sports2d 0.8.25__tar.gz → 0.8.26__tar.gz

sports2d-0.8.26/.github/workflows/sync_to_hf.yml.bak

@@ -0,0 +1,19 @@
+name: Sync to Hugging Face Space
+on:
+  push:
+    branches: [ main ]
+jobs:
+  sync:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - name: Push to Hugging Face Space
+        run: |
+          git clone https://${{ secrets.HUGGINGFACE_TOKEN }}@huggingface.co/spaces/DavidPagnon/sports2d
+          cd sports2d
+          git config --global user.name "DavidPagnon"
+          git config --global user.email "contact@david-pagnon.com"
+          cp -r ../Sports2D/* .
+          git add .
+          git commit -m "Sync from GitHub"
+          git push https://${{ secrets.HUGGINGFACE_TOKEN }}@huggingface.co/spaces/DavidPagnon/sports2d

{sports2d-0.8.25/sports2d.egg-info → sports2d-0.8.26}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sports2d
-Version: 0.8.25
+Version: 0.8.26
 Summary: Compute 2D human pose and angles from a video or a webcam.
 Author-email: David Pagnon <contact@david-pagnon.com>
 Maintainer-email: David Pagnon <contact@david-pagnon.com>
@@ -40,6 +40,8 @@ Dynamic: license-file
 [![License](https://img.shields.io/badge/License-BSD_3--Clause-blue.svg)](https://opensource.org/licenses/BSD-3-Clause)
 \
 [![Discord](https://img.shields.io/discord/1183750225471492206?logo=Discord&label=Discord%20community)](https://discord.com/invite/4mXUdSFjmt)
+[![Hugging Face Space](https://img.shields.io/badge/HuggingFace-Sports2D-yellow?logo=huggingface)](https://huggingface.co/spaces/DavidPagnon/sports2d)
+
 
 <!-- [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://bit.ly/Sports2D_Colab)-->
 
@@ -52,7 +54,7 @@ Dynamic: license-file
 </br>
 
 > **`Announcements:`**
-> - Compensate for floor angle, floor height, depth perspective effects, generate a calibration file **New in v0.9!** 
+> - Compensate for floor angle, floor height, depth perspective effects, generate a calibration file **New in v0.8.25!** 
 > - Select only the persons you want to analyze **New in v0.8!** 
 > - MarkerAugmentation and Inverse Kinematics for accurate 3D motion with OpenSim. **New in v0.7!** 
 > - Any detector and pose estimation model can be used. **New in v0.6!**
@@ -80,7 +82,7 @@ https://github.com/user-attachments/assets/2ce62012-f28c-4e23-b3b8-f68931bacb77
 <!-- https://github.com/user-attachments/assets/1c6e2d6b-d0cf-4165-864e-d9f01c0b8a0e -->
 
 `Warning:` Angle estimation is only as good as the pose estimation algorithm, i.e., it is not perfect.\
-`Warning:` Results are acceptable only if the persons move in the 2D plane (sagittal or frontal plane). The persons need to be filmed as parallel as possible to the motion plane.\
+`Warning:` Results are acceptable only if the persons move in the 2D plane (sagittal or frontal). The persons need to be filmed as parallel as possible to the motion plane.\
 If you need 3D research-grade markerless joint kinematics, consider using several cameras with **[Pose2Sim](https://github.com/perfanalytics/pose2sim)**.
 
 <!--`Warning:` Google Colab does not follow the European GDPR requirements regarding data privacy. [Install locally](#installation) if this matters.-->
@@ -90,7 +92,8 @@ If you need 3D research-grade markerless joint kinematics, consider using severa
 
 ## Contents
 1. [Installation and Demonstration](#installation-and-demonstration)
-   1. [Installation](#installation)
+   1. [Test it on Hugging face](#test-it-on-hugging-face)
+   1. [Local installation](#local-installation)
       1. [Quick install](#quick-install)
       2. [Full install](#full-install)
    2. [Demonstration](#demonstration)
@@ -119,7 +122,16 @@ If you need 3D research-grade markerless joint kinematics, consider using severa
 
 ## Installation and Demonstration
 
-### Installation
+
+### Test it on Hugging face
+
+Test an online, limited version [on Hugging Face](https://huggingface.co/spaces/DavidPagnon/sports2d): [![Hugging Face Space](https://img.shields.io/badge/HuggingFace-Sports2D-yellow?logo=huggingface)](https://huggingface.co/spaces/DavidPagnon/sports2d)
+
+<img src="Content/huggingface_demo.png" width="760">
+
+
+
+### Local installation
 
