`Pose2Sim` provides a workflow for 3D markerless kinematics, as an alternative to the more usual marker-based motion capture methods. It aims to provide a free tool to obtain research-grade results from consumer-grade equipment. Any combination of phone, webcam, GoPro, etc. can be used.
Pose2Sim stands for "OpenPose to OpenSim", as it uses OpenPose inputs (2D keypoints coordinates obtained from multiple videos) and leads to an OpenSim result (full-body 3D joint angles). Other 2D pose estimators such as BlazePose (MediaPipe), DeepLabCut, AlphaPose, can now be used as inputs.
*Tested up to v4.4-beta on Windows. Has to be compiled from source on Linux (see [there](https://simtk-confluence.stanford.edu:8443/display/OpenSim/Linux+Support)).*
3.***Optional.*** *Install Anaconda or [Miniconda](https://docs.conda.io/en/latest/miniconda.html). \
Open an Anaconda terminal and create a virtual environment with typing:*
*N.B.:* Default parameters have been provided in [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml) but can be edited.\
*N.B.:* *Try the calibration tool by changing `calibration_type` to `calculate` instead of `convert` (more info [there](#calculate-from-scratch)).*
**N.B.:** You need to proceed to the full install to import the inverse kinematic results from OpenSim. See instructions [there](https://github.com/davidpagnon/Pose2Sim_Blender?tab=readme-ov-file#full-install).
> _**Another person, hidden all along, will appear when multi-person analysis is activated!**_
Go to the Multi-participant Demo folder: `cd <path>\pose2sim\S00_Demo_Session\S00_P01_MultiParticipants`. \
Type `ipython`, and try the following code:
``` python
from Pose2Sim import Pose2Sim
Pose2Sim.personAssociation()
Pose2Sim.triangulation()
Pose2Sim.filtering()
Pose2Sim.markerAugmentation()
```
One .trc file per participant will be generated and stored in the `pose-3d` directory.\
You can then run OpenSim scaling and inverse kinematics for each resulting .trc file as in [Demonstration Part-2](#demonstration-part-2-obtain-3d-joint-angles-with-opensim).\
You can also visualize your results with Blender as in [Demonstration Part-3](#demonstration-part-3-optional-visualize-your-results-with-blender).
*N.B.:* Set *[project]*`multi_person = true` for each trial that contains multiple persons.\
Set *[triangulation]*`reorder_trc = true` if you need to run OpenSim and to match the generated .trc files with the static trials.\
Make sure that the order of *[markerAugmentation]*`participant_height` and `participant_mass` matches the order of the static trials.
*N.B.:* Note that in the case of our floating ghost participant, marker augmentation may worsen the results. See [Marker augmentation](#marker-augmentation) for instruction on when and when not to use it.
2. Copy the `S00_Demo_Session` folder wherever you like, and rename it as you wish.
3. The rest of the tutorial will explain to you how to populate the `Calibration` and `videos` folders, edit the [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml) files, and run each Pose2Sim step.
Run Pose2Sim from the `Session` folder if you want to batch process the whole session, from the `Participant` folder if you want to batch process all the trials of a participant, or from the `Trial` folder if you want to process a single trial. There should be one `Calibration` folder per session.
Global instructions are given in the Config.toml file of the `Session` folder, and can be altered for specific `Participants` or `Trials` by uncommenting keys and their values in their respective Config.toml files.\
Try uncommenting `[project]` and set `frame_range = [10,300]` for a Participant for example, or uncomment `[filtering.butterworth]` and set `cut_off_frequency = 10` for a Trial.
> _**Estimate 2D pose from images with Openpose or another pose estimation solution.**_ \
N.B.: First film a short static pose that will be used for scaling the OpenSim model (A-pose for example), and then film your motions of interest.\
N.B.: Note that the names of your camera folders must follow the same order as in the calibration file, and end with '_json'.
### With OpenPose:
The accuracy and robustness of Pose2Sim have been thoroughly assessed only with OpenPose, and especially with the BODY_25B model. Consequently, we recommend using this 2D pose estimation solution. See [OpenPose repository](https://github.com/CMU-Perceptual-Computing-Lab/openpose) for installation and running.
