Triangulation: Why does it get stuck here when using 16 cameras? It runs normally when using 8 cameras.

[GentleOstrich]

Triangulation: Why does it get stuck here when using 16 cameras? It runs normally when using 8 cameras, and runs extremely slowly when using 10 cameras. How to accelerate it ? thx !!

---------------------------------------------------------------------
Camera calibration
On Wednesday 14. January 2026, 21:53:28
Calibration directory: D:\code\pose2sim_demo\Test\calibration
---------------------------------------------------------------------


--> Residual (RMS) calibration errors for each camera are respectively [] px, 
which corresponds to [] mm.

Calibration file is stored at D:\code\pose2sim_demo\Test\calibration\Calib_easymocap.toml.

Calibration took 0.02 seconds.


---------------------------------------------------------------------
Pose estimation for Test, for all frames.
On Wednesday 14. January 2026, 21:53:28
Project directory: D:\code\pose2sim_demo\Test
---------------------------------------------------------------------

Inference run only every 2 frames. Inbetween, pose estimation tracks previously detected points.

Estimating pose...
Using HALPE_26 model (body and feet) for pose estimation in balanced mode.

Valid CUDA installation found: using ONNXRuntime backend with GPU.
Skipping pose estimation as it has already been done. Set overwrite_pose to true in Config.toml if you want to run it again.

Pose estimation took 00h00m02s.


---------------------------------------------------------------------
Associating persons for Test, for all frames.
On Wednesday 14. January 2026, 21:53:30
Project directory: D:\code\pose2sim_demo\Test
---------------------------------------------------------------------


Multi-person analysis selected.
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 82/82 [00:13<00:00,  5.87it/s]

--> A person was reconstructed if the lines from cameras to their keypoints intersected within 0.1 m and if the calculated affinity stayed above 0.2.
--> Beware that people were sorted across cameras, but not across frames. This will be done in the triangulation stage.

--> A person was reconstructed if the lines from cameras to their keypoints intersected within 0.1 m and if the calculated affinity stayed above 0.2.



--> A person was reconstructed if the lines from cameras to their keypoints intersected within 0.1 m and if the calculated affinity stayed above 0.2.
--> Beware that people were sorted across cameras, but not across frames. This will be done in the triangulation stage.

Tracked json files are stored in D:\code\pose2sim_demo\Test\pose-associated.

Associating persons took 00h00m14s.


---------------------------------------------------------------------
Triangulation of 2D points for Test, for all frames.
On Wednesday 14. January 2026, 21:53:44
Project directory: D:\code\pose2sim_demo\Test
---------------------------------------------------------------------

  0%|                                                                                                                         | 0/82 [00:00<?, ?it/s]

[davidpagnon]

The single-person mode is very slow (stupid brute-force), I'm planning to remove it in the future. Instead, I'll add a parameter to let people choose how many people they want to triangulate (1, 2, more, or all), and how to sort them (on click, largest affinity, largest displacement, etc).

So for now, my advice would be for you to change multi-person to true in your Config.toml file, and rerun person association and triangulation. You might get multiple people, you will have to choose the main one (likely the first).

Please tell me how it goes!

[GentleOstrich]

The single-person mode is very slow (stupid brute-force), I'm planning to remove it in the future. Instead, I'll add a parameter to let people choose how many people they want to triangulate (1, 2, more, or all), and how to sort them (on click, largest affinity, largest displacement, etc).

So for now, my advice would be for you to change multi-person to true in your Config.toml file, and rerun person association and triangulation. You might get multiple people, you will have to choose the main one (likely the first).

Please tell me how it goes!

