Raise error in MeanAveragePrecision if too little/many detection are provided for pycocotools backend

CHANGELOG.md

@@ -54,6 +54,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Fixed numerical stability issue in `UniversalImageQualityIndex` metric ([#2222](https://github.com/Lightning-AI/torchmetrics/pull/2222))
 
 
+- Fixed incompatibility for `MeanAveragePrecision` with `pycocotools` backend when too little `max_detection_thresholds` are provided ([#2219](https://github.com/Lightning-AI/torchmetrics/pull/2219))
+
 ## [1.2.0] - 2023-09-22
 
 ### Added

src/torchmetrics/detection/mean_ap.py

@@ -184,7 +184,8 @@ class MeanAveragePrecision(Metric):
             with step ``0.01``. Else provide a list of floats.
         max_detection_thresholds:
             Thresholds on max detections per image. If set to `None` will use thresholds ``[1, 10, 100]``.
-            Else, please provide a list of ints.
+            Else, please provide a list of ints. If the `pycocotools` backend is used then the list needs to have
+            length 3. If this is a problem, shift to `faster_coco_eval` which supports more detection thresholds.
         class_metrics:
             Option to enable per-class metrics for mAP and mAR_100. Has a performance impact that scales linearly with
             the number of classes in the dataset.
@@ -406,6 +407,11 @@ def __init__(
                 f"Expected argument `max_detection_thresholds` to either be `None` or a list of ints"
                 f" but got {max_detection_thresholds}"
             )
+        if max_detection_thresholds is not None and backend == "pycocotools" and len(max_detection_thresholds) == 3:
+            raise ValueError(
+                "When using `pycocotools` backend the number of max detection thresholds should be 3 else"
+                f"it will not work correctly with the backend. Got value {len(max_detection_thresholds)}."
+            )
         max_det_thr, _ = torch.sort(torch.tensor(max_detection_thresholds or [1, 10, 100], dtype=torch.int))
         self.max_detection_thresholds = max_det_thr.tolist()
 

tests/unittests/detection/test_map.py

@@ -854,3 +854,34 @@ def test_many_detection_thresholds(self, backend):
             assert round(res["map"].item(), 5) != 0.6
         else:
             assert round(res["map"].item(), 5) == 0.6
+
+    @pytest.mark.parametrize("max_detection_thresholds", [[1, 10], [1, 10, 50, 100]])
+    def test_with_more_and_less_detection_thresholds(self, max_detection_thresholds, backend):
+        """Test how metric is working when list of max detection thresholds is not 3.
+
+        This is a known limitation of the pycocotools where values are hardcoded to expect at least 3 elements
+        https://github.com/ppwwyyxx/cocoapi/blob/master/PythonAPI/pycocotools/cocoeval.py#L461
+
+        """
+        preds = [
+            {
+                "boxes": torch.tensor([[258.0, 41.0, 606.0, 285.0]]),
+                "scores": torch.tensor([0.536]),
+                "labels": torch.tensor([0]),
+            }
+        ]
+        target = [
+            {
+                "boxes": torch.tensor([[214.0, 41.0, 562.0, 285.0]]),
+                "labels": torch.tensor([0]),
+            }
+        ]
+
+        if backend == "pycocotools":
+            with pytest.raises(
+                ValueError, match="When using `pycocotools` backend the number of max detection thresholds should.*"
+            ):
+                metric = MeanAveragePrecision(max_detection_thresholds=max_detection_thresholds, backend=backend)
+        else:
+            metric = MeanAveragePrecision(max_detection_thresholds=max_detection_thresholds, backend=backend)
+            metric(preds, target)