 <!--- OPTION 0: **Use Colab** \
   User-friendly (but full) version, also works on a phone or a tablet.\
@@ -424,7 +436,7 @@ sports2d --video_input demo.mp4 other_video.mp4 --time_range 1.2 2.7 0 3.5
   sports2d --calculate_angles false
   ```
 - Flip angles when the person faces the other side.\
-  **N.B.:** *We consider that the person looks to the right if their toe keypoint is to the right of their heel. This is not always true when the person is sprinting, especially in the swing phase. Set it to false if you want timeseries to be continuous even when the participant switches their stance.*
+  **N.B.: Set to false when sprinting.** *We consider that each limb "looks" to the right if the toe keypoint is to the right of the heel one. This is not always true, particularly during the swing phase of sprinting. Set it to false if you want timeseries to be continuous even when the participant switches their stance.*
   ```cmd
   sports2d --flip_left_right true # Default
   ```
@@ -525,6 +537,7 @@ sports2d --help
 'calib_file': ["", "path to calibration file. '' if not specified, eg no calibration file"],
 'save_calib': ["", "save calibration file. true if not specified"],
 'feet_on_floor': ["", "offset marker augmentation results so that feet are at floor level. true if not specified"],
+'distortions': ["", "camera distortion coefficients [k1, k2, p1, p2, k3] or 'from_calib'. [0.0, 0.0, 0.0, 0.0, 0.0] if not specified"],
 'use_simple_model': ["", "IK 10+ times faster, but no muscles or flexible spine, no patella. false if not specified"],
 'close_to_zero_speed_m': ["","Sum for all keypoints: about 50 px/frame or 0.2 m/frame"], 
 'tracking_mode': ["", "'sports2d' or 'deepsort'. 'deepsort' is slower, harder to parametrize but can be more robust if correctly tuned"],

{sports2d-0.8.25 → sports2d-0.8.26}/README.md

@@ -12,6 +12,8 @@
 [![License](https://img.shields.io/badge/License-BSD_3--Clause-blue.svg)](https://opensource.org/licenses/BSD-3-Clause)
 \
 [![Discord](https://img.shields.io/discord/1183750225471492206?logo=Discord&label=Discord%20community)](https://discord.com/invite/4mXUdSFjmt)
+[![Hugging Face Space](https://img.shields.io/badge/HuggingFace-Sports2D-yellow?logo=huggingface)](https://huggingface.co/spaces/DavidPagnon/sports2d)
+
 
 <!-- [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://bit.ly/Sports2D_Colab)-->
 
@@ -24,7 +26,7 @@
 </br>
 
 > **`Announcements:`**
-> - Compensate for floor angle, floor height, depth perspective effects, generate a calibration file **New in v0.9!** 
+> - Compensate for floor angle, floor height, depth perspective effects, generate a calibration file **New in v0.8.25!** 
 > - Select only the persons you want to analyze **New in v0.8!** 
 > - MarkerAugmentation and Inverse Kinematics for accurate 3D motion with OpenSim. **New in v0.7!** 
 > - Any detector and pose estimation model can be used. **New in v0.6!**
@@ -52,7 +54,7 @@ https://github.com/user-attachments/assets/2ce62012-f28c-4e23-b3b8-f68931bacb77
 <!-- https://github.com/user-attachments/assets/1c6e2d6b-d0cf-4165-864e-d9f01c0b8a0e -->
 
 `Warning:` Angle estimation is only as good as the pose estimation algorithm, i.e., it is not perfect.\
-`Warning:` Results are acceptable only if the persons move in the 2D plane (sagittal or frontal plane). The persons need to be filmed as parallel as possible to the motion plane.\
+`Warning:` Results are acceptable only if the persons move in the 2D plane (sagittal or frontal). The persons need to be filmed as parallel as possible to the motion plane.\
 If you need 3D research-grade markerless joint kinematics, consider using several cameras with **[Pose2Sim](https://github.com/perfanalytics/pose2sim)**.
 
 <!--`Warning:` Google Colab does not follow the European GDPR requirements regarding data privacy. [Install locally](#installation) if this matters.-->
@@ -62,7 +64,8 @@ If you need 3D research-grade markerless joint kinematics, consider using severa
 
 ## Contents
 1. [Installation and Demonstration](#installation-and-demonstration)
-   1. [Installation](#installation)
+   1. [Test it on Hugging face](#test-it-on-hugging-face)
+   1. [Local installation](#local-installation)
       1. [Quick install](#quick-install)
       2. [Full install](#full-install)
    2. [Demonstration](#demonstration)
@@ -91,7 +94,16 @@ If you need 3D research-grade markerless joint kinematics, consider using severa
 
 ## Installation and Demonstration
 
-### Installation
+
+### Test it on Hugging face
+
+Test an online, limited version [on Hugging Face](https://huggingface.co/spaces/DavidPagnon/sports2d): [![Hugging Face Space](https://img.shields.io/badge/HuggingFace-Sports2D-yellow?logo=huggingface)](https://huggingface.co/spaces/DavidPagnon/sports2d)
+
+<img src="Content/huggingface_demo.png" width="760">
+
+
+
+### Local installation
 
 <!--- OPTION 0: **Use Colab** \
   User-friendly (but full) version, also works on a phone or a tablet.\
@@ -396,7 +408,7 @@ sports2d --video_input demo.mp4 other_video.mp4 --time_range 1.2 2.7 0 3.5
   sports2d --calculate_angles false
   ```
 - Flip angles when the person faces the other side.\
-  **N.B.:** *We consider that the person looks to the right if their toe keypoint is to the right of their heel. This is not always true when the person is sprinting, especially in the swing phase. Set it to false if you want timeseries to be continuous even when the participant switches their stance.*
+  **N.B.: Set to false when sprinting.** *We consider that each limb "looks" to the right if the toe keypoint is to the right of the heel one. This is not always true, particularly during the swing phase of sprinting. Set it to false if you want timeseries to be continuous even when the participant switches their stance.*
   ```cmd
   sports2d --flip_left_right true # Default
   ```
@@ -497,6 +509,7 @@ sports2d --help
 'calib_file': ["", "path to calibration file. '' if not specified, eg no calibration file"],
 'save_calib': ["", "save calibration file. true if not specified"],
 'feet_on_floor': ["", "offset marker augmentation results so that feet are at floor level. true if not specified"],
+'distortions': ["", "camera distortion coefficients [k1, k2, p1, p2, k3] or 'from_calib'. [0.0, 0.0, 0.0, 0.0, 0.0] if not specified"],
 'use_simple_model': ["", "IK 10+ times faster, but no muscles or flexible spine, no patella. false if not specified"],
 'close_to_zero_speed_m': ["","Sum for all keypoints: about 50 px/frame or 0.2 m/frame"], 
 'tracking_mode': ["", "'sports2d' or 'deepsort'. 'deepsort' is slower, harder to parametrize but can be more robust if correctly tuned"],

