sports2d 0.5.4__tar.gz → 0.5.6__tar.gz

{sports2d-0.5.4 → sports2d-0.5.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sports2d
-Version: 0.5.4
+Version: 0.5.6
 Summary: Detect pose and compute 2D joint angles from a video.
 Home-page: https://github.com/davidpagnon/Sports2D
 Author: David Pagnon
@@ -233,9 +233,10 @@ Note that it does not take distortions into account, and that it will be less ac
 ### Too slow for you?
 
 **Quick fixes:**
-- Use `--multiperson false`: Can be used if one single person is present in the video. Otherwise, persons' IDs may be mixed up.
+- Use ` --save_vid false --save_img false --show_realtime_results false`: Will not save images or videos, and will not display the results in real time. 
 - Use `--mode lightweight`: Will use a lighter version of RTMPose, which is faster but less accurate.
 - Use `--det_frequency 50`: Will detect poses only every 50 frames, and track keypoints in between, which is faster.
+- Use `--multiperson false`: Can be used if one single person is present in the video. Otherwise, persons' IDs may be mixed up.
 - Use `--load_trc <path_to_file_px.trc>`: Will use pose estimation results from a file. Useful if you want to use different parameters for pixel to meter conversion or angle calculation without running detection and pose estimation all over.
 
 <br> 

{sports2d-0.5.4 → sports2d-0.5.6}/README.md

@@ -193,9 +193,10 @@ Note that it does not take distortions into account, and that it will be less ac
 ### Too slow for you?
 
 **Quick fixes:**
-- Use `--multiperson false`: Can be used if one single person is present in the video. Otherwise, persons' IDs may be mixed up.
+- Use ` --save_vid false --save_img false --show_realtime_results false`: Will not save images or videos, and will not display the results in real time. 
 - Use `--mode lightweight`: Will use a lighter version of RTMPose, which is faster but less accurate.
 - Use `--det_frequency 50`: Will detect poses only every 50 frames, and track keypoints in between, which is faster.
+- Use `--multiperson false`: Can be used if one single person is present in the video. Otherwise, persons' IDs may be mixed up.
 - Use `--load_trc <path_to_file_px.trc>`: Will use pose estimation results from a file. Useful if you want to use different parameters for pixel to meter conversion or angle calculation without running detection and pose estimation all over.
 
 <br> 

{sports2d-0.5.4 → sports2d-0.5.6}/Sports2D/Demo/Config_demo.toml

@@ -21,7 +21,7 @@ compare = false         # Not implemented yet
 
 # Video parameters
 time_range = []   # [] for the whole video, or [start_time, end_time] (in seconds), or [[start_time1, end_time1], [start_time2, end_time2], ...]
-video_dir = ''    # If empty, result dir is current dir
+video_dir = ''    # If empty, video dir is current dir
 
 # Webcam parameters
 webcam_id = 0 # your webcam id (0 is default)

{sports2d-0.5.4 → sports2d-0.5.6}/Sports2D/process.py

@@ -1101,7 +1101,7 @@ def get_personID_with_highest_scores(all_frames_scores):
     return person_id
 
 
-def compute_floor_line(trc_data, keypoint_names = ['LBigToe', 'RBigToe'], speed_threshold = 2.5):
+def compute_floor_line(trc_data, keypoint_names = ['LBigToe', 'RBigToe'], toe_speed_below = 1.0, tot_speed_above=2.0):
     '''
     Compute the floor line equation and angle 
     from the feet keypoints when they have zero speed.
@@ -1111,19 +1111,23 @@ def compute_floor_line(trc_data, keypoint_names = ['LBigToe', 'RBigToe'], speed_
     INPUTS:
     - trc_data: pd.DataFrame. The trc data
     - keypoint_names: list of str. The names of the keypoints to use
-    - speed_threshold: float. The speed threshold below which the keypoints are considered as not moving
+    - toe_speed_below: float. The speed threshold (px/frame) below which the keypoints are considered as not moving
 
