Keypoint Communities
In this repository you will find the code to our ICCV '21 paper:
Keypoint Communities
Duncan Zauss, Sven Kreiss, Alexandre Alahi, 2021.We present a fast bottom-up method that jointly detects over 100 keypoints on humans or objects, also referred to as human/object pose estimation. We model all keypoints belonging to a human or an object -the pose- as a graph and leverage insights from community detection to quantify the independence of keypoints. We use a graph centrality measure to assign training weights to different parts of a pose. Our proposed measure quantifies how tightly a keypoint is connected to its neighborhood. Our experiments show that our method outperforms all previous methods for human pose estimation with fine-grained keypoint annotations on the face, the hands and the feet with a total of 133 keypoints. We also show that our method generalizes to car poses.
Qualitative results
Image credit: Photo by Toby Bradbury which is licensed under CC-BY-2.0.
Webcam demo. You can try it out yourself with the following command:
python -m openpifpaf.video --checkpoint=shufflenetv2k16-wholebody --show --long-edge=320
Image credit: Photo by Lokomotive74 which is licensed under CC-BY-4.0.
Created with:
python -m openpifpaf.predict docs/soccer.jpeg --checkpoint=shufflenetv2k30-wholebody --line-width=2 --show
Image credit: "Learning to surf" by fotologic which is licensed under [CC-BY-2.0].
Created with:
python3 -m openpifpaf.predict docs/000000081988.jpg --checkpoint=shufflenetv2k30-wholebody --line-width=2 --show
Installation
This project is based on OpenPifPaf. Create a virtual environment with python 3.7, 3.8 or 3.9, clone this repo and then install the required packages:
git clone https://github.com/DuncanZauss/Keypoint_Communities.git
cd Keypoint_Communities
pip install -r requirements.txt
Obtain keypoint weights
To compute the keypoint weights with our method, download the preprocessed annotations of the MS COCO WholeBody dataset and/or the ApolloCar3D dataset with the following commands:
cd Keypoint_Communities/src
wget https://github.com/DuncanZauss/Keypoint_Communities/releases/download/v0.1.0/person_keypoints_train2017_wholebody_pifpaf_style.json
wget https://github.com/DuncanZauss/Keypoint_Communities/releases/download/v0.1.0/apollo_keypoints_66_train.json
To compute the average euclidean distance in the datasets for every edge run:
python Compute_edge_weights.py
To compute training weights with centrality measures as proposed in our paper run the following command:
python Compute_training_weights.py
You will find the computed weights in the respective csv file and a visualization of the computed weights in the respective docs folder.
Training
For training you will need to download the MS COCO dataset and the WholeBody keypoint annotations as explained here. To train an OpenPifPaf model with our keypoint weighting scheme, you can use the following command:
python -m openpifpaf.train --dataset=wholebody --lr=0.0001 --momentum=0.95 --b-scale=10.0 --clip-grad-value=10 --epochs=350 --lr-decay 330 340 --lr-decay-epochs=10 --lr-warm-up-start-epoch=250 --batch-size=16 --weight-decay=1e-5 --wholebody-upsample=2 --wholebody-extended-scale --wholebody-orientation-invariant=0.1 --checkpoint=shufflenetv2k30 --head-consolidation=create --wholebody-val-annotations=
/person_keypoints_val2017_wholebody_pifpaf_style.json --wholebody-train-annotations=
/person_keypoints_train2017_wholebody_pifpaf_style.json --wholebody-apply-local-centrality-weights
Evaluation
To evaluate a trained model you first need to download the annotation file from this link and than you can use the following command to evaluate a model:
python -m openpifpaf.eval --dataset=wholebody --checkpoint=shufflenetv2k30-wholebody --force-complete-pose --seed-threshold=0.2 --force-complete-caf-th=0.001 --wholebody-val-annotations=
/coco_wholebody_val_v1.0.json
The shufflenetv2k30-wholebody is our pretrained model, which was trained with the command from the Training section and will automatically be downloaded via torchhub. If you wish to evaluate your own model you can replace it with a local path to your model.
Related projects
- AK391 created a great webdemo in Huggingface Spaces with Gradio. See demo:
Citation
If you find our research useful please cite us:
@misc{zauss2021keypoint,
title={Keypoint Communities},
author={Duncan Zauss and Sven Kreiss and Alexandre Alahi},
year={2021},
eprint={2110.00988},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
License
The code in this repository is licensed under the MIT license. For more information please refer to the LICENSE file. This project is largely based on OpenPifPaf. OpenPifPaf is licensed under the GNU AGPLv3 license, for more information please refer to OpenPifPaf's license.
I use the same command 'python -m openpifpaf.predict docs/soccer.jpeg --checkpoint=shufflenetv2k30-wholebody --line-width=2 --show
' to run the prediction but the hands of the front person are not aligned as your shown image.
Did you use a different weight? Or is there any other problems?
550 Dec 21, 2022
1 Mar 12, 2022
563 Dec 31, 2022
1 Oct 20, 2021
1 Sep 23, 2022
5 Nov 26, 2022
33 Dec 30, 2022
1k Jan 01, 2023
108 Jan 05, 2023
8 Nov 02, 2022
20 May 17, 2022
77 Dec 08, 2022
6 May 24, 2021
2.2k Jan 08, 2023
2.9k Jan 08, 2023
25 Dec 26, 2022
64 Dec 17, 2022
61 Jan 06, 2023
45 Dec 08, 2022
922 Jan 01, 2023