Thanks!
In fact, I want to reconstruct the interaction between two people and I've changed multi-person to true at begining. This is my Config.toml file:

###############################################################################
## PROJECT PARAMETERS                                                        ##
###############################################################################


# Configure your project parameters here. 
# 
# IMPORTANT:
# If a parameter is not found here, Pose2Sim will look for its value in the 
# Config.toml file of the level above. This way, you can set global 
# instructions for the Session and alter them for specific Participants or Trials.
#
# If you wish to overwrite a parameter for a specific trial or participant,  
# edit its Config.toml file by uncommenting its key (e.g., [project])
# and editing its value (e.g., frame_range = [10,300]). Or else, uncomment 
# [filtering.butterworth] and set cut_off_frequency = 10, etc.



[project]
multi_person = true # true for trials with multiple participants. If false, only the person with lowest reprojection error is analyzed.
participant_height = 'auto' # 'auto', float (eg 1.72), or list of floats (eg [1.72, 1.40])  # meters  # Only used for marker augmentation 
participant_mass = [70.0, 63.5]     # float (eg 70.0), or list of floats (eg [70.0, 63.5])  # kg      # Only used for marker augmentation and scaling, no impact on results unless you need to further compute forces

frame_rate = 'auto' # fps # int or 'auto'. If 'auto', finds from video (or defaults to 60 fps if you work with images) 
frame_range = 'auto' # 'auto', 'all', or range like [10,300]. If 'auto', will trim around the frames with low reprojection error (useful if a person enters/exits the scene)

## If cameras are not synchronized, designates the frame range of the camera with the shortest recording time
## N.B.: If you want a time range instead, use frame_range = time_range * frame_rate
## For example if you want to analyze from 0.1 to 2 seconds with a 60 fps frame rate, 
## frame_range = [0.1, 2.0]*frame_rate = [6, 120]

exclude_from_batch = [] # List of trials to be excluded from batch analysis, ['<participant_dir/trial_dir>', 'etc'].
# e.g. ['S00_P00_Participant/S00_P00_T00_StaticTrial', 'S00_P00_Participant/S00_P00_T01_BalancingTrial']


[pose]
vid_img_extension = 'mp4' # any video or image extension

pose_model = 'Body_with_feet'  #With RTMLib: 
                         # - Body_with_feet (default HALPE_26 model), 
                         # - Whole_body_wrist (COCO_133_WRIST: body + feet + 2 hand_points), 
                         # - Whole_body (COCO_133: body + feet + hands), 
                         # - Body (COCO_17). Marker augmentation won't work, Kinematic analysis will work,
                         # - Hand (HAND_21, only lightweight mode. Potentially better results with Whole_body), 
                         # - Face (FACE_106), 
                         # - Animal (ANIMAL2D_17)
                         # /!\ Only RTMPose is natively embeded in Pose2Sim. For all other pose estimation methods, you will have to run them yourself, and then refer to the documentation to convert the output files if needed
                         # /!\ For Face and Animal, use mode="""{dictionary}""", and find the corresponding .onnx model there https://github.com/open-mmlab/mmpose/tree/main/projects/rtmpose
                         
                         #With MMPose: HALPE_26, COCO_133, COCO_17, CUSTOM. See CUSTOM example at the end of the file
                         #With openpose: BODY_25B, BODY_25, BODY_135, COCO, MPII
                         #With mediapipe: BLAZEPOSE
                         #With alphapose: HALPE_26, HALPE_68, HALPE_136, COCO_133
                         #With deeplabcut: CUSTOM. See example at the end of the file

