Added with_missing in Reconstructor3DSettings

point_tracker/reconstructor3d.py

@@ -19,6 +19,8 @@ class Reconstructor3DSettings(BaseSettings):
     Class for keeping track of the settings of Reconstructor3D.
     """
 
+    with_missing: bool = False  # Whether to try to reconstruct 3d obj with less than nb_cam number of 2d points
+
     threshold_mean_repro_err: float = 5.0
     threshold_nb_camera_views: int = 2  # Minimum number of camera views that will be used for the 3d reconstruction (when with_missing = True)
     threshold_distance_tracks: float = 0.02  # max distance (in meters) between two successive 3d points to set as a track
@@ -79,8 +81,7 @@ class Reconstructor3D:
 
         return Reconstructor3D(dlt_coefs)
 
-    def recon_objs(self, xs_pts: List[int], ys_pts: List[int], frames_coord: List[int],
-                   with_missing: bool = True, sort_by_rmse: bool = True) -> Tuple[any]:
+    def recon_objs(self, xs_pts: List[int], ys_pts: List[int], frames_coord: List[int], sort_by_rmse: bool = True) -> Tuple[any]:
         """
         Reconstruct 3d coordinates of objects from 2d coordinates of points for all frames
         args are vectors of vectors (e.g. np.array([[451, 123], [78], [147, 65, 78, 79]])) of size nb_cam*nb_obj
@@ -91,7 +92,6 @@ class Reconstructor3D:
             ys_pts: y coordinates in pixel of all pts for all frames for each camera
             frames_coord: corresponding frames of all pts for each camera (can have duplicates (several pts per frames))
 
-            with_missing: Whether to try to reconstruct 3d obj with less than nb_cam number of 2d points
             sort_by_rmse: Whether to sort the results by rmse (lower error first)
 
         Returns:
@@ -142,7 +142,7 @@ class Reconstructor3D:
             xs_pt, ys_pt = np.asarray(xs_pt), np.asarray(ys_pt)
 
             if xs_pt == []: continue
-            xs_obj, ys_obj, zs_obj, index_match_pts, rmses_obj = self.match_pts_and_recon_obj(xs_pt, ys_pt, with_missing)
+            xs_obj, ys_obj, zs_obj, index_match_pts, rmses_obj = self.match_pts_and_recon_obj(xs_pt, ys_pt)
             if xs_obj == []: continue
 
             index_pts = - np.ones(np.array(index_match_pts).shape)
@@ -177,14 +177,13 @@ class Reconstructor3D:
 
         return xs_objs, ys_objs, zs_objs, frames_objs, indexes_pts, rmses_objs
 
-    def match_pts_and_recon_obj(self, xs_pts, ys_pts, with_missing=True):
+    def match_pts_and_recon_obj(self, xs_pts, ys_pts):
         """
         Match 2d pts together (by minimizing rmse) and reconstruct 3d coordinates of objects
 
         Args:
             xs_pts: x coordinates in pixel of all pts for all camera
             ys_pts: y coordinates in pixel of all pts for all camera
-            with_missing: if True will also try to reconstruct 3d obj with less than nb_cam number of 2d points
 
         Returns:
             xs_obj: x coordinates for all objects (3d reconstructed and matched to minimized rmse)
@@ -194,7 +193,7 @@ class Reconstructor3D:
             rmses_obj: minimized rmse for all objects
         """
 
-        if with_missing:
+        if self.settings.with_missing:
             # Make matrix with indexes of cameras and with None when camera is missing (nan values in xs_px)
             # (add a row of missing cam to look at combination involving less than the max number of cameras)
             indexes_frames = [np.append(np.arange(0, np.array(xs).size), None) for xs in xs_pts]
@@ -242,9 +241,9 @@ class Reconstructor3D:
             repro_pts_cams_wo_missing = \
                 np.array([repro_pts_cams[j] for j, _ in enumerate(range(1, self.nb_cam + 1)) if index_match[j] != None])
 
