Fix classwise computation in IoU metric (#1924)

Co-authored-by: Jirka Borovec <6035284+Borda@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
(cherry picked from commit eeb40e9)

Author: SkafteNicki

CHANGELOG.md

@@ -32,8 +32,12 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Fixed support for pixelwise MSE ([#2017](https://github.com/Lightning-AI/torchmetrics/pull/2017)
 
 
+- Fixed bug in detection intersection metrics when `class_metrics=True` resulting in wrong values ([#1924](https://github.com/Lightning-AI/torchmetrics/pull/1924))
+
+
 - Fixed missing attributes `higher_is_better`, `is_differentiable` for some metrics ([#2028](https://github.com/Lightning-AI/torchmetrics/pull/2028)
 
+
 ## [1.1.0] - 2023-08-22
 
 ### Added

Makefile

@@ -3,6 +3,7 @@
 export FREEZE_REQUIREMENTS=1
 # assume you have installed need packages
 export SPHINX_MOCK_REQUIREMENTS=1
+export SPHINX_FETCH_ASSETS=0
 
 clean:
 	# clean all temp runs

src/torchmetrics/detection/ciou.py

@@ -37,8 +37,6 @@ class CompleteIntersectionOverUnion(IntersectionOverUnion):
         - ``boxes`` (:class:`~torch.Tensor`): float tensor of shape ``(num_boxes, 4)`` containing ``num_boxes``
           detection boxes of the format specified in the constructor.
           By default, this method expects ``(xmin, ymin, xmax, ymax)`` in absolute image coordinates.
-        - ``scores`` (:class:`~torch.Tensor`): float tensor of shape ``(num_boxes)`` containing detection scores
-          for the boxes.
         - ``labels`` (:class:`~torch.Tensor`): integer tensor of shape ``(num_boxes)`` containing 0-indexed detection
           classes for the boxes.
 
@@ -48,14 +46,14 @@ class CompleteIntersectionOverUnion(IntersectionOverUnion):
         - ``boxes`` (:class:`~torch.Tensor`): float tensor of shape ``(num_boxes, 4)`` containing ``num_boxes`` ground
           truth boxes of the format specified in the constructor.
           By default, this method expects ``(xmin, ymin, xmax, ymax)`` in absolute image coordinates.
-        - ``labels`` (:class:`~torch.Tensor`): integer tensor of shape ``(num_boxes)`` containing 0-indexed ground truth
+        - ``labels`` (:class:`~torch.Tensor`): integer tensor of shape ``(num_boxes)`` containing 0-indexed detection
           classes for the boxes.
 
     As output of ``forward`` and ``compute`` the metric returns the following output:
 
     - ``ciou_dict``: A dictionary containing the following key-values:
 
-        - ciou: (:class:`~torch.Tensor`)
+        - ciou: (:class:`~torch.Tensor`) with overall ciou value over all classes and samples.
         - ciou/cl_{cl}: (:class:`~torch.Tensor`), if argument ``class_metrics=True``
 
     Args:
@@ -65,6 +63,9 @@ class CompleteIntersectionOverUnion(IntersectionOverUnion):
             Optional IoU thresholds for evaluation. If set to `None` the threshold is ignored.
         class_metrics:
             Option to enable per-class metrics for IoU. Has a performance impact.
+        respect_labels:
+            Ignore values from boxes that do not have the same label as the ground truth box. Else will compute Iou
+            between all pairs of boxes.
         kwargs:
             Additional keyword arguments, see :ref:`Metric kwargs` for more info.
 
