Add option to fail if preproc is ambiguous

Author: Jack Valmadre

trackeval/datasets/mot_challenge_2d_box.py

@@ -41,6 +41,7 @@ def get_default_dataset_config():
             'MIN_VIS': -1.0,  # Set to non-negative number to filter ground-truth by visibility.
             'DROP_EXTRA_TIMESTEPS': False,  # Silently ignore extra timesteps.
             'DROP_OTHER_CLASSES': False,  # Silently ignore predictions for other classes.
+            'PREPROC_REQUIRE_UNIQUE': False,  # Raise exception for identical boxes during preprocessing.
         }
         return default_config
 
@@ -401,6 +402,20 @@ def get_preprocessed_seq_data(self, raw_data, cls):
                                     ~(gt_vis[match_rows] >= self.config['MIN_VIS']))
                 to_remove_tracker = match_cols[remove_mask]
 
+                if self.config['PREPROC_REQUIRE_UNIQUE']:
+                    # Check whether predictions to be removed were non-unique.
+                    identical_rows, identical_cols = np.where(np.any(
+                            ((matching_scores[:, :, np.newaxis] == matching_scores[:, np.newaxis, :]) &
+                             np.logical_not(np.eye(matching_scores.shape[1], dtype=np.bool)) &
+                             (matching_scores[:, :, np.newaxis] != 0)),
+                            axis=-1))
+                    if np.size(identical_rows):  # np.any(np.in1d(to_remove_tracker, identical_cols))
+                        raise TrackEvalException(
+                                'Multiple predicted tracks with identical similarity to ground-truth. '
+                                'Preprocessing cannot determine which to keep. '
+                                'The following pairs were indistinguishable in timestep ' + str(t) + ': ' +
+                                ' '.join(map(str, zip(gt_ids[identical_rows], tracker_ids[identical_cols]))))
+
             # Apply preprocessing to remove all unwanted tracker dets.
             data['tracker_ids'][t] = np.delete(tracker_ids, to_remove_tracker, axis=0)
             data['tracker_dets'][t] = np.delete(tracker_dets, to_remove_tracker, axis=0)