-            # To correct the error reduction when we use less points than the max number of cams
-            correction_factor = count_missing * np.sqrt(2) if count_missing > 0 else 1
-            mean_repro_err = np.sqrt(np.mean(np.sum((repro_pts_cams_wo_missing - pts_cams) ** 2, 1))) * correction_factor
+            # To correct the error reduction when we use fewer points than the max number of cams
+            correction_factor = np.sqrt(1 + count_missing) if count_missing > 0 else 1
+            mean_repro_err = np.mean(np.sqrt(np.sum((repro_pts_cams_wo_missing - pts_cams) ** 2, 1))) * correction_factor
 
             if mean_repro_err < self.settings.threshold_mean_repro_err:
                 # Fill tracking results

process/dlc_postprocessing.py

@@ -493,7 +493,7 @@ def save_stroboscopic_images(image_sequence: MultiImageSequence, frames: List[in
 #                     frames_coord.append([frame, frame])
 #
 #                 xs_pose, ys_pose, zs_pose, frames_pose, indexes, rmses, = \
-#                     reconstructor.recon_objs(xs_coord, ys_coord, frames_coord, with_missing=False, sort_by_rmse=False)
+#                     reconstructor.recon_objs(xs_coord, ys_coord, frames_coord, sort_by_rmse=False)
 #
 #                 # if len(xs_pose) == 0:
 #                 #     xs_pose, ys_pose, zs_pose = [[np.nan, np.nan]], [[np.nan, np.nan]], [[np.nan, np.nan]]
@@ -626,7 +626,7 @@ def save_stroboscopic_images(image_sequence: MultiImageSequence, frames: List[in
 #                     frames_coord.append([frame])
 #
 #                 xs_pose, ys_pose, zs_pose, frames_pose, indexes, rmses, = \
-#                     reconstructor.recon_objs(xs_coord, ys_coord, frames_coord, with_missing=False)
+#                     reconstructor.recon_objs(xs_coord, ys_coord, frames_coord)
 #
 #                 if len(xs_pose) == 0:
 #                     skeleton3d[obj_name][label]['x'][i], skeleton3d[obj_name][label]['y'][i], \
@@ -786,7 +786,7 @@ def reverse_processes_2d_points(skeleton2d: Dict[str, any], settings: Settings,
 #                     frames_coord.append([frame])
 #
 #                 xs_pose, ys_pose, zs_pose, frames_pose, indexes, rmses, = \
-#                     reconstructor.recon_objs(xs_coord, ys_coord, frames_coord, with_missing=False)
+#                     reconstructor.recon_objs(xs_coord, ys_coord, frames_coord)
 #
 #                 if len(xs_pose) == 0:
 #                     skeleton3d[obj_name][label]['x'][i], skeleton3d[obj_name][label]['y'][i], \
@@ -871,7 +871,7 @@ def recon3d_dlc(skeleton2d: Dict[str, any], dlt_path: str,
                     frames_coord.append([frame])
 
                 xs_pose, ys_pose, zs_pose, frames_pose, indexes, rmses = \
-                    reconstructor.recon_objs(xs_coord, ys_coord, frames_coord, with_missing=True)
+                    reconstructor.recon_objs(xs_coord, ys_coord, frames_coord)
 
                 if len(xs_pose) == 0:
                     skeleton3d[obj_name][label]['x'][i], skeleton3d[obj_name][label]['y'][i], \

process/point_tracking.py

@@ -168,7 +168,7 @@ def reconstruct_3d_points(pts_dict: Dict[any, Dict[str, List[float]]], dlt_path:
     frames_coord = np.array([pts_dict[camn]['frame'] for camn in range(1, nb_cam + 1)])
 
     xs_pose, ys_pose, zs_pose, frames_pose, indexes, rmses, = \
-        reconstructor.recon_objs(xs_coord, ys_coord, frames_coord, with_missing=False)
+        reconstructor.recon_objs(xs_coord, ys_coord, frames_coord)
 
     save_names_list = [save_names[camn] for camn in range(1, nb_cam + 1)]
     save_paths_list = [save_paths[camn] for camn in range(1, nb_cam + 1)]
@@ -204,7 +204,7 @@ def reconstruct_3d_tracks(pts_dict: Dict[any, Dict[str, List[float]]], dlt_path:
     frames_coord = np.array([pts_dict[camn]['frame'] for camn in range(1, nb_cam + 1)])
 
     xs_pose, ys_pose, zs_pose, frames_pose, indexes_pts, _, = \
-        reconstructor.recon_objs(xs_coord, ys_coord, frames_coord, with_missing=False)
+        reconstructor.recon_objs(xs_coord, ys_coord, frames_coord)
 
     tracks_dict = reconstructor.recon_tracks(xs_pose, ys_pose, zs_pose, frames_pose, indexes_pts)