MediaPipe multi pose detection example. The graph has been adapted from the HandLandmarkTrackingCpu example by mediapipe. This is a proof of concept and is provided for educational purposes only.

Source: Allan Mas

Currently, this is only tested on Windows and macOS. It's recommended to use Python3 (>=3.7) and a virtual environment.

python install -r requirements.txt

To run an example use the basic python command to start up the script.

# start pose detection with webcam
python pose.py

# pose detection with single image
python pose.py --image images/pexels-allan-mas-5368956.jpg

There is also a pre-compiled pose detection graph which includes the detection to ROI conversion. It is possible to run the example with the following command:

python detection.py

To build the graphs, download the mediapipe repository (>=0.8.7) and set the path to this repository as well as to the mediapipe repo in the build_custom_graphs.sh. After that run the build_custom_graphs.sh script and let it compile. If you need help to setup your development environment, have a look at Building MediaPipe Python Package.

The roi, which is detected either by the pose detection or the landmark to pose roi node, is quite large for a person. The overlap of different pose rois leads to difficulties in tracking. This is because the node was made for single pose detection. It is possible to increase the min-similarity-threshold parameter to let multiple overlapping pose not be count as one. If another pose detection would be applied, which is way more close to the landmarks detected, this could already solve the problem.

The solution_base file has been copied into mpx/solution_base.py to adapt the path, where the resources are loaded and set the _input_stream_type_info manually due to a not registered type error which could not be resolved for now.

At the moment the landmark filter is not implemented into the graph because of difficulties to map filter to pose.

MIT License - Copyright (c) 2022 Florian Bruggisser

Based on mediapipe-osc.