@@ -86,7 +87,7 @@ class CompleteIntersectionOverUnion(IntersectionOverUnion):
         ... ]
         >>> metric = CompleteIntersectionOverUnion()
         >>> metric(preds, target)
-        {'ciou': tensor(-0.5694)}
+        {'ciou': tensor(0.8611)}
 
     Raises:
         ModuleNotFoundError:
@@ -105,14 +106,15 @@ def __init__(
         box_format: str = "xyxy",
         iou_threshold: Optional[float] = None,
         class_metrics: bool = False,
+        respect_labels: bool = True,
         **kwargs: Any,
     ) -> None:
         if not _TORCHVISION_GREATER_EQUAL_0_13:
             raise ModuleNotFoundError(
                 f"Metric `{self._iou_type.upper()}` requires that `torchvision` version 0.13.0 or newer is installed."
                 " Please install with `pip install torchvision>=0.13` or `pip install torchmetrics[detection]`."
             )
-        super().__init__(box_format, iou_threshold, class_metrics, **kwargs)
+        super().__init__(box_format, iou_threshold, class_metrics, respect_labels, **kwargs)
 
     @staticmethod
     def _iou_update_fn(*args: Any, **kwargs: Any) -> Tensor:

src/torchmetrics/detection/diou.py

@@ -37,8 +37,6 @@ class DistanceIntersectionOverUnion(IntersectionOverUnion):
         - ``boxes`` (:class:`~torch.Tensor`): float tensor of shape ``(num_boxes, 4)`` containing ``num_boxes``
           detection boxes of the format specified in the constructor.
           By default, this method expects ``(xmin, ymin, xmax, ymax)`` in absolute image coordinates.
-        - ``scores`` (:class:`~torch.Tensor`): float tensor of shape ``(num_boxes)`` containing detection scores
-          for the boxes.
         - ``labels`` (:class:`~torch.Tensor`): integer tensor of shape ``(num_boxes)`` containing 0-indexed detection
           classes for the boxes.
 
@@ -55,7 +53,7 @@ class DistanceIntersectionOverUnion(IntersectionOverUnion):
 
     - ``diou_dict``: A dictionary containing the following key-values:
 
-        - diou: (:class:`~torch.Tensor`)
+        - diou: (:class:`~torch.Tensor`) with overall diou value over all classes and samples.
         - diou/cl_{cl}: (:class:`~torch.Tensor`), if argument ``class_metrics=True``
 
     Args:
@@ -65,6 +63,9 @@ class DistanceIntersectionOverUnion(IntersectionOverUnion):
             Optional IoU thresholds for evaluation. If set to `None` the threshold is ignored.
         class_metrics:
             Option to enable per-class metrics for IoU. Has a performance impact.
+        respect_labels:
+            Ignore values from boxes that do not have the same label as the ground truth box. Else will compute Iou
+            between all pairs of boxes.
         kwargs:
             Additional keyword arguments, see :ref:`Metric kwargs` for more info.
 
@@ -86,7 +87,7 @@ class DistanceIntersectionOverUnion(IntersectionOverUnion):
         ... ]
         >>> metric = DistanceIntersectionOverUnion()
         >>> metric(preds, target)
-        {'diou': tensor(-0.0694)}
+        {'diou': tensor(0.8611)}
 
     Raises:
         ModuleNotFoundError:
@@ -105,14 +106,15 @@ def __init__(
         box_format: str = "xyxy",
         iou_threshold: Optional[float] = None,
         class_metrics: bool = False,
+        respect_labels: bool = True,
         **kwargs: Any,
     ) -> None:
         if not _TORCHVISION_GREATER_EQUAL_0_13:
             raise ModuleNotFoundError(
                 f"Metric `{self._iou_type.upper()}` requires that `torchvision` version 0.13.0 or newer is installed."
                 " Please install with `pip install torchvision>=0.13` or `pip install torchmetrics[detection]`."
             )
-        super().__init__(box_format, iou_threshold, class_metrics, **kwargs)
+        super().__init__(box_format, iou_threshold, class_metrics, respect_labels, **kwargs)
 
     @staticmethod
     def _iou_update_fn(*args: Any, **kwargs: Any) -> Tensor:

src/torchmetrics/detection/giou.py

@@ -37,8 +37,6 @@ class GeneralizedIntersectionOverUnion(IntersectionOverUnion):
         - ``boxes`` (:class:`~torch.Tensor`): float tensor of shape ``(num_boxes, 4)`` containing ``num_boxes``
           detection boxes of the format specified in the constructor.
           By default, this method expects ``(xmin, ymin, xmax, ymax)`` in absolute image coordinates.
-        - ``scores`` (:class:`~torch.Tensor`): float tensor of shape ``(num_boxes)`` containing detection scores
-          for the boxes.
         - ``labels`` (:class:`~torch.Tensor`): integer tensor of shape ``(num_boxes)`` containing 0-indexed detection
           classes for the boxes.
 