mode = 'balanced' # 'lightweight', 'balanced', 'performance', 
                  # or """{dictionary}""" (see below)
# A dictionary (WITHIN THREE DOUBLE QUOTES) allows you to manually select the person detection (if top_down approach) and/or pose estimation models (see https://github.com/Tau-J/rtmlib). 
# Models can be local paths or URLs.
# Make sure the input_sizes are within square brackets, and that they are in the opposite order from the one in the model path (for example, it would be [192,256] for rtmpose-m_simcc-body7_pt-body7-halpe26_700e-256x192-4d3e73dd_20230605.zip). 
# If your pose_model is not provided in skeletons.py, you may have to create your own one (see example at the end of the file).
# Example, equivalent to mode='balanced':
# mode = """{'det_class':'YOLOX',
#          'det_model':'https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/onnx_sdk/yolox_m_8xb8-300e_humanart-c2c7a14a.zip',
#          'det_input_size':[640, 640],
#          'pose_class':'RTMPose',
#          'pose_model':'https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/onnx_sdk/rtmpose-m_simcc-body7_pt-body7-halpe26_700e-256x192-4d3e73dd_20230605.zip',
#          'pose_input_size':[192,256]}"""
# Example with one-stage RTMO model (Requires pose_model = 'Body'):
# mode = """{'pose_class':'RTMO', 
#          'pose_model':'https://download.openmmlab.com/mmpose/v1/projects/rtmo/onnx_sdk/rtmo-m_16xb16-600e_body7-640x640-39e78cc4_20231211.zip', 
#          'pose_input_size':[640, 640]}"""
# Example with animal pose estimation:
# mode = """{'pose_class':'RTMPose',
#          'pose_model':'https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/onnx_sdk/rtmpose-m_simcc-ap10k_pt-aic-coco_210e-256x256-7a041aa1_20230206.zip',
#          'pose_input_size':[256,256]}"""


det_frequency = 2 # Run person detection only every N frames, and inbetween track previously detected bounding boxes (pose estimation is still run on all frames). 
                  # Equal to or greater than 1, can be as high as you want in simple uncrowded cases. Much faster, but might be less accurate. 
device = 'auto' # 'auto', 'CPU', 'CUDA', 'MPS', 'ROCM'
backend = 'auto' # 'auto', 'openvino', 'onnxruntime', 'opencv'

tracking_mode = 'sports2d' # 'none', 'sports2d' or 'deepsort'. 'deepsort' is slower but more robust in difficult configurations
max_distance_px = 100  # px # If a person is detected further than max_distance from its position on the previous frame, it will be considered as a new one
deepsort_params = """{'max_age':30, 'n_init':3, 'nms_max_overlap':0.8, 'max_cosine_distance':0.3, 'nn_budget':200, 'max_iou_distance':0.8}""" # """{dictionary between 3 double quotes}"""
                  # More robust in crowded scenes but Can be tricky to parametrize. More information there: https://github.com/levan92/deep_sort_realtime/blob/master/deep_sort_realtime/deepsort_tracker.py#L51
                  # Note: For faster and more robust tracking, use {'embedder_gpu': True, embedder':'torchreid'}, which uses the GPU and runs osnet_ain_x1_0 by default. requires `pip install torch torchvision torchreid gdown tensorboard`

display_detection = true
overwrite_pose = false # set to false if you don't want to recalculate pose estimation when it has already been done
save_video = 'to_video' # 'to_video' or 'to_images', 'none', or ['to_video', 'to_images']
output_format = 'openpose' # 'openpose', 'mmpose', 'deeplabcut', 'none' or a list of them # /!\ only 'openpose' is supported for now


[synchronization]
synchronization_gui = true # true or false. If true, a player will popup and let the user determine synchronization parameters. If false, it will be done automatically based on the parameters below
display_sync_plots = true # true or false (lowercase)
save_sync_plots = true # true or false (lowercase) # save synchronization plots
keypoints_to_consider = ['RWrist'] # 'all' if all points should be considered, for example if the participant did not perform any particicular sharp movement. In this case, the capture needs to be 5-10 seconds long at least
                           # ['RWrist', 'RElbow'] list of keypoint names if you want to specify keypoints with a sharp vertical motion.
approx_time_maxspeed = 'auto' # 'auto' if you want to consider the whole capture (default, slower if long sequences)
                           # [10.0, 2.0, 8.0, 11.0] list of times (seconds) if you want to specify the approximate time of a clear vertical event for each camera
time_range_around_maxspeed = 2.0 # Search for best correlation in the range [approx_time_maxspeed - time_range_around_maxspeed, approx_time_maxspeed  + time_range_around_maxspeed]
likelihood_threshold = 0.4 # Keypoints whose likelihood is below likelihood_threshold are filtered out
filter_cutoff = 6 # time series are smoothed to get coherent time-lagged correlation
filter_order = 4


