perf: majority vote for best label and average score

biotrack/track.py

@@ -1,6 +1,8 @@
 # biotrack, Apache-2.0 license
 # Filename: biotrack/tracker/track.py
 # Description:  Basic track object to contain and update tracks
+from collections import Counter
+
 from biotrack.logger import info, debug
 import numpy as np
 
@@ -19,6 +21,7 @@ def __init__(self, track_id: int, label: str, pt: np.array, emb: np.array, frame
         self.box = {frame: box}
         self.emb = emb
         self.best_label = label
+        self.best_score = score
         self.start_frame = frame
         self.last_updated_frame = frame
         self.x_scale = x_scale
@@ -67,10 +70,9 @@ def get_best(self, rescale=True) -> (int, np.array, str, np.array, float):
             frame_num = self.last_updated_frame
             box = self.box[frame_num]
             pt = self.pt[frame_num]
-        max_score = max(self.score.values())
         if rescale:
             pt, box = self.rescale(pt, box)
-        return frame_num, pt, self.best_label, box, max_score
+        return frame_num, pt, self.best_label, box, self.best_score
 
     def get(self, frame_num: int, rescale=True) -> (np.array, str, np.array, float):
         if frame_num not in self.pt.keys():
@@ -118,20 +120,41 @@ def update(self, label: str, pt: np.array, emb: np.array, frame_num: int, box:np
             self.emb = emb
         self.last_updated_frame = frame_num
 
-        # Update the best_label with that of the highest scoring label
-        # Only choose scores that are greater than 0
-        valid_scores = {k: v for k, v in self.score.items() if v > 0}
-        if len(valid_scores) == 0:
-            self.best_label = "marine organism"
+        # Update the best_label with that which occurs the most that has a score > 0. This is a simple majority vote
+        data = [(pred, score) for pred, score in zip(self.label.values(), self.score.values()) if float(score) > 0.]
+
+        if len(data) > 0:
+            p, s = zip(*data)
+            model_predictions = list(p)
+            model_scores = list(s)
+
+            # Count occurrences of each prediction in the top lists
+            counter = Counter(model_predictions)
+
+            majority_count = (len(data) // 2) + 1
+
+            majority_predictions = [pred for pred, count in counter.items() if count >= majority_count]
+
+            # If there are no majority predictions
+            if len(majority_predictions) == 0:
+                # Pick the prediction with the highest score
+                # best_pred, max_score = max_score_p(model_predictions, model_scores)
+                self.best_label = "marine organism"
+                self.best_score = 0.0
+            else:
+                self.best_label = majority_predictions[0]
+                best_score = 0.0
+                num_majority = 0
+                # Sum all the scores for the majority predictions
+                for pred, score in data:
+                    if pred in majority_predictions:
+                        best_score += float(score)
+                        num_majority += 1
+                self.best_score /= num_majority
         else:
-            max_score = max(valid_scores.values())
-            for key, value in valid_scores.items():
-                if value == max_score:
-                    max_frame = key
-                    break
-            self.best_label = self.label[max_frame]
-
-        pts_pretty = [f"{pt[0]:.2f},{pt[1]:.2f},{label}" for pt, label in zip(self.pt.values(), self.label.values())]
-        best_label = max(set(self.label.values()), key=list(self.label.values()).count)
+            self.best_label = "marine organism"
+            self.best_score = 0.0
+
+        pts_pretty = [f"{pt[0]:.2f},{pt[1]:.2f},{label},{score}" for pt, label, score in zip(self.pt.values(), self.label.values(), self.score.values())]
         total_frames = len(self.pt)
-        info(f"Updating tracker {self.id} total_frames {total_frames} updated start {self.start_frame} to {self.last_updated_frame} {pts_pretty} with label {best_label}")
+        info(f"Updating tracker {self.id} total_frames {total_frames} updated start {self.start_frame} to {self.last_updated_frame} {pts_pretty} with label {self.best_label}, score {self.best_score}")