Smarter 'auto" number of workers for pose estimation. Allows for Ctrl+C interruption. Implemented parallelization of kinematics.

Author: davidpagnon

Pose2Sim/Demo_Batch/Config.toml

@@ -79,8 +79,8 @@ det_frequency = 4 # Run person detection only every N frames, and inbetween trac
                   # 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'
-parallel_pose = 'auto' # 'auto' or integer (number of threads). 'auto': one thread per video (ie 4 on the Demo data).
-                            # Requires display_detection = false, otherwise falls back to sequential (ie parallel_pose = 1).
+parallel_workers_pose = 'auto' # 'auto', int, or false # One worker per video, using GPU if available
+                               # Requires display_detection = false, otherwise falls back to sequential (ie parallel_workers_pose = 1/false).
 
 display_detection = false # Real-time display of the videos with pose estimation
 overwrite_pose = false # Set to false if you don't want to recalculate pose estimation when it has already been done
@@ -193,7 +193,8 @@ calibration_type = 'convert' # 'convert' or 'calculate'
 # for each frame, each keypoint: if results are not good, we remove a camera and try again
 reproj_error_threshold_triangulation = 15 # px # if reprojection error is above, triangulation results won't be accepted
 likelihood_threshold_triangulation= 0.3 # if 2D likelihood is below, estimations for this camera won't be accepted
-min_cameras_for_triangulation = 2 # won't try below N cameras if the triangulation is still not good
+min_cameras_for_triangulation = 2 # stops trying to triangulate if results still not good with min_cameras_for_triangulation cameras. 
+                                  # Increase if you have many cameras and want to be more robust to occlusions
 
 # gap filling
 max_distance_m = 1.0 # m # max_distance a person can jump from its previous position before being considered as a new one
@@ -263,6 +264,7 @@ make_c3d = true # save triangulated data in c3d format in addition to trc
 
 
 [kinematics]
+parallel_workers_kinematics = 'auto' # 'auto', int, or false # One worker per person, uses CPU. Only available in multi-person mode.
 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)

Pose2Sim/Demo_Batch/Trial_1/Config.toml

@@ -78,8 +78,8 @@
 #                   # 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'
-# parallel_pose = 'auto' # 'auto' or integer (number of threads). 'auto': one thread per video (ie 4 on the Demo data).
-#                             # Requires display_detection = false, otherwise falls back to sequential (ie parallel_pose = 1).
+# parallel_workers_pose = 'auto' # 'auto', int, or false # One worker per video, using GPU if available
+#                                # Requires display_detection = false, otherwise falls back to sequential (ie parallel_workers_pose = 1/false).
 
 # display_detection = true # Real-time display of the videos with pose estimation
 # overwrite_pose = false # Set to false if you don't want to recalculate pose estimation when it has already been done
@@ -193,7 +193,8 @@
 # # for each frame, each keypoint: if results are not good, we remove a camera and try again
 # reproj_error_threshold_triangulation = 15 # px # if reprojection error is above, triangulation results won't be accepted
 # likelihood_threshold_triangulation= 0.3 # if 2D likelihood is below, estimations for this camera won't be accepted
-# min_cameras_for_triangulation = 2 # won't try below N cameras if the triangulation is still not good
+# min_cameras_for_triangulation = 2 # stops trying to triangulate if results still not good with min_cameras_for_triangulation cameras. 
+#                                   Increase if you have many cameras and want to be more robust to occlusions
 
 # # gap filling
 # max_distance_m = 1.0 # m # max_distance a person can jump from its previous position before being considered as a new one
@@ -263,6 +264,7 @@
 
 
 # [kinematics]
+# parallel_workers_kinematics = 'auto' # 'auto', int, or false # One worker per person, uses CPU. Only available in multi-person mode.
 # 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)

Pose2Sim/Demo_Batch/Trial_2/Config.toml

@@ -77,8 +77,8 @@ participant_mass = [70.0, 63.5]     # float (eg 70.0), or list of floats (eg [70
 #                   # 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'
-# parallel_pose = 'auto' # 'auto' or integer (number of threads). 'auto': one thread per video (ie 4 on the Demo data).
-#                             # Requires display_detection = false, otherwise falls back to sequential (ie parallel_pose = 1).
+# parallel_workers_pose = 'auto' # 'auto', int, or false # One worker per video, using GPU if available
+#                                # Requires display_detection = false, otherwise falls back to sequential (ie parallel_workers_pose = 1/false).
 
 # display_detection = true # Real-time display of the videos with pose estimation
 # overwrite_pose = false # Set to false if you don't want to recalculate pose estimation when it has already been done
@@ -191,7 +191,8 @@ keypoints_to_consider = ['RWrist'] # 'all' if all points should be considered, f
 # # for each frame, each keypoint: if results are not good, we remove a camera and try again
 # reproj_error_threshold_triangulation = 15 # px # if reprojection error is above, triangulation results won't be accepted
 # likelihood_threshold_triangulation= 0.3 # if 2D likelihood is below, estimations for this camera won't be accepted
-# min_cameras_for_triangulation = 2 # won't try below N cameras if the triangulation is still not good
+# min_cameras_for_triangulation = 2 # stops trying to triangulate if results still not good with min_cameras_for_triangulation cameras. 
+#                                   Increase if you have many cameras and want to be more robust to occlusions
 
 # # gap filling
 # max_distance_m = 1.0 # m # max_distance a person can jump from its previous position before being considered as a new one
@@ -261,6 +262,7 @@ keypoints_to_consider = ['RWrist'] # 'all' if all points should be considered, f
 