# Take heart, calibration is not that complicated once you get the hang of it!
[calibration]
calibration_type = 'convert' # 'convert' or 'calculate'

   [calibration.convert]
   convert_from = 'easymocap' # 'caliscope', 'qualisys', 'optitrack', vicon', 'opencap', 'easymocap', 'biocv', 'anipose', or 'freemocap'
      [calibration.convert.caliscope]  # No parameter needed
      [calibration.convert.qualisys]
      binning_factor = 1 # Usually 1, except when filming in 540p where it usually is 2
      [calibration.convert.optitrack]  # See readme for instructions
      [calibration.convert.vicon]      # No parameter needed
      [calibration.convert.opencap]    # No parameter needed
      [calibration.convert.easymocap]  # No parameter needed
      [calibration.convert.biocv]      # No parameter needed
      [calibration.convert.anipose]    # No parameter needed
      [calibration.convert.freemocap]  # No parameter needed
  

   [calibration.calculate] 
      save_debug_images = true       # Save images with clicked points (and their reprojections for extrinsic calibration)
      # Camera properties, theoretically need to be calculated only once in a camera lifetime
      [calibration.calculate.intrinsics]
      overwrite_intrinsics = false # set to false if you don't want to recalculate intrinsic parameters
      intrinsics_extension = 'jpg' # any video or image extension
      extract_every_N_sec = 1 # if video, extract frames every N seconds (can be <1 )
      intrinsics_corners_nb = [4,7] 
      intrinsics_square_size = 60 # mm
      show_detection_intrinsics = true # true or false (lowercase)

      # Camera placements, need to be done before every session
      [calibration.calculate.extrinsics]
      calculate_extrinsics = true # true or false (lowercase) 
      extrinsics_method = 'scene' # 'board', 'scene', 'keypoints'
      extrinsics_extension = 'png' # any video or image extension
      # 'board' should be large enough to be detected when laid on the floor. Not recommended.
      # 'scene' involves manually clicking any point of know coordinates on scene. Usually more accurate if points are spread out.
      # 'keypoints' uses automatic pose estimation of a person freely walking and waving arms in the scene. Slighlty less accurate, requires synchronized cameras.
      show_reprojection_error = true # true or false (lowercase)
      moving_cameras = false # Not implemented yet

         [calibration.calculate.extrinsics.board]
         board_position = 'vertical' # 'vertical' or 'horizontal'
         extrinsics_corners_nb = [4,7] # [H,W] rather than [w,h]
         extrinsics_square_size = 60 # mm # [h,w] if square is actually a rectangle

         [calibration.calculate.extrinsics.scene]
         # list of 3D coordinates to be manually labelled on images. Can also be a 2 dimensional plane. 
         # in m -> unlike for intrinsics, NOT in mm!
         # [X,Y,Z], which becomes [Z,X,Y] in the TRC reference frame, and [-Y,X,Z] in the Blender one
         object_coords_3d =   [[-2.0,  0.3,  0.0], 
                              [-2.0 , 0.0,  0.0], 
                              [-2.0, 0.0,  0.05], 
                              [-2.0, -0.3 ,  0.0], 
                              [0.0,  0.3,  0.0], 
                              [0.0, 0.0,  0.0], 
                              [0.0, 0.0,  0.05], 
                              [0.0, -0.3,  0.0]]
        
         [calibration.calculate.extrinsics.keypoints]
         # Coming soon!


