feat: Update LoudFE & VAFE to not keep adjacent features

Otherwise features tend to cluster

pipeline/feature_extractors.py

@@ -200,7 +200,7 @@ class LoudAudioFeatureExtractor(FeatureExtractor):
 
      teardown() is used to clean up temporary files created during setup (if specified by config)
     """
-    _CONFIG_DEFAULT_NUM_FEATURES = 5 # keep the top 5 loudnesses
+    _CONFIG_DEFAULT_NUM_FEATURES = 15 # keep the top 5 loudnesses
     _CONFIG_DEFAULT_MIN_DURATION = 5.00 # seconds
     def __init__(self, input_files=None, config=None,
                  num_features=_CONFIG_DEFAULT_NUM_FEATURES,
@@ -255,28 +255,33 @@ class LoudAudioFeatureExtractor(FeatureExtractor):
 
         return loudness_features
 
-    def _keep_num(self, features, num=_CONFIG_DEFAULT_NUM_FEATURES) -> list:
+    def _keep_num(self, features, num=_CONFIG_DEFAULT_NUM_FEATURES, margin=10.0) -> list:
         """Keep the top n features (default: 5)
 
         Approach:
          - for range in 0-n
            + expand the nth top feature to min duration
             (move start back by 0.5*min_duration, end forward by 0.5*min_duration)
-           + drop any features that are now in that feature's range
+           + drop any features that are now in that feature's range (plus margin)
          - return the top n features
 
         Each feature is a Feature object, with an Interval object
         """
+        keep_features = []
+        # ensure features are sorted by score
+        features = sorted(features, key=lambda x: x.score, reverse=True)
         for i in range(num):
+            current_feature = features.pop(0)
             # expand the feature to min_duration
-            features[i].interval.move_start(-0.5*self._min_duration, relative=True)
-            features[i].interval.move_end(0.5*self._min_duration, relative=True)
-            # drop any features that are now in that feature's range
+            current_feature.interval.move_start(-0.5*self._min_duration, relative=True)
+            current_feature.interval.move_end(0.5*self._min_duration, relative=True)
+            keep_features.append(current_feature)
+            # drop any features that are now in that feature's range (plus margin)
             features = [f for f in features if
-                        f.interval.start < features[i].interval.start or
-                        f.interval.end > features[i].interval.end]
+                        (f.interval.end < current_feature.interval.start-margin or
+                         f.interval.start > current_feature.interval.end+margin)]
 
-        return features[:num]
+        return keep_features
 
     def setup(self):
         """extract audio from video files to be processed by pyloudnorm
@@ -322,7 +327,7 @@ class VideoActivityFeatureExtractor(FeatureExtractor):
 
     #TODO: minimum duration -- consider whether to do here, or expand duration post-consolidation
     """
-    _CONFIG_DEFAULT_NUM_FEATURES = 5 # keep the top 5 activity moments
+    _CONFIG_DEFAULT_NUM_FEATURES = 15 # keep the top 5 activity moments
     _CONFIG_DEFAULT_MIN_DURATION = 5.00 # seconds
     def __init__(self, input_files=None, config=None,
                  num_features=_CONFIG_DEFAULT_NUM_FEATURES,
@@ -379,28 +384,34 @@ class VideoActivityFeatureExtractor(FeatureExtractor):
         scores = sorted(scores, key=lambda x: x[1], reverse=True)
         return scores[:int(len(scores) * (percent / 100))]
 
-    def _keep_num(self, features, num=_CONFIG_DEFAULT_NUM_FEATURES) -> list:
+    def _keep_num(self, features, num=_CONFIG_DEFAULT_NUM_FEATURES, margin=10.0) -> list:
         """Keep the top n features (default: 5)
 
         Approach:
          - for range in 0-n
            + expand the nth top feature to min duration
             (move start back by 0.5*min_duration, end forward by 0.5*min_duration)
-           + drop any features that are now in that feature's range
+           + drop any features that are now in that feature's range (plus margin)
          - return the top n features
 
         Each feature is a Feature object, with an Interval object
         """
+        keep_features = []
+        # ensure features are sorted by score
+        features = sorted(features, key=lambda x: x.score, reverse=True)
         for i in range(num):
+            current_feature = features.pop(0)
             # expand the feature to min_duration
-            features[i].interval.move_start(-0.5*self._min_duration, relative=True)
-            features[i].interval.move_end(0.5*self._min_duration, relative=True)
-            # drop any features that are now in that feature's range
+            current_feature.interval.move_start(-0.5*self._min_duration, relative=True)
+            current_feature.interval.move_end(0.5*self._min_duration, relative=True)
+            keep_features.append(current_feature)
+            # drop any features that are now in that feature's range (plus margin)
             features = [f for f in features if
-                        f.interval.start < features[i].interval.start or
-                        f.interval.end > features[i].interval.end]
+                        (f.interval.end < current_feature.interval.start-margin or
+                         f.interval.start > current_feature.interval.end+margin)]
+
+        return keep_features
 
-        return features[:num]
 
     def setup(self):
         pass