* Open a command prompt in your **OpenPose** directory. \
* The [BODY_25B model](https://github.com/CMU-Perceptual-Computing-Lab/openpose_train/tree/master/experimental_models) has more accurate results than the standard BODY_25 one and has been extensively tested for Pose2Sim. \
You can also use the [BODY_135 model](https://github.com/CMU-Perceptual-Computing-Lab/openpose_train/tree/master/experimental_models), which allows for the evaluation of pronation/supination, wrist flexion, and wrist deviation.\
All other OpenPose models (BODY_25, COCO, MPII) are also supported.\
* Use one of the `json_display_with_img.py` or `json_display_with_img.py` scripts (see [Utilities](#utilities)) if you want to display 2D pose detections.
**N.B.:** *OpenPose BODY_25B is the default 2D pose estimation model used in Pose2Sim. However, other skeleton models from other 2D pose estimation solutions can be used alternatively.*
[Mediapipe BlazePose](https://google.github.io/mediapipe/solutions/pose.html) is very fast, fully runs under Python, handles upside-down postures and wrist movements (but no subtalar ankle angles). \
However, it is less robust and accurate than OpenPose, and can only detect a single person.
* Use the script `Blazepose_runsave.py` (see [Utilities](#utilities)) to run BlazePose under Python, and store the detected coordinates in OpenPose (json) or DeepLabCut (h5 or csv) format:
``` cmd
python -m Blazepose_runsave -i input_file -dJs
```
Type in `python -m Blazepose_runsave -h` for explanation on parameters.
* Make sure you changed the `pose_model` and the `tracked_keypoint` in the [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml) file.
If you need to detect specific points on a human being, an animal, or an object, you can also train your own model with [DeepLabCut](https://github.com/DeepLabCut/DeepLabCut). In this case, Pose2Sim is used as an alternative to [AniPose](https://github.com/lambdaloop/anipose), but it may yield better results since 3D reconstruction takes confidence into account (see [this article](https://doi.org/10.1080/21681163.2023.2292067)).
1. Train your DeepLabCut model and run it on your images or videos (more instruction on their repository)
2. Translate the h5 2D coordinates to json files (with `DLC_to_OpenPose.py` script, see [Utilities](#utilities)):
3. Edit `pose.CUSTOM` in [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml), and edit the node IDs so that they correspond to the column numbers of the 2D pose file, starting from zero. Make sure you also changed the `pose_model` and the `tracked_keypoint`.\
You can visualize your skeleton's hierarchy by changing pose_model to CUSTOM and writing these lines:
``` python
config_path = r'path_to_Config.toml'
import toml, anytree
config = toml.load(config_path)
pose_model = config.get('pose').get('pose_model')
model = anytree.importer.DictImporter().import_(config.get('pose').get(pose_model))
for pre, _, node in anytree.RenderTree(model):
print(f'{pre}{node.name} id={node.id}')
```
4. Create an OpenSim model if you need inverse kinematics.
### With AlphaPose:
[AlphaPose](https://github.com/MVIG-SJTU/AlphaPose) is one of the main competitors of OpenPose, and its accuracy is comparable. As a top-down approach (unlike OpenPose which is bottom-up), it is faster on single-person detection, but slower on multi-person detection.\
All AlphaPose models are supported (HALPE_26, HALPE_68, HALPE_136, COCO_133, COCO, MPII). For COCO and MPII, AlphaPose must be run with the flag "--format cmu".
* Install and run AlphaPose on your videos (more instruction on their repository)
* Translate the AlphaPose single json file to OpenPose frame-by-frame files (with `AlphaPose_to_OpenPose.py` script, see [Utilities](#utilities)):
* Make sure you changed the `pose_model` and the `tracked_keypoint` in the [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml) file.
> _**Cameras need to be synchronized, so that 2D points correspond to the same position across cameras.**_\
***N.B.:** Skip this step if your cameras are already synchronized.*
If your cameras are not natively synchronized, you can use [this script](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/Utilities/synchronize_cams_draft.py). This is still a draft, and will be updated in the future.\
Alternatively, use a flashlight or a clap to synchronize them. GoPro cameras can also be synchronized with a timecode, by GPS (outdoors) or with a remote control (slightly less reliable).
If you already have a calibration file, set `calibration_type` type to `convert` in your [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/Empty_project/User/Config.toml) file.