[personAssociation]
   likelihood_threshold_association = 0.3 # should be in single_person section

   [personAssociation.single_person]
   reproj_error_threshold_association = 20 # px
   tracked_keypoint = 'Neck' # If the neck is not detected by the pose_model, check skeleton.py 
               # and choose a stable point for tracking the person of interest (e.g., 'right_shoulder' or 'RShoulder')
   
   [personAssociation.multi_person]
   reconstruction_error_threshold = 0.1 # 0.1 = 10 cm
   min_affinity = 0.2 # affinity below which a correspondence is ignored. 
                      # affinity is high when reconstruction_error << reconstruction_error_threshold
   # if reconstruction_error_threshold too low or min_affinity too high: no correspondences will be found
   # if reconstruction_error_threshold too high or min_affinity too low: wrong correspondences will be found. If one person is associated to several ones on a different view, only the first is kept
                      

[triangulation]
reproj_error_threshold_triangulation = 15 # px # keypoint not triangulated if reprojection error is above 
likelihood_threshold_triangulation= 0.3 # keypoint ignored if 2D likelihood is below
min_cameras_for_triangulation = 2 # Won't triangulate a point if less than N cameras see it correctly
interp_if_gap_smaller_than = 20 # do not interpolate larger gaps
max_distance_m = 1.0 # m # if a person is detected further than max_distance from its position on the previous frame, it will be considered as a new one

interpolation = 'linear' # linear, slinear, quadratic, cubic, or none
                        # 'none' if you don't want to interpolate missing points
remove_incomplete_frames = false # true or false (lowercase) # If true, if any keypoint cannot be triangulated, the full time frame is removed
sections_to_keep = 'all' # 'all', 'largest', 'first', 'last'
                        # keep 'all' valid sections even when they are interspersed with untriangulated chunks, or the 'largest' valid section, or the 'first' one, or the 'last' one
min_chunk_size = 10 # Minimum number of valid frames in a row to keep a chunk of data for a person
fill_large_gaps_with = 'last_value' # 'last_value', 'nan', or 'zeros' 

show_interp_indices = true # true or false (lowercase). For each keypoint, return the frames that need to be interpolated
make_c3d = true # save triangulated data in c3d format in addition to trc


[filtering]
reject_outliers = true      # Hampel filter for outlier rejection before other filtering methods. Can be slow. Rejects outliers that are outside of a 95% confidence interal from the median in a sliding window of size 7.
filter = true               # Proceed to further filtering after outlier rejection.
type = 'butterworth' # butterworth, kalman, gvc_spline, gaussian, LOESS, median, butterworth_on_speed

display_figures = true # true or false (lowercase) 
save_filt_plots = true # true or false (lowercase) # save filtering plots
make_c3d = true # also save triangulated data in c3d format

   # Most intuitive and standard filter in biomechanics
   [filtering.butterworth]
   cut_off_frequency = 6 # Hz # Will be divided by slowmo_factor to be equivalent to non slowed-down video
   order = 4 

   # Used in countless applications, this one is a simplified Kalman filter 
   [filtering.kalman]
   # How much more do you trust triangulation results (measurements), than the assumption of constant acceleration(process)?
   trust_ratio = 500 # = measurement_trust/process_trust ~= process_noise/measurement_noise
   smooth = true # should be true, unless you need real-time filtering

   # Automatically determines optimal parameters for each point, which is good when some move faster than others (eg fingers vs hips).
   [filtering.gcv_spline]
   cut_off_frequency = 'auto' # 'auto' or int # If int, behaves like a Butterworth filter. 'auto' is usually better, unless the signal is too short (noise can then be considered as signal -> trajectories not filtered)
   smoothing_factor = 1.0 # >=0, ignored if cut_off_frequency != 'auto'. Biases results towards more smoothing (>1) or more fidelity to data (<1) 

