using inverse distance as matching method for points

weecology · Jan 9, 2025 · c8e4d0f · c8e4d0f
1 parent c374375
commit c8e4d0f
Show file tree

Hide file tree

Showing 4 changed files with 161 additions and 447 deletions.
diff --git a/docs/examples/baseline_boxes.ipynb b/docs/examples/baseline_boxes.ipynb
diff --git a/docs/examples/baseline_points.ipynb b/docs/examples/baseline_points.ipynb
diff --git a/milliontrees/common/metrics/all_metrics.py b/milliontrees/common/metrics/all_metrics.py
@@ -461,7 +461,7 @@ class KeypointAccuracy(ElementwiseMetric):
     """Given a specific Intersection over union threshold, determine the
     accuracy achieved for a one-class detector."""
 
-    def __init__(self, distance_threshold=5, score_threshold=5, name=None, geometry_name="y"):
+    def __init__(self, distance_threshold=0.1, score_threshold=0.1, name=None, geometry_name="y"):
         self.distance_threshold = distance_threshold
         self.score_threshold = score_threshold
         self.geometry_name = geometry_name
@@ -478,13 +478,18 @@ def _compute_element_wise(self, y_pred, y_true):
 
             gt_boxes = gt[self.geometry_name]
             pred_boxes = target_boxes[target_scores > self.score_threshold]
-            det_accuracy = torch.mean(torch.stack([ self._accuracy(gt_boxes,pred_boxes,iou_thr) for iou_thr in np.arange(0.5,0.51,0.05)]))
+            det_accuracy = self._accuracy(gt_boxes,pred_boxes,self.distance_threshold)
             batch_results.append(det_accuracy)
 
         return torch.tensor(batch_results)
 
-    def _point_iou(self, src_keypoints, pred_keypoints):
-        return torch.cdist(src_keypoints, pred_keypoints, p=2)
+    def _point_nearness(self, src_keypoints, pred_keypoints):
+        distance = torch.cdist(src_keypoints, pred_keypoints, p=2)
+
+        # Inverson of distance to get relative distance
+        relative_distance = 1/distance
+
+        return relative_distance
 
     def _accuracy(self, src_keypoints, pred_keypoints, distance_threshold):
         total_gt = len(src_keypoints)
@@ -494,7 +499,7 @@ def _accuracy(self, src_keypoints, pred_keypoints, distance_threshold):
             matcher = Matcher(distance_threshold,
                               distance_threshold,
                               allow_low_quality_matches=False)
-            match_quality_matrix = self._point_iou(src_keypoints, pred_keypoints)
+            match_quality_matrix = self._point_nearness(src_keypoints, pred_keypoints)
             results = matcher(match_quality_matrix)
             true_positive = torch.count_nonzero(results.unique() != -1)
             matched_elements = results[results > -1]

diff --git a/milliontrees/datasets/TreePoints.py b/milliontrees/datasets/TreePoints.py
@@ -48,10 +48,12 @@ def __init__(self,
                  root_dir='data',
                  download=False,
                  split_scheme='official',
-                 geometry_name='y'):
+                 geometry_name='y',
+                 distance_threshold=0.1):
         self._version = version
         self._split_scheme = split_scheme
         self.geometry_name = geometry_name
+        self.distance_threshold = distance_threshold
 
         if self._split_scheme not in ['official', 'random']:
             raise ValueError(
@@ -120,7 +122,7 @@ def __init__(self,
         self._metadata_array = torch.tensor(unique_sources.values.astype('int'))
         self._metadata_fields = ['filename_id','source_id']
 
-        self._metric = KeypointAccuracy()
+        self._metric = KeypointAccuracy(distance_threshold=distance_threshold)
         self._collate = TreePointsDataset._collate_fn
 
         # eval grouper