{sports2d-0.8.25 → sports2d-0.8.26}/Sports2D/Demo/Config_demo.toml

@@ -106,13 +106,16 @@ to_meters = true
 make_c3d = true
 save_calib = true
 
+# Compensate for camera horizon 
+floor_angle = 'auto' # float, 'from_kinematics', 'from_calib', or 'auto' # 'auto' is equivalent to 'from_kinematics', ie angle calculated from foot contacts. 'from_calib' calculates it from a toml calibration file. Use float to manually specify it in degrees
+xy_origin = ['auto'] # [px_x,px_y], or ['from kinematics'], ['from_calib'], or ['auto']. # BETWEEN BRACKETS! # ['auto'] is equivalent to ['from_kinematics'], ie origin estimated at first foot contact, direction is direction of motion. ['from_calib'] calculates it from a calibration file. Use [px_x,px_y] to manually specify it in pixels (px_y points downwards)
+
 # Compensate for perspective effects, which make the further limb look smaller. 1-2% coordinate error at 10 m, less if the camera is further away
 perspective_value = 10 # Either camera-to-person distance (m), or focal length (px), or field-of-view (degrees or radians), or '' if perspective_unit=='from_calib'
 perspective_unit = 'distance_m' # 'distance_m', 'f_px', 'fov_deg', 'fov_rad', or 'from_calib'
 
-# Compensate for camera horizon 
-floor_angle = 'auto' # float, 'from_kinematics', 'from_calib', or 'auto' # 'auto' is equivalent to 'from_kinematics', ie angle calculated from foot contacts. 'from_calib' calculates it from a toml calibration file. Use float to manually specify it in degrees
-xy_origin = ['auto'] # [px_x,px_y], or ['from kinematics'], ['from_calib'], or ['auto']. # BETWEEN BRACKETS! # ['auto'] is equivalent to ['from_kinematics'], ie origin estimated at first foot contact, direction is direction of motion. ['from_calib'] calculates it from a calibration file. Use [px_x,px_y] to manually specify it in pixels (px_y points downwards)
+# Optional distortion coefficients
+distortions = [0.0, 0.0, 0.0, 0.0, 0.0] # [k1, k2, p1, p2, k3] or 'from_calib' (not implemented yet)
 
 # Optional calibration file
 calib_file = ''      # Calibration file in the Pose2Sim toml format, or '' if not available
@@ -130,7 +133,7 @@ joint_angles = ['Right ankle', 'Left ankle', 'Right knee', 'Left knee', 'Right h
 segment_angles = ['Right foot', 'Left foot', 'Right shank', 'Left shank', 'Right thigh', 'Left thigh', 'Pelvis', 'Trunk', 'Shoulders', 'Head', 'Right arm', 'Left arm', 'Right forearm', 'Left forearm']
 
 # Processing parameters
-flip_left_right = true  # Same angles whether the participant faces left/right. Set it to false if you want timeseries to be continuous even when the participant switches their stance.
+flip_left_right = false  # Same angles whether the participant faces left/right. Set it to false if you want timeseries to be continuous even when the participant switches their stance.
 correct_segment_angles_with_floor_angle = true # If the camera is tilted, corrects segment angles as regards to the floor angle. Set to false if it is the floor which is actually tilted
 
 
@@ -212,7 +215,7 @@ use_custom_logging = false # if integrated in an API that already has logging
 # 
 # Check your model hierarchy with:  for pre, _, node in RenderTree(model): 
 #                                      print(f'{pre}{node.name} id={node.id}')
-[pose.CUSTOM]
+[[pose.CUSTOM]]
 name = "Hip"
 id = 19
   [[pose.CUSTOM.children]]

{sports2d-0.8.25 → sports2d-0.8.26}/Sports2D/Sports2D.py

@@ -197,6 +197,7 @@ DEFAULT_CONFIG =   {'base': {'video_input': ['demo.mp4'],
                                 'save_calib': True,
                                 'perspective_value': 10.0,
                                 'perspective_unit': 'distance_m',
+                                'distortions': [0.0, 0.0, 0.0, 0.0, 0.0],
                                 'floor_angle': 'auto',
                                 'xy_origin': ['auto'],
                                 'calib_file': '',
@@ -311,6 +312,7 @@ CONFIG_HELP =   {'config': ["C", "path to a toml configuration file"],
                 'calib_file': ["", "path to calibration file. '' if not specified, eg no calibration file"],
                 'save_calib': ["", "save calibration file. true if not specified"],
                 'feet_on_floor': ["", "offset marker augmentation results so that feet are at floor level. true if not specified"],
+                'distortions': ["", "camera distortion coefficients [k1, k2, p1, p2, k3] or 'from_calib'. [0.0, 0.0, 0.0, 0.0, 0.0] if not specified"],
                 'use_simple_model': ["", "IK 10+ times faster, but no muscles or flexible spine, no patella. false if not specified"],
                 'close_to_zero_speed_m': ["","Sum for all keypoints: about 50 px/frame or 0.2 m/frame"], 
                 'tracking_mode': ["", "'sports2d' or 'deepsort'. 'deepsort' is slower, harder to parametrize but can be more robust if correctly tuned"],

{sports2d-0.8.25 → sports2d-0.8.26}/Sports2D/process.py

@@ -242,7 +242,7 @@ def setup_model_class_mode(pose_model, mode, config_dict={}):
         try: # from Config.toml
             from anytree.importer import DictImporter
             model_name = pose_model.upper()
-            pose_model = DictImporter().import_(config_dict.get('pose').get(pose_model))
+            pose_model = DictImporter().import_(config_dict.get('pose').get(pose_model)[0])
             if pose_model.id == 'None':
                 pose_model.id = None
             logging.info(f"Using model {model_name} for pose estimation.")
@@ -798,7 +798,6 @@ def make_mot_with_angles(angles, time, mot_path):
 def pose_plots(trc_data_unfiltered, trc_data, person_id, show=True):
     '''
     Displays trc filtered and unfiltered data for comparison
-    ⚠ Often crashes on the third window...
 