   [filtering.loess]
   nb_values_used = 5 # = fraction of data used * nb frames
   [filtering.gaussian]
   sigma_kernel = 1 #px
   [filtering.median]
   kernel_size = 3
   [filtering.butterworth_on_speed]
   order = 4 
   cut_off_frequency = 10 # Hz


[markerAugmentation] 
## Requires the following markers: ["Neck", "RShoulder", "LShoulder", "RHip", "LHip", "RKnee", "LKnee",
##        "RAnkle", "LAnkle", "RHeel", "LHeel", "RSmallToe", "LSmallToe",
##        "RBigToe", "LBigToe", "RElbow", "LElbow", "RWrist", "LWrist"]
feet_on_floor = false # true or false (lowercase) # If true, markers will be translated so that the feet are on the floor (assuming that the floor is flat). Useful if you want to estimate ground reaction forces or joint loads.
make_c3d = true # save triangulated data in c3d format in addition to trc


[kinematics]
use_augmentation = true  # true or false (lowercase) # Set to true if you want to use the model with augmented markers
use_simple_model = false # true or false # >10 times faster IK if true. No muscles, no constraints (eg stiff spine and shoulders, no patella)
right_left_symmetry = true # true or false (lowercase) # Set to false only if you have good reasons to think the participant is not symmetrical (e.g. prosthetic limb)

default_height = 1.7 # meters # If automatic height calculation did not work, this value is used to scale the model
remove_individual_scaling_setup = true # true or false (lowercase) # If true, the individual scaling setup files are removed to avoid cluttering
remove_individual_ik_setup = true # true or false (lowercase) # If true, the individual IK setup files are removed to avoid cluttering
fastest_frames_to_remove_percent = 0.1 # Frames with high speed are considered as outliers
close_to_zero_speed_m = 0.2 # Sum for all keypoints: about 50 px/frame or 0.2 m/frame
large_hip_knee_angles = 45 # Hip and knee angles below this value are considered as imprecise
trimmed_extrema_percent = 0.5 # Proportion of the most extreme segment values to remove before calculating their mean)


[logging]
use_custom_logging = false # if integrated in an API that already has logging



# CUSTOM skeleton
# If you use a model with different keypoints and/or different ordering
# Useful if you trained your own model, from DeepLabCut or MMPose for example. 
# Make sure the ids are set in the right order and start from zero.
# 
# If you want to perform inverse kinematics, you will also need to create an OpenSim model
# and add to its markerset the location where you expect the triangulated keypoints to be detected.
# 
# In this example, CUSTOM reproduces the HALPE_26 skeleton (default skeletons are stored in skeletons.py).
# You can create as many custom skeletons as you want, just add them further down and rename them.
# 
# Check your model hierarchy with:  
# from anytree import Node, RenderTree
# for pre, _, node in RenderTree(model): 
#     print(f'{pre}{node.name} id={node.id}')
[[pose.CUSTOM]]
name = "Hip"
id = 19
  [[pose.CUSTOM.children]]
  name = "RHip"
  id = 12
     [[pose.CUSTOM.children.children]]
     name = "RKnee"
     id = 14
        [[pose.CUSTOM.children.children.children]]
        name = "RAnkle"
        id = 16
           [[pose.CUSTOM.children.children.children.children]]
           name = "RBigToe"
           id = 21
              [[pose.CUSTOM.children.children.children.children.children]]
              name = "RSmallToe"
              id = 23
           [[pose.CUSTOM.children.children.children.children]]
           name = "RHeel"
           id = 25
  [[pose.CUSTOM.children]]
  name = "LHip"
  id = 11
     [[pose.CUSTOM.children.children]]
     name = "LKnee"
     id = 13
        [[pose.CUSTOM.children.children.children]]
        name = "LAnkle"
        id = 15
           [[pose.CUSTOM.children.children.children.children]]
           name = "LBigToe"
           id = 20
              [[pose.CUSTOM.children.children.children.children.children]]
              name = "LSmallToe"
              id = 22
           [[pose.CUSTOM.children.children.children.children]]
           name = "LHeel"
           id = 24
  [[pose.CUSTOM.children]]
  name = "Neck"
  id = 18
     [[pose.CUSTOM.children.children]]
     name = "Head"
     id = 17
        [[pose.CUSTOM.children.children.children]]
        name = "Nose"
        id = 0
     [[pose.CUSTOM.children.children]]
     name = "RShoulder"
     id = 6
        [[pose.CUSTOM.children.children.children]]
        name = "RElbow"
        id = 8
           [[pose.CUSTOM.children.children.children.children]]
           name = "RWrist"
           id = 10
     [[pose.CUSTOM.children.children]]
     name = "LShoulder"
     id = 5
        [[pose.CUSTOM.children.children.children]]
        name = "LElbow"
        id = 7
           [[pose.CUSTOM.children.children.children.children]]
           name = "LWrist"
           id = 9