     OUTPUT:
     - angle: float. The angle of the floor line in radians
     - xy_origin: list. The origin of the floor line
     '''
 
+    # Remove frames where the person is mostly not moving (outlier)
+    av_speeds = np.nanmean([np.insert(np.linalg.norm(trc_data[kpt].diff(), axis=1)[1:],0,0) for kpt in trc_data.columns.unique()[1:]], axis=0)
+    trc_data = trc_data[av_speeds>tot_speed_above]
+
     # Retrieve zero-speed coordinates for the foot
     low_speeds_X, low_speeds_Y = [], []
     for kpt in keypoint_names:
         speeds = np.linalg.norm(trc_data[kpt].diff(), axis=1)
         
-        low_speed_frames = trc_data[speeds<speed_threshold].index
+        low_speed_frames = trc_data[speeds<toe_speed_below].index
         low_speeds_coords = trc_data[kpt].loc[low_speed_frames]
         low_speeds_coords = low_speeds_coords[low_speeds_coords!=0]
 
@@ -1416,7 +1420,6 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
     gaussian_filter_kernel = config_dict.get('post-processing').get('gaussian').get('sigma_kernel')
     loess_filter_kernel = config_dict.get('post-processing').get('loess').get('nb_values_used')
     median_filter_kernel = config_dict.get('post-processing').get('median').get('kernel_size')
-    butterworth_filter_cutoff /= slowmo_factor
     filter_options = [do_filter, filter_type,
                            butterworth_filter_order, butterworth_filter_cutoff, frame_rate,
                            gaussian_filter_kernel, loess_filter_kernel, median_filter_kernel]
@@ -1453,6 +1456,7 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
         logging.warning('Webcam input: the framerate may vary. If results are filtered, Sports2D will use the average framerate as input.')
     else:
         cap, out_vid, cam_width, cam_height, fps = setup_video(video_file_path, save_vid, vid_output_path)
+        fps *= slowmo_factor
         start_time = get_start_time_ffmpeg(video_file_path)
         frame_range = [int((time_range[0]-start_time) * frame_rate), int((time_range[1]-start_time) * frame_rate)] if time_range else [0, int(cap.get(cv2.CAP_PROP_FRAME_COUNT))]
         frame_iterator = tqdm(range(*frame_range)) # use a progress bar
@@ -1466,7 +1470,7 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
     if load_trc:
         if not '_px' in str(load_trc): 
             logging.error(f'\n{load_trc} file needs to be in px, not in meters.')
-        logging.info(f'\nUsing a pose file instead of running pose tracking {load_trc}.')
+        logging.info(f'\nUsing a pose file instead of running pose estimation and tracking: {load_trc}.')
         # Load pose file in px
         Q_coords, _, _, keypoints_names, _ = read_trc(load_trc)
         keypoints_ids = [i for i in range(len(keypoints_names))]
@@ -1501,7 +1505,7 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
     frame_count = 0
     while cap.isOpened():
         # Skip to the starting frame
-        if frame_count < frame_range[0]:
+        if frame_count < frame_range[0] and not load_trc:
             cap.read()
             frame_count += 1
             continue
@@ -1528,6 +1532,8 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
 