     INPUTS:
     - trc_data_unfiltered: pd.DataFrame. The unfiltered trc data
@@ -809,23 +808,26 @@ def pose_plots(trc_data_unfiltered, trc_data, person_id, show=True):
     OUTPUT:
     - matplotlib window with tabbed figures for each keypoint
     '''
-
+    
     os_name = platform.system()
-    if os_name == 'Windows':
-        mpl.use('qt5agg') # windows
     mpl.rc('figure', max_open_warning=0)
+    if show:
+        if os_name == 'Windows':
+            mpl.use('qt5agg') # windows
+        pw = plotWindow()
+        pw.MainWindow.setWindowTitle('Person'+ str(person_id) + ' coordinates')
+    else:
+        mpl.use('Agg') # Otherwise fails on Hugging-face
+        figures_list = []    
 
     keypoints_names = trc_data.columns[1::3]
-    
-    pw = plotWindow()
-    pw.MainWindow.setWindowTitle('Person'+ str(person_id) + ' coordinates') # Main title
-
     for id, keypoint in enumerate(keypoints_names):
         f = plt.figure()
-        if os_name == 'Windows':
-            f.canvas.manager.window.setWindowTitle(keypoint + ' Plot') # windows
-        elif os_name == 'Darwin':  # macOS
-            f.canvas.manager.set_window_title(keypoint + ' Plot') # mac
+        if show:
+            if os_name == 'Windows':
+                f.canvas.manager.window.setWindowTitle(keypoint + ' Plot')
+            elif os_name == 'Darwin':
+                f.canvas.manager.set_window_title(keypoint + ' Plot')
 
         axX = plt.subplot(211)
         plt.plot(trc_data_unfiltered.iloc[:,0], trc_data_unfiltered.iloc[:,id*3+1], label='unfiltered')
@@ -840,18 +842,21 @@ def pose_plots(trc_data_unfiltered, trc_data, person_id, show=True):
         axY.set_xlabel('Time (seconds)')
         axY.set_ylabel(keypoint+' Y')
 
-        pw.addPlot(keypoint, f)
+        if show:
+            pw.addPlot(keypoint, f)
+        else:
+            figures_list.append((keypoint, f))
     
     if show:
         pw.show()
-
-    return pw
-
+        return pw
+    else:
+        return figures_list
+    
 
 def angle_plots(angle_data_unfiltered, angle_data, person_id, show=True):
     '''
     Displays angle filtered and unfiltered data for comparison
-    ⚠ Often crashes on the third window...
 
     INPUTS:
     - angle_data_unfiltered: pd.DataFrame. The unfiltered angle data
@@ -862,21 +867,24 @@ def angle_plots(angle_data_unfiltered, angle_data, person_id, show=True):
     '''
 
     os_name = platform.system()
-    if os_name == 'Windows':
-        mpl.use('qt5agg') # windows
     mpl.rc('figure', max_open_warning=0)
+    if show:
+        if os_name == 'Windows':
+            mpl.use('qt5agg') # windows
+        pw = plotWindow()
+        pw.MainWindow.setWindowTitle('Person'+ str(person_id) + ' angles')
+    else:
+        mpl.use('Agg') # Otherwise fails on Hugging-face
+        figures_list = []
 
     angles_names = angle_data.columns[1:]
-    
-    pw = plotWindow()
-    pw.MainWindow.setWindowTitle('Person'+ str(person_id) + ' angles') # Main title
-
     for id, angle in enumerate(angles_names):
         f = plt.figure()
-        if os_name == 'Windows':
-            f.canvas.manager.window.setWindowTitle(angle + ' Plot') # windows
-        elif os_name == 'Darwin':  # macOS
-            f.canvas.manager.set_window_title(angle + ' Plot') # mac
+        if show:
+            if os_name == 'Windows':
+                f.canvas.manager.window.setWindowTitle(angle + ' Plot') # windows
+            elif os_name == 'Darwin':  # macOS
+                f.canvas.manager.set_window_title(angle + ' Plot') # mac
         
         ax = plt.subplot(111)
         plt.plot(angle_data_unfiltered.iloc[:,0], angle_data_unfiltered.iloc[:,id+1], label='unfiltered')
@@ -886,12 +894,16 @@ def angle_plots(angle_data_unfiltered, angle_data, person_id, show=True):
         ax.set_ylabel(angle+' (°)')
         plt.legend()
 
-        pw.addPlot(angle, f)
-
+        if show:
+            pw.addPlot(angle, f)
+        else:
+            figures_list.append((angle, f))
+    
     if show:
         pw.show()
-
-    return pw
+        return pw
+    else:
+        return figures_list
 
 
 def get_personIDs_with_highest_scores(all_frames_scores, nb_persons_to_detect):
@@ -1306,7 +1318,7 @@ def compute_floor_line(trc_data, score_data, keypoint_names = ['LBigToe', 'RBigT
         trc_data_kpt_trim = trc_data_kpt.iloc[start:end].reset_index(drop=True)
         score_data_kpt_trim = score_data_kpt.iloc[start:end].reset_index(drop=True)
 
-        # Compute speeds
+        # Compute euclidean speed
         speeds = np.linalg.norm(trc_data_kpt_trim.diff(), axis=1)
 
         # Remove speeds with low confidence
@@ -1450,7 +1462,8 @@ def get_floor_params(floor_angle='auto', xy_origin=['auto'],
     except:
         floor_angle_kin = 0
         xy_origin_kin = cam_width/2, cam_height/2
-        logging.warning(f'Could not estimate the floor angle and xy_origin from person {0}. Make sure that the full body is visible. Using floor angle = 0° and xy_origin = [{cam_width/2}, {cam_height/2}] px.')
+        gait_direction = 1
+        logging.warning(f'Could not estimate the floor angle, xy_origin, and visible from person {0}. Make sure that the full body is visible. Using floor angle = 0°, xy_origin = [{cam_width/2}, {cam_height/2}] px, and visible_side = right.')
 