 
 # [kinematics]
+# parallel_workers_kinematics = 'auto' # 'auto', int, or false # One worker per person, uses CPU. Only available in multi-person mode.
 # 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)

Pose2Sim/Demo_MultiPerson/Config.toml

@@ -79,8 +79,8 @@ det_frequency = 4 # Run person detection only every N frames, and inbetween trac
                   # 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'
-parallel_pose = 'auto' # 'auto' or integer (number of threads). 'auto': one thread per video (ie 4 on the Demo data).
-                            # Requires display_detection = false, otherwise falls back to sequential (ie parallel_pose = 1).
+parallel_workers_pose = 'auto' # 'auto', int, or false # One worker per video, using GPU if available
+                               # Requires display_detection = false, otherwise falls back to sequential (ie parallel_workers_pose = 1/false).
 
 display_detection = true # Real-time display of the videos with pose estimation
 overwrite_pose = false # Set to false if you don't want to recalculate pose estimation when it has already been done
@@ -193,7 +193,8 @@ calibration_type = 'convert' # 'convert' or 'calculate'
 # for each frame, each keypoint: if results are not good, we remove a camera and try again
 reproj_error_threshold_triangulation = 15 # px # if reprojection error is above, triangulation results won't be accepted
 likelihood_threshold_triangulation= 0.3 # if 2D likelihood is below, estimations for this camera won't be accepted
-min_cameras_for_triangulation = 2 # won't try below N cameras if the triangulation is still not good
+min_cameras_for_triangulation = 2 # stops trying to triangulate if results still not good with min_cameras_for_triangulation cameras. 
+                                  # Increase if you have many cameras and want to be more robust to occlusions
 
 # gap filling
 max_distance_m = 1.0 # m # max_distance a person can jump from its previous position before being considered as a new one
@@ -265,6 +266,7 @@ 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)
+parallel_workers_kinematics = 'auto' # 'auto', int, or false # One worker per person, uses CPU. Only available in multi-person mode.
 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

Pose2Sim/Demo_SinglePerson/Config.toml

@@ -77,10 +77,9 @@ det_frequency = 4 # Run person detection only every N frames, and inbetween trac
                   # 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'
-parallel_pose = 'auto' # 'auto' or integer (number of threads). 'auto': one thread per video (ie 4 on the Demo data).
-                            # Requires display_detection = false, otherwise falls back to sequential (ie parallel_pose = 1).
-
-display_detection = true # Real-time display of the videos with pose estimation
+parallel_workers_pose = 'auto' # 'auto', int, or false # One worker per video, using GPU if available
+                               # Requires display_detection = false, otherwise falls back to sequential (ie parallel_workers_pose = 1/false).
+display_detection = true # Real-time display of the videos with pose estimation. Set to false for parallel processing.
 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
@@ -192,7 +191,8 @@ calibration_type = 'convert' # 'convert' or 'calculate'
 # for each frame, each keypoint: if results are not good, we remove a camera and try again
 reproj_error_threshold_triangulation = 15 # px # if reprojection error is above, triangulation results won't be accepted
 likelihood_threshold_triangulation= 0.3 # if 2D likelihood is below, estimations for this camera won't be accepted
-min_cameras_for_triangulation = 2 # won't try below N cameras if the triangulation is still not good
+min_cameras_for_triangulation = 2 # stops trying to triangulate if results still not good with min_cameras_for_triangulation cameras. 
+                                  # Increase if you have many cameras and want to be more robust to occlusions
 
 # gap filling
 max_distance_m = 1.0 # m # max_distance a person can jump from its previous position before being considered as a new one
@@ -264,6 +264,7 @@ 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)
+parallel_workers_kinematics = 'auto' # 'auto', int, or false # One worker per person, uses CPU. Only available in multi-person mode.
 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

Pose2Sim/Utilities/tests.py

@@ -142,7 +142,7 @@ def test_workflow(self, mock_input):
         config_dict.get("project").update({"frame_rate":60})
         config_dict.get("pose").update({"det_frequency":10})
         config_dict.get("pose").update({"mode":'balanced'})
-        config_dict.get("pose").update({"parallel_pose":1})
+        config_dict.get("pose").update({"parallel_workers_pose":2})
         config_dict.get("pose").update({"overwrite_pose":True})
         config_dict.get("pose").update({"save_video":'none'})
         config_dict.get('synchronization').update({'keypoints_to_consider':['RWrist']})

Pose2Sim/common.py

@@ -190,6 +190,37 @@ def is_image_file(path):
     return os.path.isfile(path) and os.path.splitext(path)[1].lower() in image_extensions
 
 
+def get_max_workers(device='cpu'):
+    '''
+    Determine the number of parallel workers for pose estimation.
+    
+    INPUTS:
+        device: 'cpu', 'cuda', 'mps', 'rocm'
+    
+    OUTPUTS:
+        max_workers: Number of workers (>= 1)
+    '''
+    
+    if device == 'cpu':
+        # Half of the CPU cores
+        cpu_count = os.cpu_count() or 4
+        max_workers = max(1, cpu_count // 2)
+    
+    else:
+        # ~1 GB per worker, 20% headroom
+        try: 
+            # 'cuda' or 'rocm'
+            import torch
+            vram_gb = torch.cuda.mem_get_info(0)[0] / 1e9
+            max_workers = max(1, int(vram_gb * 0.8)) 
+        except Exception: 
+            # 'mps' or fallback
+            max_workers = 4
+
+    return max_workers
+
+
+
 def read_trc(trc_path):
     '''
     Read a TRC file and extract its contents.