- set `convert_from` to 'qualisys' in your [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml) file. Change `binning_factor` to 2 if you film in 540p.
- Calculate intrinsics with a board (see next section).
- Use their C++ API [to retrieve extrinsic properties](https://docs.optitrack.com/developer-tools/motive-api/motive-api-function-reference#tt_cameraxlocation). Translation can be copied as is in your `Calib.toml` file, but TT_CameraOrientationMatrix first needs to be [converted to a Rodrigues vector](https://docs.opencv.org/3.4/d9/d0c/group__calib3d.html#ga61585db663d9da06b68e70cfbf6a1eac) with OpenCV. See instructions [here](https://github.com/perfanalytics/pose2sim/issues/28).
- set `convert_from` to 'vicon' in your [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml) file. No other setting is needed.
- set `convert_from` to 'opencap' in your [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml) file. No other setting is needed.
- set `convert_from` to 'easymocap' in your [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml) file. No other setting is needed.
- set `convert_from` to 'biocv' in your [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml) file. No other setting is needed.
> *N.B.:* Try the calibration tool on the Demo by changing `calibration_type` to `calculate` in [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml).\
For the sake of practicality, there are voluntarily few board images for intrinsic calibration, and few points to click for extrinsic calibration. In spite of this, your reprojection error should be under 1-2 cm, which [does not hinder the quality of kinematic results in practice](https://www.mdpi.com/1424-8220/21/19/6530/htm#:~:text=Angle%20results%20were,Table%203).).
> *N.B.:* _Intrinsic parameters:_ camera properties (focal length, optical center, distortion), usually need to be calculated only once in their lifetime. In theory, cameras with same model and same settings will have identical intrinsic parameters.\
> *N.B.:* _Extrinsic parameters:_ camera placement in space (position and orientation), need to be calculated every time a camera is moved. Can be calculated from a board, or from points in the scene with known coordinates.\
> *N.B.:* If there is no measurable item in the scene, you can temporarily bring something in (a table, for example), perform calibration, and then remove it before you start capturing motion.
Manually measure the 3D coordinates of 10 or more points in the scene (tiles, lines on wall, boxes, treadmill dimensions...). These points should be as spread out as possible. Replace `object_coords_3d` by these coordinates in Config.toml.\
Check printed output. If results are not satisfying, try and release the constraints in the [Config.toml](https://github.com/perfanalytics/pose2sim/blob/main/Pose2Sim/S00_Demo_Session/Config.toml) file.
If the reprojection error is above a threshold, right and left sides are swapped; if it is still above, cameras are removed until the threshold is met. If more cameras are removed than threshold, triangulation is skipped for this point and this frame. In the end, missing values are interpolated.
Check your filtration with the displayed figures, and visualize your .trc file in OpenSim. If your filtering is not satisfying, try and change the parameters in the `Config.toml` file.
- Open an Anaconda terminal in your OpenSim/bin directory, typically `C:\OpenSim <Version>\bin`.\
You'll need to adjust the `time_range`, `output_motion_file`, and enter the full paths to the input and output `.osim`, `.trc`, and `.mot` files in your setup file.
- Or take advantage of the full the OpenSim Python API. See [there](https://simtk-confluence.stanford.edu:8443/display/OpenSim/Scripting+in+Python) for installation instructions (conda install may take a while).\
If you run into a DLL error while importing opensim, open the file `<Pose2Sim-env>\Lib\opensim\__init__.py` and replace `conda`by `conda-meta` line 4. `<Pose2Sim-env>` location can be found with `conda env list`.\
You can also run other API commands. See [there](https://simtk-confluence.stanford.edu:8443/display/OpenSim/Common+Scripting+Commands#CommonScriptingCommands-UsingtheTools) for more instructions on how to use it.
A list of standalone tools (see [Utilities](https://github.com/perfanalytics/pose2sim/tree/main/Pose2Sim/Utilities)), which can be either run as scripts, or imported as functions. Check usage in the docstrings of each Python file. The figure below shows how some of these toolscan be used to further extend Pose2Sim usage.
Converts a Qualisys .qca.txt calibration file to the Pose2Sim .toml calibration file (similar to what is used in [AniPose](https://anipose.readthedocs.io/en/latest/)).