             # Retrieve pose or Estimate pose and track people
             if load_trc: 
+                if frame_nb >= len(keypoints_all):
+                    break
                 keypoints = keypoints_all[frame_nb]
                 scores = scores_all[frame_nb]
             else: 
@@ -1545,18 +1551,23 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
             valid_X, valid_Y, valid_scores = [], [], []
             valid_X_flipped, valid_angles = [], []
             for person_idx in range(len(keypoints)):
-                # Retrieve keypoints and scores for the person, remove low-confidence keypoints
-                person_X, person_Y = np.where(scores[person_idx][:, np.newaxis] < keypoint_likelihood_threshold, np.nan, keypoints[person_idx]).T
-                person_scores = np.where(scores[person_idx] < keypoint_likelihood_threshold, np.nan, scores[person_idx])
-
-                # Skip person if the fraction of valid detected keypoints is too low
-                enough_good_keypoints = len(person_scores[~np.isnan(person_scores)]) >= len(person_scores) * keypoint_number_threshold
-                scores_of_good_keypoints = person_scores[~np.isnan(person_scores)]
-                average_score_of_remaining_keypoints_is_enough = (np.nanmean(scores_of_good_keypoints) if len(scores_of_good_keypoints)>0 else 0) >= average_likelihood_threshold
-                if not enough_good_keypoints or not average_score_of_remaining_keypoints_is_enough:
-                    person_X = np.full_like(person_X, np.nan)
-                    person_Y = np.full_like(person_Y, np.nan)
-                    person_scores = np.full_like(person_scores, np.nan)
+                if load_trc:
+                    person_X = keypoints[person_idx][:,0]
+                    person_Y = keypoints[person_idx][:,1]
+                    person_scores = scores[person_idx]
+                else:
+                    # Retrieve keypoints and scores for the person, remove low-confidence keypoints
+                    person_X, person_Y = np.where(scores[person_idx][:, np.newaxis] < keypoint_likelihood_threshold, np.nan, keypoints[person_idx]).T
+                    person_scores = np.where(scores[person_idx] < keypoint_likelihood_threshold, np.nan, scores[person_idx])
+
+                    # Skip person if the fraction of valid detected keypoints is too low
+                    enough_good_keypoints = len(person_scores[~np.isnan(person_scores)]) >= len(person_scores) * keypoint_number_threshold
+                    scores_of_good_keypoints = person_scores[~np.isnan(person_scores)]
+                    average_score_of_remaining_keypoints_is_enough = (np.nanmean(scores_of_good_keypoints) if len(scores_of_good_keypoints)>0 else 0) >= average_likelihood_threshold
+                    if not enough_good_keypoints or not average_score_of_remaining_keypoints_is_enough:
+                        person_X = np.full_like(person_X, np.nan)
+                        person_Y = np.full_like(person_Y, np.nan)
+                        person_scores = np.full_like(person_scores, np.nan)
                 valid_X.append(person_X)
                 valid_Y.append(person_Y)
                 valid_scores.append(person_scores)
@@ -1632,9 +1643,10 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
     all_frames_scores = make_homogeneous(all_frames_scores)
 
     frame_range = [0,frame_count] if video_file == 'webcam' else frame_range
-    all_frames_time = pd.Series(np.linspace(frame_range[0]/fps/slowmo_factor, frame_range[1]/fps/slowmo_factor, frame_count+1), name='time')
+    all_frames_time = pd.Series(np.linspace(frame_range[0]/fps, frame_range[1]/fps, frame_count+1), name='time')
     if not multiperson:
-        detected_persons = [get_personID_with_highest_scores(all_frames_scores)]
+        calib_on_person_id = get_personID_with_highest_scores(all_frames_scores)
+        detected_persons = [calib_on_person_id]
     else:
         detected_persons = range(all_frames_X_homog.shape[1])
 
@@ -1685,15 +1697,13 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
                 else:
                     filter_type = filter_options[1]
                     if filter_type == 'butterworth':
+                        cutoff = filter_options[3]
                         if video_file == 'webcam':
-                            cutoff = filter_options[3]
                             if cutoff / (fps / 2) >= 1:
                                 cutoff_old = cutoff
                                 cutoff = fps/(2+0.001)
                                 args = f'\n{cutoff_old:.1f} Hz cut-off framerate too large for a real-time framerate of {fps:.1f} Hz. Using a cut-off framerate of {cutoff:.1f} Hz instead.'
                                 filter_options[3] = cutoff
-                        else: 
-                            args = ''
                         args = f'Butterworth filter, {filter_options[2]}th order, {filter_options[3]} Hz.'
                         filter_options[4] = fps
                     if filter_type == 'gaussian':
@@ -1730,38 +1740,44 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
 