     # Determine final floor angle estimation
     if floor_angle == 'from_calib':
@@ -1578,7 +1591,7 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
     
     # Base parameters
     video_dir = Path(config_dict.get('base').get('video_dir'))
-
+    
     nb_persons_to_detect = config_dict.get('base').get('nb_persons_to_detect')
     if nb_persons_to_detect != 'all': 
         try:
@@ -1825,6 +1838,7 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
         keypoints_names = [node.name for _, _, node in RenderTree(pose_model) if node.id!=None]
         t0 = 0
         tf = (cap.get(cv2.CAP_PROP_FRAME_COUNT)-1) / fps if cap.get(cv2.CAP_PROP_FRAME_COUNT)>0 else float('inf')
+        kpt_id_max = max(keypoints_ids)+1
 
         # Set up pose tracker
         try:
@@ -1913,60 +1927,64 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
                 if video_file == "webcam":
                     out_vid.write(frame)
 
-                # Detect poses
-                keypoints, scores = pose_tracker(frame)
-
-                # Non maximum suppression (at pose level, not detection, and only using likely keypoints)
-                frame_shape = frame.shape
-                mask_scores = np.mean(scores, axis=1) > 0.2
-
-                likely_keypoints = np.where(mask_scores[:, np.newaxis, np.newaxis], keypoints, np.nan)
-                likely_scores = np.where(mask_scores[:, np.newaxis], scores, np.nan)
-                likely_bboxes = bbox_xyxy_compute(frame_shape, likely_keypoints, padding=0)
-                score_likely_bboxes = np.nanmean(likely_scores, axis=1)
-
-                valid_indices = np.where(~np.isnan(score_likely_bboxes))[0]
-                if len(valid_indices) > 0:
-                    valid_bboxes = likely_bboxes[valid_indices]
-                    valid_scores = score_likely_bboxes[valid_indices]
-                    keep_valid = nms(valid_bboxes, valid_scores, nms_thr=0.45)
-                    keep = valid_indices[keep_valid]
-                else:
-                    keep = []
-                keypoints, scores = likely_keypoints[keep], likely_scores[keep]
-
-                # # Debugging: display detected keypoints on the frame
-                # colors = [(255,0,0), (0,255,0), (0,0,255), (255,255,0), (255,0,255), (0,255,255), (128,0,0), (0,128,0), (0,0,128), (128,128,0), (128,0,128), (0,128,128)]
-                # bboxes = likely_bboxes[keep]
-                # for person_idx in range(len(keypoints)):
-                #     for kpt_idx, kpt in enumerate(keypoints[person_idx]):
-                #         if not np.isnan(kpt).any():
-                #             cv2.circle(frame, (int(kpt[0]), int(kpt[1])), 3, colors[person_idx%len(colors)], -1)
-                #     if not np.isnan(bboxes[person_idx]).any():
-                #         cv2.rectangle(frame, (int(bboxes[person_idx][0]), int(bboxes[person_idx][1])), (int(bboxes[person_idx][2]), int(bboxes[person_idx][3])), colors[person_idx%len(colors)], 1)
-                #     cv2.imshow(f'{video_file} Sports2D', frame)
-
-                # Track poses across frames
-                if tracking_mode == 'deepsort':
-                    keypoints, scores = sort_people_deepsort(keypoints, scores, deepsort_tracker, frame, frame_count)
-                if tracking_mode == 'sports2d': 
-                    if 'prev_keypoints' not in locals(): prev_keypoints = keypoints
-                    prev_keypoints, keypoints, scores = sort_people_sports2d(prev_keypoints, keypoints, scores=scores, max_dist=max_distance)
-                else:
-                    pass
-                
-                # # Debugging: display detected keypoints on the frame
-                # colors = [(255,0,0), (0,255,0), (0,0,255), (255,255,0), (255,0,255), (0,255,255), (128,0,0), (0,128,0), (0,0,128), (128,128,0), (128,0,128), (0,128,128)]
-                # for person_idx in range(len(keypoints)):
-                #     for kpt_idx, kpt in enumerate(keypoints[person_idx]):
-                #         if not np.isnan(kpt).any():
-                #             cv2.circle(frame, (int(kpt[0]), int(kpt[1])), 3, colors[person_idx%len(colors)], -1)
-                #         # if not np.isnan(bboxes[person_idx]).any():
-                #         #     cv2.rectangle(frame, (int(bboxes[person_idx][0]), int(bboxes[person_idx][1])), (int(bboxes[person_idx][2]), int(bboxes[person_idx][3])), colors[person_idx%len(colors)], 1)
-                #     cv2.imshow(f'{video_file} Sports2D', frame)
-                # # if (cv2.waitKey(1) & 0xFF) == ord('q') or (cv2.waitKey(1) & 0xFF) == 27:
-                # #     break
-                # # input()
+                try: # Frames with no detection cause errors on MacOS CoreMLExecutionProvider
+                    # Detect poses
+                    keypoints, scores = pose_tracker(frame)
+
+                    # Non maximum suppression (at pose level, not detection, and only using likely keypoints)
+                    frame_shape = frame.shape
+                    mask_scores = np.mean(scores, axis=1) > 0.2
+
+                    likely_keypoints = np.where(mask_scores[:, np.newaxis, np.newaxis], keypoints, np.nan)
+                    likely_scores = np.where(mask_scores[:, np.newaxis], scores, np.nan)
+                    likely_bboxes = bbox_xyxy_compute(frame_shape, likely_keypoints, padding=0)
+                    score_likely_bboxes = np.nanmean(likely_scores, axis=1)
+
+                    valid_indices = np.where(~np.isnan(score_likely_bboxes))[0]
+                    if len(valid_indices) > 0:
+                        valid_bboxes = likely_bboxes[valid_indices]
+                        valid_scores = score_likely_bboxes[valid_indices]
+                        keep_valid = nms(valid_bboxes, valid_scores, nms_thr=0.45)
+                        keep = valid_indices[keep_valid]
+                    else:
+                        keep = []
+                    keypoints, scores = likely_keypoints[keep], likely_scores[keep]
+
+                    # # Debugging: display detected keypoints on the frame
+                    # colors = [(255,0,0), (0,255,0), (0,0,255), (255,255,0), (255,0,255), (0,255,255), (128,0,0), (0,128,0), (0,0,128), (128,128,0), (128,0,128), (0,128,128)]
+                    # bboxes = likely_bboxes[keep]
+                    # for person_idx in range(len(keypoints)):
+                    #     for kpt_idx, kpt in enumerate(keypoints[person_idx]):
+                    #         if not np.isnan(kpt).any():
+                    #             cv2.circle(frame, (int(kpt[0]), int(kpt[1])), 3, colors[person_idx%len(colors)], -1)
+                    #     if not np.isnan(bboxes[person_idx]).any():
+                    #         cv2.rectangle(frame, (int(bboxes[person_idx][0]), int(bboxes[person_idx][1])), (int(bboxes[person_idx][2]), int(bboxes[person_idx][3])), colors[person_idx%len(colors)], 1)
+                    #     cv2.imshow(f'{video_file} Sports2D', frame)
+
+                    # Track poses across frames
+                    if tracking_mode == 'deepsort':
+                        keypoints, scores = sort_people_deepsort(keypoints, scores, deepsort_tracker, frame, frame_count)
+                    if tracking_mode == 'sports2d': 
+                        if 'prev_keypoints' not in locals(): prev_keypoints = keypoints
+                        prev_keypoints, keypoints, scores = sort_people_sports2d(prev_keypoints, keypoints, scores=scores, max_dist=max_distance)
+                    else:
+                        pass
+                    
+                    # # Debugging: display detected keypoints on the frame
+                    # colors = [(255,0,0), (0,255,0), (0,0,255), (255,255,0), (255,0,255), (0,255,255), (128,0,0), (0,128,0), (0,0,128), (128,128,0), (128,0,128), (0,128,128)]
+                    # for person_idx in range(len(keypoints)):
+                    #     for kpt_idx, kpt in enumerate(keypoints[person_idx]):
+                    #         if not np.isnan(kpt).any():
+                    #             cv2.circle(frame, (int(kpt[0]), int(kpt[1])), 3, colors[person_idx%len(colors)], -1)
+                    #         # if not np.isnan(bboxes[person_idx]).any():
+                    #         #     cv2.rectangle(frame, (int(bboxes[person_idx][0]), int(bboxes[person_idx][1])), (int(bboxes[person_idx][2]), int(bboxes[person_idx][3])), colors[person_idx%len(colors)], 1)
+                    #     cv2.imshow(f'{video_file} Sports2D', frame)
+                    # # if (cv2.waitKey(1) & 0xFF) == ord('q') or (cv2.waitKey(1) & 0xFF) == 27:
+                    # #     break
+                    # # input()
+                except:
+                    keypoints = np.full((1,kpt_id_max,2), fill_value=np.nan)
+                    scores = np.full((1,kpt_id_max), fill_value=np.nan)
             