Converts 3D point data of a .c3d file to a .trc file compatible with OpenSim. No analog data (force plates, emg) nor computed data (angles, powers, etc.) are retrieved.
Detects gait events from point coordinates according to [Zeni et al. (2008)](https://www.sciencedirect.com/science/article/abs/pii/S0966636207001804?via%3Dihub).
If you use this code or data, please cite [Pagnon et al., 2022b](https://doi.org/10.21105/joss.04362), [Pagnon et al., 2022a](https://www.mdpi.com/1424-8220/22/7/2712), or [Pagnon et al., 2021](https://www.mdpi.com/1424-8220/21/19/6530).
@Article{Pagnon_2022_JOSS,
AUTHOR = {Pagnon, David and Domalain, Mathieu and Reveret, Lionel},
TITLE = {Pose2Sim: An open-source Python package for multiview markerless kinematics},
✔**Pose:** Support OpenPose [body_25b](https://github.com/CMU-Perceptual-Computing-Lab/openpose_train/tree/master/experimental_models#body_25b-model---option-2-recommended) for more accuracy, [body_135](https://github.com/CMU-Perceptual-Computing-Lab/openpose_train/tree/master/experimental_models#single-network-whole-body-pose-estimation-model) for pronation/supination.
✔**Pose:** Support [BlazePose](https://developers.google.com/mediapipe/solutions/vision/pose_landmarker) for faster inference (on mobile device).
✔**Pose:** Support [DeepLabCut](http://www.mackenziemathislab.org/deeplabcut) for training on custom datasets.
✔**Pose:** Support [AlphaPose](https://github.com/MVIG-SJTU/AlphaPose) as an alternative to OpenPose.
▢**Pose:** Directly reading from DeepLabCut .csv or .h5 files instead of converting to .json (triangulation, person association, calibration, synchronization...)
▢**Pose:** GUI help for DeepLabCut model creation.
✔**Calibration:** Easier and clearer calibration procedure: separate intrinsic and extrinsic parameter calculation, edit corner detection if some are wrongly detected (or not visible).
✔**Calibration:** Possibility to evaluate extrinsic parameters from cues on scene.
▢**Calibration:** Once object points have been detected or clicked once, track them for live calibration of moving cameras. Propose to click again when they are lost.
▢**Calibration:** Calibrate cameras by pairs and compute average extrinsic calibration with [aniposelib](https://github.com/lambdaloop/aniposelib/blob/d03b485c4e178d7cff076e9fe1ac36837db49158/aniposelib/utils.py#L167).
▢**Calibration:** Fine-tune calibration with bundle adjustment.
▢**Calibration:** Support ChArUco board detection (see [there](https://mecaruco2.readthedocs.io/en/latest/notebooks_rst/Aruco/sandbox/ludovic/aruco_calibration_rotation.html)).
▢**Calibration:** Calculate calibration with points rather than board. (1) SBA calibration with wand (cf [Argus](https://argus.web.unc.edu), see converter [here](https://github.com/backyardbiomech/DLCconverterDLT/blob/master/DLTcameraPosition.py)). Set world reference frame in the end.
▢**Calibration:** Alternatively, self-calibrate with [OpenPose keypoints](https://ietresearch.onlinelibrary.wiley.com/doi/full/10.1049/cvi2.12130). Set world reference frame in the end.
✔**Person Association:** Multiple persons association. 1. Triangulate all the persons whose reprojection error is below a certain threshold (instead of only the one with minimum error), and then track in time with speed cf [Slembrouck 2020](https://link.springer.com/chapter/10.1007/978-3-030-40605-9_15)? or 2. Based on affinity matrices [Dong 2021](https://arxiv.org/pdf/1901.04111.pdf)? or 3. Based on occupancy maps [Yildiz 2012](https://link.springer.com/chapter/10.1007/978-3-642-35749-7_10)? or 4. With a neural network [Huang 2023](https://arxiv.org/pdf/2304.09471.pdf)?
✔**Triangulation:** Set a likelihood threshold below which a camera should not be used, a reprojection error threshold, and a minimum number of remaining cameras below which triangulation is skipped for this frame.
✔**Triangulation:** Solve limb swapping (although not really an issue with Body_25b). Try triangulating with opposite side if reprojection error too large. Alternatively, ignore right and left sides, use RANSAC or SDS triangulation, and then choose right or left by majority voting. More confidence can be given to cameras whose plane is the most coplanar to the right/left line.