So it is very slow when doing multi-person task. And this is my python code to run the reconstruction process:

import os
os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"
from Pose2Sim import Pose2Sim
Pose2Sim.calibration()
Pose2Sim.poseEstimation()
# Pose2Sim.synchronization()
Pose2Sim.personAssociation()
Pose2Sim.triangulation()
Pose2Sim.filtering()
Pose2Sim.markerAugmentation()
Pose2Sim.kinematics()

[davidpagnon]

Hm okay, so it is very slow even in multi-person mode... There are probably ways to further optimize it but I won't be able to investigate for now.

The truth is that 16 cameras is quite a lot, are your results worse when using fewer?

[GentleOstrich]

Hm okay, so it is very slow even in multi-purpose mode... There are probably ways to further optimize it but I won't be able to investigate for now.

The truth is that 16 cameras is quite a lot, are your results worse when using fewer?

The result is not much better when using 10 cameras than 8 cameras, the people's hand's positions are wrong when i want to reconstruct the interaction of handovering something.

[davidpagnon]

Oh okay, I think we found the problem! Do the points from the trc file look okay?

If so, this is not a problem in triangulation, but in the OpenSim model, which does not have mobile hands. 2 options:

• either you compute the finger angles and positions from the trc file
• or you add a hand model to the current OpenSim model. It sounds intimidating but I believe an LLM (ChatGPT or anything else) could help you

Whatever you do, could you keep me posted? It would be useful for the next users and I'd like to improve the software for everyone.
If you need help from other users, feel free to come to the Discord server: https://discord.com/invite/4mXUdSFjmt

[GentleOstrich]

Oh okay, I think we found the problem! Do the points from the trc file look okay?

If so, this is not a problem in triangulation, but in the OpenSim model, which does not have mobile hands. 2 options:

• either you compute the finger angles and positions from the trc file
• or you add a hand model to the current OpenSim model. It sounds intimidating but I believe an LLM (ChatGPT or anything else) could help you

Whatever you do, could you keep me posted? It would be useful for the next users and I'd like to improve the software for everyone. If you need help from other users, feel free to come to the Discord server: https://discord.com/invite/4mXUdSFjmt

OK, I'll try it. Thank you

[GentleOstrich]

I used pose_model = 'Whole_body' in Config.toml and 5 cameras, getting the result like this:

I think it is much better than before.

Besides, is there any method to convert an opensim file to smplx formart?

Tanks ~

[davidpagnon]

Okay, great!

I've wanted to animate a mesh based on the opensim skeleton or on the trc coordinates for a while, and there has been some work done on the Discord server. I'm pretty sure @carlosedubarreto has managed it so you could ask him, but I believe he is quite busy at the moment :)

[carlosedubarreto]

Hello @GentleOstrich ,
Last year I did some blender scripts to load the opensim data in blender.
Actually I did a post on my site, take a look here https://www.patreon.com/posts/endorfina-125755950
There is also a video explaining how to use it. Hope it helps