             # Process coordinates and compute angles
             valid_X, valid_Y, valid_scores = [], [], []
@@ -2058,6 +2076,10 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
                 if (cv2.waitKey(1) & 0xFF) == ord('q') or (cv2.waitKey(1) & 0xFF) == 27:
                     break
 
+                # # Debugging
+                # img_output_path = img_output_dir / f'{video_file_stem}_frame{frame_nb:06d}.png'
+                # cv2.imwrite(str(img_output_path), img)
+
             all_frames_X.append(np.array(valid_X))
             all_frames_X_flipped.append(np.array(valid_X_flipped))
             all_frames_Y.append(np.array(valid_Y))
@@ -2260,12 +2282,20 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
             if not to_meters and (show_plots or save_plots):
                 pw = pose_plots(trc_data_unfiltered_i, trc_data_i, i, show=show_plots)
                 if save_plots:
-                    for n, f in enumerate(pw.figure_handles):
-                        dpi = pw.canvases[i].figure.dpi
-                        f.set_size_inches(1280/dpi, 720/dpi)
-                        title = pw.tabs.tabText(n)
-                        plot_path = plots_output_dir / (pose_output_path.stem + f'_person{i:02d}_px_{title.replace(" ","_").replace("/","_")}.png')
-                        f.savefig(plot_path, dpi=dpi, bbox_inches='tight')
+                    if show_plots:
+                        for n, f in enumerate(pw.figure_handles):
+                            dpi = pw.canvases[n].figure.dpi
+                            f.set_size_inches(1280/dpi, 720/dpi)
+                            title = pw.tabs.tabText(n)
+                            plot_path = plots_output_dir / (pose_output_path.stem + f'_person{i:02d}_px_{title.replace(" ","_").replace("/","_")}.png')
+                            f.savefig(plot_path, dpi=dpi, bbox_inches='tight')
+                    else: # Tabbed plots not used
+                        for title, f in pw:
+                            dpi = f.dpi
+                            f.set_size_inches(1280/dpi, 720/dpi)
+                            plot_path = plots_output_dir / (pose_output_path.stem + f'_person{i:02d}_px_{title.replace(" ","_").replace("/","_")}.png')
+                            f.savefig(plot_path, dpi=dpi, bbox_inches='tight')
+                            plt.close(f)
                     logging.info(f'Pose plots (px) saved in {plots_output_dir}.')
                     
             all_frames_X_processed[:,idx_person,:], all_frames_Y_processed[:,idx_person,:] = all_frames_X_person_filt, all_frames_Y_person_filt
@@ -2334,41 +2364,40 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
             message = get_correction_message(xy_origin)
             logging.info(f'Floor level: {cy:.2f} px (from the top of the image), gait starting at {cx:.2f} px in the {direction_person0} direction for the first person. Corrected using {message}\n')
 
+            # Prepare calibration data
+            R90z = np.array([[0.0, -1.0, 0.0],
+                            [1.0, 0.0, 0.0],
+                            [0.0, 0.0, 1.0]])
+            R270x = np.array([[1.0, 0.0, 0.0],
+                            [0.0, 0.0, 1.0],
+                            [0.0, -1.0, 0.0]])
+            
+            calib_file_path = output_dir / f'{video_file_stem}_Sports2D_calib.toml'
+            
+            # name, size, distortions
+            N = [video_file_stem]
+            S = [[cam_width, cam_height]]
+            D = [[0.0, 0.0, 0.0, 0.0]]
+
+            # Intrinsics
+            f = height_px / first_person_height * distance_m
+            cu = cam_width/2
+            cv = cam_height/2
+            K = np.array([[[f, 0.0, cu], [0.0, f, cv], [0.0, 0.0, 1.0]]])
+
+            # Extrinsics
+            Rfloory = np.array([[np.cos(floor_angle_estim), 0.0, np.sin(floor_angle_estim)],
+                                [0.0, 1.0, 0.0],
+                                [-np.sin(floor_angle_estim), 0.0, np.cos(floor_angle_estim)]])
+            R_world = R90z @ Rfloory @ R270x
+            T_world = R90z @ np.array([-(cx-cu)/f*distance_m, -distance_m, (cy-cv)/f*distance_m])
+            
+            R_cam, T_cam = world_to_camera_persp(R_world, T_world)
+            Tvec_cam = T_cam.reshape(1,3).tolist()
+            Rvec_cam = cv2.Rodrigues(R_cam)[0].reshape(1,3).tolist()
 