✔**Triangulation:** [Undistort](https://docs.opencv.org/3.4/da/d54/group__imgproc__transform.html#ga887960ea1bde84784e7f1710a922b93c) 2D points before triangulating (and [distort](https://github.com/lambdaloop/aniposelib/blob/d03b485c4e178d7cff076e9fe1ac36837db49158/aniposelib/cameras.py#L301) them before computing reprojection error).
✔**Triangulation:** Offer the possibility to augment the triangulated data with [the OpenCap LSTM](https://github.com/stanfordnmbl/opencap-core/blob/main/utilsAugmenter.py). Create "BODY_25_AUGMENTED" model, Scaling_setup, IK_Setup.
✔**Triangulation:** Multiple person kinematics (output multiple .trc coordinates files). Triangulate all persons with reprojection error above threshold, and identify them by minimizing their displacement across frames.
▢**Triangulation:** Pre-compile weighted_triangulation and reprojection with @jit(nopython=True, parallel=True) for faster execution.
▢**Triangulation:** Offer the possibility of triangulating with Sparse Bundle Adjustment (SBA), Extended Kalman Filter (EKF), Full Trajectory Estimation (FTE) (see [AcinoSet](https://github.com/African-Robotics-Unit/AcinoSet)).
▢**Triangulation:** Implement normalized DLT and RANSAC triangulation, Outlier rejection (sliding z-score?), as well as a [triangulation refinement step](https://doi.org/10.1109/TMM.2022.3171102).
✔**OpenSim:** Integrate better spine from [lifting fullbody model](https://pubmed.ncbi.nlm.nih.gov/30714401) to the [gait full-body model](https://nmbl.stanford.edu/wp-content/uploads/07505900.pdf), more accurate for the knee.
✔**OpenSim:** Optimize model marker positions as compared to ground-truth marker-based positions.
✔**OpenSim:** Add scaling and inverse kinematics setup files.
✔**OpenSim:** Add full model with contact spheres ([SmoothSphereHalfSpaceForce](https://simtk.org/api_docs/opensim/api_docs/classOpenSim_1_1SmoothSphereHalfSpaceForce.html#details)) and full-body muscles ([DeGrooteFregly2016Muscle](https://simtk.org/api_docs/opensim/api_docs/classOpenSim_1_1DeGrooteFregly2016Muscle.html#details)), for [Moco](https://opensim-org.github.io/opensim-moco-site/) for example.
✔**OpenSim:** Add model with [ISB shoulder](https://github.com/stanfordnmbl/opencap-core/blob/main/opensimPipeline/Models/LaiUhlrich2022_shoulder.osim).
✔**GUI:** Blender add-on (cf [MPP2SOS](https://blendermarket.com/products/mocap-mpp2soss)), or webapp (e.g., with [Napari](https://napari.org/stable). See my draft project [Maya-Mocap](https://github.com/davidpagnon/Maya-Mocap) and [BlendOsim](https://github.com/JonathanCamargo/BlendOsim).
▢**Doc:** Use [Sphinx](https://www.sphinx-doc.org/en/master), [MkDocs](https://www.mkdocs.org), or (maybe better), [github.io](https://docs.github.com/fr/pages/quickstart) for clearer documentation.
✔**Bug:** calibration.py. FFMPEG error message when calibration files are images. See [there](https://github.com/perfanalytics/pose2sim/issues/33#:~:text=In%20order%20to%20check,filter%20this%20message%20yet.).
▢**Bug:** common.py, class plotWindow(). Python crashes after a few runs of `Pose2Sim.filtering()` when `display_figures=true`. See [there](https://github.com/superjax/plotWindow/issues/7).
- Supervised my PhD: [@lreveret](https://github.com/lreveret) (INRIA, Université Grenoble Alpes), and [@mdomalai](https://github.com/mdomalai) (Université de Poitiers).
- Provided the Demo data: [@aaiaueil](https://github.com/aaiaueil) from Université Gustave Eiffel.
- Tested the code and provided feedback: [@simonozan](https://github.com/simonozan), [@daeyongyang](https://github.com/daeyongyang), [@ANaaim](https://github.com/ANaaim), [@rlagnsals](https://github.com/rlagnsals)
