1 1

ShuffleNet models are trained without ImageNet pretraining:

time CUDA_VISIBLE_DEVICES=0,1 python3 -m openpifpaf.train \
  --batch-size=64 \
  --basenet=shufflenetv2x2 \
  --head-quad=1 \
  --epochs=150 \
  --momentum=0.9 \
  --headnets pif paf paf25 \
  --lambdas 30 2 2 50 3 3 50 3 3 \
  --loader-workers=16 \
  --lr=0.1 \
  --lr-decay 120 140 \
  --no-pretrain \
  --weight-decay=1e-5 \
  --update-batchnorm-runningstatistics \

You can refine an existing model with the --checkpoint option.

To produce evaluations at every epoch, check the directory for new snapshots every 5 minutes:

while true; do \
  CUDA_VISIBLE_DEVICES=0 find outputs/ -name "resnet101block5-pif-paf-l1-190109-113346.pkl.epoch???" -exec \
    python3 -m openpifpaf.eval_coco --checkpoint {} -n 500 --long-edge=641 --skip-existing \; \
  ; \
  sleep 300; \

Person Skeletons

COCO / kinematic tree / dense:

Created with python3 -m openpifpaf.data.


Processing a video frame by frame from video.avi to video.pose.mp4 using ffmpeg:

export VIDEO=video.avi  # change to your video file

mkdir ${VIDEO}.images
ffmpeg -i ${VIDEO} -qscale:v 2 -vf scale=641:-1 -f image2 ${VIDEO}.images/%05d.jpg
python3 -m openpifpaf.predict --checkpoint resnet152 --glob "${VIDEO}.images/*.jpg"
ffmpeg -framerate 24 -pattern_type glob -i ${VIDEO}.images/'*.jpg.skeleton.png' -vf scale=640:-2 -c:v libx264 -pix_fmt yuv420p ${VIDEO}.pose.mp4

In this process, ffmpeg scales the video to 641px which can be adjusted.

Documentation Pages

Related Projects

  • monoloco: “Monocular 3D Pedestrian Localization and Uncertainty Estimation” which uses OpenPifPaf for poses.
  • openpifpafwebdemo: web front-end.


  author = {Kreiss, Sven and Bertoni, Lorenzo and Alahi, Alexandre},
  title = {PifPaf: Composite Fields for Human Pose Estimation},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2019}
comments powered by Disqus