@@ -55,7 +53,7 @@ class GeneralizedIntersectionOverUnion(IntersectionOverUnion):
 
     - ``giou_dict``: A dictionary containing the following key-values:
 
-        - giou: (:class:`~torch.Tensor`)
+        - giou: (:class:`~torch.Tensor`) with overall giou value over all classes and samples.
         - giou/cl_{cl}: (:class:`~torch.Tensor`), if argument ``class metrics=True``
 
     Args:
@@ -65,6 +63,9 @@ class GeneralizedIntersectionOverUnion(IntersectionOverUnion):
             Optional IoU thresholds for evaluation. If set to `None` the threshold is ignored.
         class_metrics:
             Option to enable per-class metrics for IoU. Has a performance impact.
+        respect_labels:
+            Ignore values from boxes that do not have the same label as the ground truth box. Else will compute Iou
+            between all pairs of boxes.
         kwargs:
             Additional keyword arguments, see :ref:`Metric kwargs` for more info.
 
@@ -86,7 +87,7 @@ class GeneralizedIntersectionOverUnion(IntersectionOverUnion):
         ... ]
         >>> metric = GeneralizedIntersectionOverUnion()
         >>> metric(preds, target)
-        {'giou': tensor(-0.0694)}
+        {'giou': tensor(0.8613)}
 
     Raises:
         ModuleNotFoundError:
@@ -105,9 +106,10 @@ def __init__(
         box_format: str = "xyxy",
         iou_threshold: Optional[float] = None,
         class_metrics: bool = False,
+        respect_labels: bool = True,
         **kwargs: Any,
     ) -> None:
-        super().__init__(box_format, iou_threshold, class_metrics, **kwargs)
+        super().__init__(box_format, iou_threshold, class_metrics, respect_labels, **kwargs)
 
     @staticmethod
     def _iou_update_fn(*args: Any, **kwargs: Any) -> Tensor:

src/torchmetrics/detection/helpers.py

@@ -20,6 +20,7 @@ def _input_validator(
     preds: Sequence[Dict[str, Tensor]],
     targets: Sequence[Dict[str, Tensor]],
     iou_type: Union[Literal["bbox", "segm"], Tuple[Literal["bbox", "segm"]]] = "bbox",
+    ignore_score: bool = False,
 ) -> None:
     """Ensure the correct input format of `preds` and `targets`."""
     if isinstance(iou_type, str):
@@ -39,7 +40,7 @@ def _input_validator(
             f"Expected argument `preds` and `target` to have the same length, but got {len(preds)} and {len(targets)}"
         )
 
-    for k in [*item_val_name, "scores", "labels"]:
+    for k in [*item_val_name, "labels"] + (["scores"] if not ignore_score else []):
         if any(k not in p for p in preds):
             raise ValueError(f"Expected all dicts in `preds` to contain the `{k}` key")
 
@@ -50,7 +51,7 @@ def _input_validator(
     for ivn in item_val_name:
         if any(type(pred[ivn]) is not Tensor for pred in preds):
             raise ValueError(f"Expected all {ivn} in `preds` to be of type Tensor")
-    if any(type(pred["scores"]) is not Tensor for pred in preds):
+    if not ignore_score and any(type(pred["scores"]) is not Tensor for pred in preds):
         raise ValueError("Expected all scores in `preds` to be of type Tensor")
     if any(type(pred["labels"]) is not Tensor for pred in preds):
         raise ValueError("Expected all labels in `preds` to be of type Tensor")
@@ -67,6 +68,8 @@ def _input_validator(
                     f"Input '{ivn}' and labels of sample {i} in targets have a"
                     f" different length (expected {item[ivn].size(0)} labels, got {item['labels'].size(0)})"
                 )
+    if ignore_score:
+        return
     for i, item in enumerate(preds):
         for ivn in item_val_name:
             if not (item[ivn].size(0) == item["labels"].size(0) == item["scores"].size(0)):