The code has been migrated directly into the main OpenPifPaf repository. The documentation hasn't been moved yet from this repo to the guide in OpenPifPaf so I am keeping this repository for now.

This is the tracking plugin for OpenPifPaf.
New 2021 paper:

OpenPifPaf: Composite Fields for Semantic Keypoint Detection and Spatio-Temporal Association
Sven Kreiss, Lorenzo Bertoni, Alexandre Alahi, 2021.

Many image-based perception tasks can be formulated as detecting, associating and tracking semantic keypoints, e.g., human body pose estimation and tracking. In this work, we present a general framework that jointly detects and forms spatio-temporal keypoint associations in a single stage, making this the first real-time pose detection and tracking algorithm. We present a generic neural network architecture that uses Composite Fields to detect and construct a spatio-temporal pose which is a single, connected graph whose nodes are the semantic keypoints (e.g., a person's body joints) in multiple frames. For the temporal associations, we introduce the Temporal Composite Association Field (TCAF) which requires an extended network architecture and training method beyond previous Composite Fields. Our experiments show competitive accuracy while being an order of magnitude faster on multiple publicly available datasets such as COCO, CrowdPose and the PoseTrack 2017 and 2018 datasets. We also show that our method generalizes to any class of semantic keypoints such as car and animal parts to provide a holistic perception framework that is well suited for urban mobility such as self-driving cars and delivery robots.

Previous CVPR 2019 paper.

Install

pip install 'openpifpaf_posetrack[test,train]'

# from source:
pip install --editable '.[test,train]'

Prediction

The standard openpifpaf.video still works exactly the same way. With this plugin installed, you can use --checkpoint=tshufflenetv2k16 (with a t at the beginning). This model can be decoded in multiple ways and you should pick one decoder. To get started, we recommend --decoder=trackingpose:0. Putting it all together, an example command to process a video stream from a camera is:

MPLBACKEND=macosx python3 -m openpifpaf.video --show --long-edge=321 --checkpoint=tshufflenetv2k16 --decoder=trackingpose:0 --source 0 --horizontal-flip

Posetrack Dataset

Data. Follow the Posetrack instructions to download and untar the images. Labels:

mkdir data-posetrack
cd data-posetrack
wget https://posetrack.net/posetrack18-data/posetrack18_v0.45_public_labels.tar.gz
tar -xvf posetrack18_v0.45_public_labels.tar.gz
mv posetrack_data/* .
rm -r posetrack_data

Generate PoseTrack2017 json data of the ground truth. Usage of octave instead of matlab is not documented, but this seems to work:

cd matlab
octave --no-gui --eval "addpath('./external/jsonlab'); mat2json('your_relative_path/data-posetrack2017/annotations/val/'); quit"

This takes a long time. It is faster on the test set:

octave --no-gui --eval "addpath('./external/jsonlab'); mat2json('your_relative_path/data-posetrack2017/annotations/test/'); quit"

The Posetrack poses look like these:

Created with python -m openpifpaf_posetrack.draw_poses.

Train posetrack2018-cocokpst

# 210226

# first convert from single image to tracking model
python3 -m openpifpaf_posetrack.imagetotracking --checkpoint shufflenetv2k30

# train
time python3 -m torch.distributed.launch --nproc_per_node=4 \
  -m openpifpaf.train --ddp \
  --lr=0.0003 --momentum=0.95 --b-scale=10.0 \
  --epochs=50 --lr-decay 40 45 --lr-decay-epochs=5 \
  --batch-size=8 \
  --weight-decay=1e-5 \
  --dataset=posetrack2018-cocokpst --dataset-weights 1 1 --stride-apply=2 \
  --posetrack-upsample=2 \
  --cocokp-upsample=2 --cocokp-orientation-invariant=0.1 --cocokp-blur=0.1 \
  --checkpoint outputs/tshufflenetv2k30-210217-075056-cocokp-o10s-6f9daa84.pkl

CUDA_VISIBLE_DEVICES=3 python -m openpifpaf.eval \
  --watch --checkpoint "outputs/tshufflenetv2k??-20121?-*-posetrack2018-*.pkl.epoch??[0,5]" \
  --dataset=posetrack2018 \
  --loader-workers=8 \
  --decoder=trackingpose:0 \
  --write-predictions

The training script supports --train-annotations and --val-annotations to restrict the used annotation files. This is useful for local testing.

To produce submissions to the 2018 test server:

CUDA_VISIBLE_DEVICES=0 python -m openpifpaf.eval \
  --checkpoint outputs/tshufflenetv2k30-210222-112623-posetrack2018-cocokpst-o10-123ec670.pkl \
  --dataset=posetrack2018 --posetrack2018-eval-annotations="data-posetrack2018/annotations/test/*.json" \
  --loader-workers=8 \
  --decoder=trackingpose:0 \
  --write-predictions

For the 2017 test server:

CUDA_VISIBLE_DEVICES=1 python -m openpifpaf.eval \
  --checkpoint outputs/tshufflenetv2k30-210222-112623-posetrack2018-cocokpst-o10-123ec670.pkl \
  --dataset=posetrack2017 --posetrack2017-eval-annotations="data-posetrack2017/annotations/test/*.json" \
  --loader-workers=8 \
  --decoder=trackingpose:0 \
  --write-predictions

Citation