Bugfix for empty preds or target in iou scores (#2806)

Co-authored-by: Jirka Borovec <[email protected]>
Co-authored-by: Jirka B <[email protected]>

(cherry picked from commit fbc7877)

CHANGELOG.md

@@ -6,6 +6,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 **Note: we move fast, but still we preserve 0.1 version (one feature release) back compatibility.**
 
+---
+
+## [UnReleased] - 2024-MM-DD
+
+### Fixed
+
+- Fixed iou scores in detection for either empty predictions/targets leading to wrong scores ([#2805](https://github.com/Lightning-AI/torchmetrics/pull/2805))
+
+
 ---
 
 ## [1.5.1] - 2024-10-22

src/torchmetrics/detection/iou.py

@@ -182,14 +182,17 @@ def update(self, preds: List[Dict[str, Tensor]], target: List[Dict[str, Tensor]]
         """Update state with predictions and targets."""
         _input_validator(preds, target, ignore_score=True)
 
-        for p, t in zip(preds, target):
-            det_boxes = self._get_safe_item_values(p["boxes"])
-            gt_boxes = self._get_safe_item_values(t["boxes"])
-            self.groundtruth_labels.append(t["labels"])
+        for p_i, t_i in zip(preds, target):
+            det_boxes = self._get_safe_item_values(p_i["boxes"])
+            gt_boxes = self._get_safe_item_values(t_i["boxes"])
+            self.groundtruth_labels.append(t_i["labels"])
 
             iou_matrix = self._iou_update_fn(det_boxes, gt_boxes, self.iou_threshold, self._invalid_val)  # N x M
             if self.respect_labels:
-                label_eq = p["labels"].unsqueeze(1) == t["labels"].unsqueeze(0)  # N x M
+                if det_boxes.numel() > 0 and gt_boxes.numel() > 0:
+                    label_eq = p_i["labels"].unsqueeze(1) == t_i["labels"].unsqueeze(0)  # N x M
+                else:
+                    label_eq = torch.eye(iou_matrix.shape[0], dtype=bool, device=iou_matrix.device)  # type: ignore[call-overload]
                 iou_matrix[~label_eq] = self._invalid_val
             self.iou_matrix.append(iou_matrix)
 

src/torchmetrics/functional/detection/ciou.py

@@ -31,6 +31,11 @@ def _ciou_update(
 
     from torchvision.ops import complete_box_iou
 
+    if preds.numel() == 0:  # if no boxes are predicted
+        return torch.zeros(target.shape[0], target.shape[0], device=target.device, dtype=torch.float32)
+    if target.numel() == 0:  # if no boxes are true
+        return torch.zeros(preds.shape[0], preds.shape[0], device=preds.device, dtype=torch.float32)
+
     iou = complete_box_iou(preds, target)
     if iou_threshold is not None:
         iou[iou < iou_threshold] = replacement_val

src/torchmetrics/functional/detection/diou.py

@@ -31,6 +31,11 @@ def _diou_update(
 
     from torchvision.ops import distance_box_iou
 
+    if preds.numel() == 0:  # if no boxes are predicted
+        return torch.zeros(target.shape[0], target.shape[0], device=target.device, dtype=torch.float32)
+    if target.numel() == 0:  # if no boxes are true
+        return torch.zeros(preds.shape[0], preds.shape[0], device=preds.device, dtype=torch.float32)
+
     iou = distance_box_iou(preds, target)
     if iou_threshold is not None:
         iou[iou < iou_threshold] = replacement_val

src/torchmetrics/functional/detection/giou.py

@@ -31,6 +31,11 @@ def _giou_update(
 
     from torchvision.ops import generalized_box_iou
 
+    if preds.numel() == 0:  # if no boxes are predicted
+        return torch.zeros(target.shape[0], target.shape[0], device=target.device, dtype=torch.float32)
+    if target.numel() == 0:  # if no boxes are true
+        return torch.zeros(preds.shape[0], preds.shape[0], device=preds.device, dtype=torch.float32)
+
     iou = generalized_box_iou(preds, target)
     if iou_threshold is not None:
         iou[iou < iou_threshold] = replacement_val

src/torchmetrics/functional/detection/iou.py

@@ -32,6 +32,11 @@ def _iou_update(
 
     from torchvision.ops import box_iou
 
+    if preds.numel() == 0:  # if no boxes are predicted
+        return torch.zeros(target.shape[0], target.shape[0], device=target.device, dtype=torch.float32)
+    if target.numel() == 0:  # if no boxes are true
+        return torch.zeros(preds.shape[0], preds.shape[0], device=preds.device, dtype=torch.float32)
+
     iou = box_iou(preds, target)
     if iou_threshold is not None:
         iou[iou < iou_threshold] = replacement_val

tests/unittests/detection/test_intersection.py

@@ -355,6 +355,43 @@ def test_corner_case_only_one_empty_prediction(self, class_metric, functional_me
         for val in res.values():
             assert val == torch.tensor(0.0)
 
+    def test_empty_preds_and_target(self, class_metric, functional_metric, reference_metric):
+        """Check that for either empty preds and targets that the metric returns 0 in these cases before averaging."""
+        x = [
+            {
+                "boxes": torch.empty(size=(0, 4), dtype=torch.float32),
+                "labels": torch.tensor([], dtype=torch.long),
+            },
+            {
+                "boxes": torch.FloatTensor([[0.1, 0.1, 0.2, 0.2], [0.3, 0.3, 0.4, 0.4]]),
+                "labels": torch.LongTensor([1, 2]),
+            },
+        ]
+
+        y = [
+            {
+                "boxes": torch.FloatTensor([[0.1, 0.1, 0.2, 0.2], [0.3, 0.3, 0.4, 0.4]]),
+                "labels": torch.LongTensor([1, 2]),
+                "scores": torch.FloatTensor([0.9, 0.8]),
+            },
+            {
+                "boxes": torch.FloatTensor([[0.1, 0.1, 0.2, 0.2], [0.3, 0.3, 0.4, 0.4]]),
+                "labels": torch.LongTensor([1, 2]),
+                "scores": torch.FloatTensor([0.9, 0.8]),
+            },
+        ]
+        metric = class_metric()
+        metric.update(x, y)
+        res = metric.compute()
+        for val in res.values():
+            assert val == torch.tensor(0.5)
+
+        metric = class_metric()
+        metric.update(y, x)
+        res = metric.compute()
+        for val in res.values():
+            assert val == torch.tensor(0.5)
+
 
 def test_corner_case():
     """See issue: https://github.com/Lightning-AI/torchmetrics/issues/1921."""