documenting

docs/source/user_guide/evaluation.rst

@@ -729,8 +729,9 @@ The only difference between each task type is in how the IoU between objects is
 calculated:
 
 -   For object detections, IoUs are computed between each pair of bounding boxes
--   For instance segmentations and polygons, IoUs are computed between the
-    polygonal shapes rather than their rectangular bounding boxes
+-   For instance segmentations, when ``use_masks=True``, IoUs are computed
+    between the dense pixel masks rather than their rectangular bounding boxes
+-   For polygons, IoUs are computed between the polygonal shapes
 -   For keypoint tasks,
     `object keypoint similarity <https://cocodataset.org/#keypoints-eval>`_
     is computed for each pair of objects, using the extent of the ground truth
@@ -744,8 +745,7 @@ stored in |Detections| format.
 
 For instance segmentation tasks, the ground truth and predicted objects should
 be stored in |Detections| format, and each |Detection| instance should have its
-:attr:`mask <fiftyone.core.labels.Detection.mask>` attribute populated to
-define the extent of the object within its bounding box.
+mask populated to define the extent of the object within its bounding box.
 
 .. note::
 

fiftyone/utils/eval/coco.py

@@ -48,7 +48,8 @@ class COCOEvaluationConfig(DetectionEvaluationConfig):
             of the provided :class:`fiftyone.core.labels.Polyline` instances
             rather than using their actual geometries
         tolerance (None): a tolerance, in pixels, when generating approximate
-            polylines for instance masks. Typical values are 1-3 pixels
+            polylines for instance masks. Typical values are 1-3 pixels. By
+            default, IoUs are computed directly on the dense pixel masks
         compute_mAP (False): whether to perform the necessary computations so
             that mAP and PR curves can be generated
         iou_threshs (None): a list of IoU thresholds to use when computing mAP

fiftyone/utils/eval/openimages.py

@@ -43,7 +43,8 @@ class OpenImagesEvaluationConfig(DetectionEvaluationConfig):
             of the provided :class:`fiftyone.core.labels.Polyline` instances
             rather than using their actual geometries
         tolerance (None): a tolerance, in pixels, when generating approximate
-            polylines for instance masks. Typical values are 1-3 pixels
+            polylines for instance masks. Typical values are 1-3 pixels. By
+            default, IoUs are computed directly on the dense pixel masks
         max_preds (None): the maximum number of predicted objects to evaluate
             when computing mAP and PR curves
         error_level (1): the error level to use when manipulating instance

fiftyone/utils/iou.py

@@ -74,7 +74,8 @@ def compute_ious(
             of the provided :class:`fiftyone.core.labels.Polyline` instances
             rather than using their actual geometries
         tolerance (None): a tolerance, in pixels, when generating approximate
-            polylines for instance masks. Typical values are 1-3 pixels
+            polylines for instance masks. Typical values are 1-3 pixels. By
+            default, IoUs are computed directly on the dense pixel masks
         sparse (False): whether to return a sparse dict of non-zero IoUs rather
             than a full matrix
         error_level (1): the error level to use when manipulating instance