             else:
                 # Compute calibration parameters
+                if not multiperson: 
+                    selected_person_id = calib_on_person_id
+                    calib_on_person_id = 0
                 height_px = compute_height(trc_data[calib_on_person_id].iloc[:,1:], keypoints_names,
                                             fastest_frames_to_remove_percent=fastest_frames_to_remove_percent, close_to_zero_speed=close_to_zero_speed_px, large_hip_knee_angles=large_hip_knee_angles, trimmed_extrema_percent=trimmed_extrema_percent)
 
                 if floor_angle == 'auto' or xy_origin == 'auto':
                     # estimated from the line formed by the toes when they are on the ground (where speed = 0)
-                    floor_angle_estim, xy_origin_estim = compute_floor_line(trc_data[calib_on_person_id], keypoint_names=['LBigToe', 'RBigToe'])
+                    toe_speed_below = 1 # m/s (below which the foot is considered to be stationary)
+                    px_per_m = height_px/person_height_m
+                    toe_speed_below_px_frame = toe_speed_below * px_per_m / fps
+                    floor_angle_estim, xy_origin_estim = compute_floor_line(trc_data[calib_on_person_id], keypoint_names=['LBigToe', 'RBigToe'], toe_speed_below=toe_speed_below_px_frame)
                 if not floor_angle == 'auto':
                     floor_angle_estim = floor_angle
-                floor_angle_estim = -floor_angle_estim # Y points downwards
                 if xy_origin == 'auto':
                     cx, cy = xy_origin_estim
                 else:
                     cx, cy = xy_origin
                 logging.info(f'Using height of person #{calib_on_person_id} ({person_height_m}m) to convert coordinates in meters. '
-                             f'Floor angle: {np.degrees(-floor_angle_estim) if not floor_angle=="auto" else f"auto (estimation: {round(np.degrees(-floor_angle_estim),2)}°)"}, '
-                             f'xy_origin: {xy_origin if not xy_origin=="auto" else f"auto (estimation: {[round(c,2) for c in xy_origin_estim]})"}.')
+                             f'Floor angle: {np.degrees(floor_angle_estim) if not floor_angle=="auto" else f"auto (estimation: {round(np.degrees(floor_angle_estim),2)}°)"}, '
+                             f'xy_origin: {xy_origin if not xy_origin=="auto" else f"auto (estimation: {[round(c) for c in xy_origin_estim]})"}.')
 
             # Coordinates in m
             for i in range(len(trc_data)):
                 if not np.array(trc_data[i].iloc[:,1:] ==0).all():
-                    trc_data_m_i = pd.concat([convert_px_to_meters(trc_data[i][kpt_name], person_height_m, height_px, cx, cy, floor_angle_estim) for kpt_name in keypoints_names], axis=1)
+                    trc_data_m_i = pd.concat([convert_px_to_meters(trc_data[i][kpt_name], person_height_m, height_px, cx, cy, -floor_angle_estim) for kpt_name in keypoints_names], axis=1)
                     trc_data_m_i.insert(0, 't', all_frames_time)
-                    trc_data_unfiltered_m_i = pd.concat([convert_px_to_meters(trc_data_unfiltered[i][kpt_name], person_height_m, height_px, cx, cy, floor_angle_estim) for kpt_name in keypoints_names], axis=1)
+                    trc_data_unfiltered_m_i = pd.concat([convert_px_to_meters(trc_data_unfiltered[i][kpt_name], person_height_m, height_px, cx, cy, -floor_angle_estim) for kpt_name in keypoints_names], axis=1)
                     trc_data_unfiltered_m_i.insert(0, 't', all_frames_time)
 
                     if to_meters and show_plots:
                         pose_plots(trc_data_unfiltered_m_i, trc_data_m_i, i)
                     
                     # Write to trc file
-                    pose_path_person_m_i = (pose_output_path.parent / (pose_output_path_m.stem + f'_person{i:02d}.trc'))
+                    idx_path = selected_person_id if not multiperson and not calib_file else i
+                    pose_path_person_m_i = (pose_output_path.parent / (pose_output_path_m.stem + f'_person{idx_path:02d}.trc'))
                     make_trc_with_trc_data(trc_data_m_i, pose_path_person_m_i)
-                    logging.info(f'Person {i}: Pose in meters saved to {pose_path_person_m_i.resolve()}.')
+                    logging.info(f'Person {idx_path}: Pose in meters saved to {pose_path_person_m_i.resolve()}.')
                     