             # Save calibration file
             if save_calib and not calib_file:
-                R90z = np.array([[0.0, -1.0, 0.0],
-                                [1.0, 0.0, 0.0],
-                                [0.0, 0.0, 1.0]])
-                R270x = np.array([[1.0, 0.0, 0.0],
-                                [0.0, 0.0, 1.0],
-                                [0.0, -1.0, 0.0]])
-                
-                calib_file_path = output_dir / f'{video_file_stem}_Sports2D_calib.toml'
-                
-                # name, size, distortions
-                N = [video_file_stem]
-                S = [[cam_width, cam_height]]
-                D = [[0.0, 0.0, 0.0, 0.0]]
-
-                # Intrinsics
-                f = height_px / first_person_height * distance_m
-                cu = cam_width/2
-                cv = cam_height/2
-                K = np.array([[[f, 0.0, cu], [0.0, f, cv], [0.0, 0.0, 1.0]]])
-
-                # Extrinsics
-                Rfloory = np.array([[np.cos(floor_angle_estim), 0.0, np.sin(floor_angle_estim)],
-                                    [0.0, 1.0, 0.0],
-                                    [-np.sin(floor_angle_estim), 0.0, np.cos(floor_angle_estim)]])
-                R_world = R90z @ Rfloory @ R270x
-                T_world = R90z @ np.array([-(cx-cu)/f*distance_m, -distance_m, (cy-cv)/f*distance_m])
-                
-                R_cam, T_cam = world_to_camera_persp(R_world, T_world)
-                Tvec_cam = T_cam.reshape(1,3).tolist()
-                Rvec_cam = cv2.Rodrigues(R_cam)[0].reshape(1,3).tolist()
-
-                # Write calibration file
                 toml_write(calib_file_path, N, S, D, K, Rvec_cam, Tvec_cam)
                 logging.info(f'Calibration saved to {calib_file_path}.')
 
@@ -2417,12 +2446,20 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
                     if to_meters and (show_plots or save_plots):
                         pw = pose_plots(trc_data_unfiltered_m_i, trc_data_m_i, i, show=show_plots)
                         if save_plots:
-                            for n, f in enumerate(pw.figure_handles):
-                                dpi = pw.canvases[i].figure.dpi
-                                f.set_size_inches(1280/dpi, 720/dpi)
-                                title = pw.tabs.tabText(n)
-                                plot_path = plots_output_dir / (pose_output_path_m.stem + f'_person{i:02d}_m_{title.replace(" ","_").replace("/","_")}.png')
-                                f.savefig(plot_path, dpi=dpi, bbox_inches='tight')
+                            if show_plots:
+                                for n, f in enumerate(pw.figure_handles):
+                                    dpi = pw.canvases[n].figure.dpi
+                                    f.set_size_inches(1280/dpi, 720/dpi)
+                                    title = pw.tabs.tabText(n)
+                                    plot_path = plots_output_dir / (pose_output_path.stem + f'_person{i:02d}_m_{title.replace(" ","_").replace("/","_")}.png')
+                                    f.savefig(plot_path, dpi=dpi, bbox_inches='tight')
+                            else: # Tabbed plots not used
+                                for title, f in pw:
+                                    dpi = f.dpi
+                                    f.set_size_inches(1280/dpi, 720/dpi)
+                                    plot_path = plots_output_dir / (pose_output_path.stem + f'_person{i:02d}_m_{title.replace(" ","_").replace("/","_")}.png')
+                                    f.savefig(plot_path, dpi=dpi, bbox_inches='tight')
+                                    plt.close(f)
                             logging.info(f'Pose plots (m) saved in {plots_output_dir}.')
 
                     # Write to trc file
@@ -2553,12 +2590,20 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
             if show_plots or save_plots:
                 pw = angle_plots(all_frames_angles_person, angle_data, i, show=show_plots) # i = current person
                 if save_plots:
-                    for n, f in enumerate(pw.figure_handles):
-                        dpi = pw.canvases[i].figure.dpi
-                        f.set_size_inches(1280/dpi, 720/dpi)
-                        title = pw.tabs.tabText(n)
-                        plot_path = plots_output_dir / (pose_output_path_m.stem + f'_person{i:02d}_ang_{title.replace(" ","_").replace("/","_")}.png')
-                        f.savefig(plot_path, dpi=dpi, bbox_inches='tight')
+                    if show_plots:
+                        for n, f in enumerate(pw.figure_handles):
+                            dpi = pw.canvases[n].figure.dpi
+                            f.set_size_inches(1280/dpi, 720/dpi)
+                            title = pw.tabs.tabText(n)
+                            plot_path = plots_output_dir / (pose_output_path.stem + f'_person{i:02d}_ang_{title.replace(" ","_").replace("/","_")}.png')
+                            f.savefig(plot_path, dpi=dpi, bbox_inches='tight')
+                    else: # Tabbed plots not used
+                        for title, f in pw:
+                            dpi = f.dpi
+                            f.set_size_inches(1280/dpi, 720/dpi)
+                            plot_path = plots_output_dir / (pose_output_path.stem + f'_person{i:02d}_ang_{title.replace(" ","_").replace("/","_")}.png')
+                            f.savefig(plot_path, dpi=dpi, bbox_inches='tight')
+                            plt.close(f)
                     logging.info(f'Pose plots (m) saved in {plots_output_dir}.')
 
