Merge branch 'main' into check_empty_subgraph

src/traccuracy/matchers/_iou.py

@@ -30,7 +30,10 @@ def _match_nodes(gt, res, threshold=0.5, one_to_one=False):
         gtcells (np arr): Array of overlapping ids in the gt frame.
         rescells (np arr): Array of overlapping ids in the res frame.
     """
-    iou = np.zeros((np.max(gt) + 1, np.max(res) + 1))
+    if threshold == 0.0 and not one_to_one:
+        raise ValueError("Threshold of 0 is not valid unless one_to_one is True")
+    # casting to int to avoid issue #152 (result is float with numpy<2, dtype=uint64)
+    iou = np.zeros((int(np.max(gt) + 1), int(np.max(res) + 1)))
 
     overlapping_gt_labels, overlapping_res_labels, _ = get_labels_with_overlap(gt, res)
 
@@ -39,12 +42,14 @@ def _match_nodes(gt, res, threshold=0.5, one_to_one=False):
         iou_res_idx = overlapping_res_labels[index]
         intersection = np.logical_and(gt == iou_gt_idx, res == iou_res_idx)
         union = np.logical_or(gt == iou_gt_idx, res == iou_res_idx)
-        iou[iou_gt_idx, iou_res_idx] = intersection.sum() / union.sum()
+        iou_value = intersection.sum() / union.sum()
+        if iou_value >= threshold:
+            iou[iou_gt_idx, iou_res_idx] = iou_value
 
     if one_to_one:
         pairs = _one_to_one_assignment(iou)
     else:
-        pairs = np.where(iou >= threshold)
+        pairs = np.where(iou)
 
     # Catch the case where there are no overlaps
     if len(pairs) < 2:
@@ -76,6 +81,8 @@ def _one_to_one_assignment(iou, unmapped_cost=4):
 
     # Assign 1 - iou to top left and bottom right
     cost = 1 - iou[1:, 1:]
+    # increase the cost for those with no IOU to higher than the unmapped cost
+    cost[cost == 1] = unmapped_cost + 1
     matrix[:n0, :n1] = cost
     matrix[n_obj - n1 :, n_obj - n0 :] = cost.T
 

tests/matchers/test_iou.py

@@ -34,6 +34,22 @@ def get_two_to_one(w, h, imw, imh):
     return merge.astype("int"), split.astype("int")
 
 
+def get_no_overlap(imw, imh):
+    """two non-overlapping segmentations that start at high number labels
+
+    Mapping []
+
+    """
+
+    im1 = np.zeros((imw, imh))
+    im1[0:2, 0:2] = 100
+
+    im2 = np.zeros((imw, imh))
+    im2[4:6, 4:6] = 222
+
+    return im1.astype("int"), im2.astype("int")
+
+
 def test__match_nodes():
     # creat dummy image to test against
     num_labels = 5
@@ -62,6 +78,48 @@ def test__match_nodes():
     # Check that only one of the merge matches is present
     assert ((2, 4) in matches) != ((2, 5) in matches)
 
+    with pytest.raises(ValueError):
+        # Test that threshold 0 is not valid when not one-to-one
+        gtcells, rescells = _match_nodes(im1, im2, threshold=0.0)
+
+
+def test__match_nodes_threshold():
+    im1, im2 = get_two_to_one(10, 10, 30, 30)
+    # Test high threshold
+    gtcells, rescells = _match_nodes(im1, im2, threshold=1)
+    # Create match tuples
+    matches = list(zip(gtcells, rescells))
+    # Check that nothing is matched
+    assert len(matches) == 1
+
+    # Test for high threshold and one to one
+    gtcells, rescells = _match_nodes(im1, im2, threshold=0.7, one_to_one=True)
+    # Create match tuples
+    matches = list(zip(gtcells, rescells))
+    # Check that nothing is matched
+    assert len(matches) == 1
+
+
+def test__match_nodes_non_sequential():
+    # test when the segmentation ids are high numbers (the lower numbers should never appear)
+
+    im1, im2 = get_no_overlap(30, 30)
+
+    # Test that phantom segmentations are not matched
+    gtcells, rescells = _match_nodes(im1, im2, threshold=0.1)
+    # Create match tuples
+    matches = list(zip(gtcells, rescells))
+    # Check that nothing is matched
+    assert len(matches) == 0
+
+    # Test that with one-to-one, phantom segmentations are not matched,
+    # even with threshold 0
+    gtcells, rescells = _match_nodes(im1, im2, threshold=0.0, one_to_one=True)
+    # Create match tuples
+    matches = list(zip(gtcells, rescells))
+    # Check that nothing is matched
+    assert len(matches) == 0
+
 
 def test__construct_time_to_seg_id_map():
     # Test 2d data