                 
             
@@ -1815,6 +1831,13 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
     if save_angles and calculate_angles:
         logging.info('\nPost-processing angles:')
         all_frames_angles = make_homogeneous(all_frames_angles)
+        
+        # unwrap angles
+        # all_frames_angles = np.unwrap(all_frames_angles, axis=0, period=180) # This give all nan values -> need to mask nans
+        for i in range(all_frames_angles.shape[1]):  # for each person
+            for j in range(all_frames_angles.shape[2]):  # for each angle
+                valid_mask = ~np.isnan(all_frames_angles[:, i, j])
+                all_frames_angles[valid_mask, i, j] = np.unwrap(all_frames_angles[valid_mask, i, j], period=180)
 
         # Process angles for each person
         for i in detected_persons:
@@ -1846,16 +1869,14 @@ def process_fun(config_dict, video_file, time_range, frame_rate, result_dir):
                 else:
                     filter_type = filter_options[1]
                     if filter_type == 'butterworth':
+                        cutoff = filter_options[3]
                         if video_file == 'webcam':
-                            cutoff = filter_options[3]
                             if cutoff / (fps / 2) >= 1:
                                 cutoff_old = cutoff
                                 cutoff = fps/(2+0.001)
                                 args = f'\n{cutoff_old:.1f} Hz cut-off framerate too large for a real-time framerate of {fps:.1f} Hz. Using a cut-off framerate of {cutoff:.1f} Hz instead.'
                                 filter_options[3] = cutoff
-                        else: 
-                            args = ''
-                        args = f'Butterworth filter, {filter_options[2]}th order, {filter_options[3]} Hz. ' + args
+                        args = f'Butterworth filter, {filter_options[2]}th order, {filter_options[3]} Hz.'
                         filter_options[4] = fps
                     if filter_type == 'gaussian':
                         args = f'Gaussian filter, Sigma kernel {filter_options[5]}.'

{sports2d-0.5.4 → sports2d-0.5.6}/setup.cfg

@@ -1,6 +1,6 @@
 [metadata]
 name = sports2d
-version = 0.5.4
+version = 0.5.6
 author = David Pagnon
 author_email = contact@david-pagnon.com
 description = Detect pose and compute 2D joint angles from a video.

{sports2d-0.5.4 → sports2d-0.5.6}/sports2d.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sports2d
-Version: 0.5.4
+Version: 0.5.6
 Summary: Detect pose and compute 2D joint angles from a video.
 Home-page: https://github.com/davidpagnon/Sports2D
 Author: David Pagnon
@@ -233,9 +233,10 @@ Note that it does not take distortions into account, and that it will be less ac
 ### Too slow for you?
 
 **Quick fixes:**
-- Use `--multiperson false`: Can be used if one single person is present in the video. Otherwise, persons' IDs may be mixed up.
+- Use ` --save_vid false --save_img false --show_realtime_results false`: Will not save images or videos, and will not display the results in real time. 
 - Use `--mode lightweight`: Will use a lighter version of RTMPose, which is faster but less accurate.
 - Use `--det_frequency 50`: Will detect poses only every 50 frames, and track keypoints in between, which is faster.
+- Use `--multiperson false`: Can be used if one single person is present in the video. Otherwise, persons' IDs may be mixed up.
 - Use `--load_trc <path_to_file_px.trc>`: Will use pose estimation results from a file. Useful if you want to use different parameters for pixel to meter conversion or angle calculation without running detection and pose estimation all over.
 
 <br>