 

{sports2d-0.8.25 → sports2d-0.8.26/sports2d.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sports2d
-Version: 0.8.25
+Version: 0.8.26
 Summary: Compute 2D human pose and angles from a video or a webcam.
 Author-email: David Pagnon <contact@david-pagnon.com>
 Maintainer-email: David Pagnon <contact@david-pagnon.com>
@@ -40,6 +40,8 @@ Dynamic: license-file
 [![License](https://img.shields.io/badge/License-BSD_3--Clause-blue.svg)](https://opensource.org/licenses/BSD-3-Clause)
 \
 [![Discord](https://img.shields.io/discord/1183750225471492206?logo=Discord&label=Discord%20community)](https://discord.com/invite/4mXUdSFjmt)
+[![Hugging Face Space](https://img.shields.io/badge/HuggingFace-Sports2D-yellow?logo=huggingface)](https://huggingface.co/spaces/DavidPagnon/sports2d)
+
 
 <!-- [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://bit.ly/Sports2D_Colab)-->
 
@@ -52,7 +54,7 @@ Dynamic: license-file
 </br>
 
 > **`Announcements:`**
-> - Compensate for floor angle, floor height, depth perspective effects, generate a calibration file **New in v0.9!** 
+> - Compensate for floor angle, floor height, depth perspective effects, generate a calibration file **New in v0.8.25!** 
 > - Select only the persons you want to analyze **New in v0.8!** 
 > - MarkerAugmentation and Inverse Kinematics for accurate 3D motion with OpenSim. **New in v0.7!** 
 > - Any detector and pose estimation model can be used. **New in v0.6!**
@@ -80,7 +82,7 @@ https://github.com/user-attachments/assets/2ce62012-f28c-4e23-b3b8-f68931bacb77
 <!-- https://github.com/user-attachments/assets/1c6e2d6b-d0cf-4165-864e-d9f01c0b8a0e -->
 
 `Warning:` Angle estimation is only as good as the pose estimation algorithm, i.e., it is not perfect.\
-`Warning:` Results are acceptable only if the persons move in the 2D plane (sagittal or frontal plane). The persons need to be filmed as parallel as possible to the motion plane.\
+`Warning:` Results are acceptable only if the persons move in the 2D plane (sagittal or frontal). The persons need to be filmed as parallel as possible to the motion plane.\
 If you need 3D research-grade markerless joint kinematics, consider using several cameras with **[Pose2Sim](https://github.com/perfanalytics/pose2sim)**.
 
 <!--`Warning:` Google Colab does not follow the European GDPR requirements regarding data privacy. [Install locally](#installation) if this matters.-->
@@ -90,7 +92,8 @@ If you need 3D research-grade markerless joint kinematics, consider using severa
 
 ## Contents
 1. [Installation and Demonstration](#installation-and-demonstration)
-   1. [Installation](#installation)
+   1. [Test it on Hugging face](#test-it-on-hugging-face)
+   1. [Local installation](#local-installation)
       1. [Quick install](#quick-install)
       2. [Full install](#full-install)
    2. [Demonstration](#demonstration)
@@ -119,7 +122,16 @@ If you need 3D research-grade markerless joint kinematics, consider using severa
 
 ## Installation and Demonstration
 
-### Installation
+
+### Test it on Hugging face
+
+Test an online, limited version [on Hugging Face](https://huggingface.co/spaces/DavidPagnon/sports2d): [![Hugging Face Space](https://img.shields.io/badge/HuggingFace-Sports2D-yellow?logo=huggingface)](https://huggingface.co/spaces/DavidPagnon/sports2d)
+
+<img src="Content/huggingface_demo.png" width="760">
+
+
+
+### Local installation
 
 <!--- OPTION 0: **Use Colab** \
   User-friendly (but full) version, also works on a phone or a tablet.\
@@ -424,7 +436,7 @@ sports2d --video_input demo.mp4 other_video.mp4 --time_range 1.2 2.7 0 3.5
   sports2d --calculate_angles false
   ```
 - Flip angles when the person faces the other side.\
-  **N.B.:** *We consider that the person looks to the right if their toe keypoint is to the right of their heel. This is not always true when the person is sprinting, especially in the swing phase. Set it to false if you want timeseries to be continuous even when the participant switches their stance.*
+  **N.B.: Set to false when sprinting.** *We consider that each limb "looks" to the right if the toe keypoint is to the right of the heel one. This is not always true, particularly during the swing phase of sprinting. Set it to false if you want timeseries to be continuous even when the participant switches their stance.*
   ```cmd
   sports2d --flip_left_right true # Default
   ```
@@ -525,6 +537,7 @@ sports2d --help
 'calib_file': ["", "path to calibration file. '' if not specified, eg no calibration file"],
 'save_calib': ["", "save calibration file. true if not specified"],
 'feet_on_floor': ["", "offset marker augmentation results so that feet are at floor level. true if not specified"],
+'distortions': ["", "camera distortion coefficients [k1, k2, p1, p2, k3] or 'from_calib'. [0.0, 0.0, 0.0, 0.0, 0.0] if not specified"],
 'use_simple_model': ["", "IK 10+ times faster, but no muscles or flexible spine, no patella. false if not specified"],
 'close_to_zero_speed_m': ["","Sum for all keypoints: about 50 px/frame or 0.2 m/frame"], 
 'tracking_mode': ["", "'sports2d' or 'deepsort'. 'deepsort' is slower, harder to parametrize but can be more robust if correctly tuned"],

{sports2d-0.8.25 → sports2d-0.8.26}/sports2d.egg-info/SOURCES.txt

@@ -6,11 +6,13 @@ pyproject.toml
 .github/workflows/continuous-integration.yml
 .github/workflows/joss_pdf.yml
 .github/workflows/publish-on-release.yml
+.github/workflows/sync_to_hf.yml.bak
 Content/Demo_plots.png
 Content/Demo_results.png
 Content/Demo_terminal.png
 Content/Person_selection.png
 Content/Video_tuto_Sports2D_Colab.png
+Content/huggingface_demo.png
 Content/joint_convention.png
 Content/paper.bib
 Content/paper.md