Add a quality setting for avoiding using bbox in point group matching

cvat/apps/quality_control/migrations/0004_qualitysettings_use_image_space_for_point_group_comparisons.py

@@ -0,0 +1,18 @@
+# Generated by Django 4.2.15 on 2024-11-01 11:59
+
+from django.db import migrations, models
+
+
+class Migration(migrations.Migration):
+
+    dependencies = [
+        ("quality_control", "0003_qualityreport_assignee_last_updated_and_more"),
+    ]
+
+    operations = [
+        migrations.AddField(
+            model_name="qualitysettings",
+            name="use_image_space_for_point_group_comparisons",
+            field=models.BooleanField(default=False),
+        ),
+    ]

cvat/apps/quality_control/models.py

@@ -205,6 +205,8 @@ class QualitySettings(models.Model):
 
     low_overlap_threshold = models.FloatField()
 
+    use_image_space_for_point_group_comparisons = models.BooleanField(default=False)
+
     compare_line_orientation = models.BooleanField()
     line_orientation_threshold = models.FloatField()
 

cvat/apps/quality_control/quality_reports.py

@@ -187,6 +187,9 @@ class ComparisonParameters(_Serializable):
     oks_sigma: float = 0.09
     "Like IoU threshold, but for points, % of the bbox area to match a pair of points"
 
+    use_image_space_for_point_group_comparisons: bool = False
+    "Compare points in the image space, instead of using the point group bbox"
+
     line_thickness: float = 0.01
     "Thickness of polylines, relatively to the (image area) ^ 0.5"
 
@@ -955,6 +958,7 @@ def __init__(
         # https://cocodataset.org/#keypoints-eval
         # https://github.com/cocodataset/cocoapi/blob/8c9bcc3cf640524c4c20a9c40e89cb6a2f2fa0e9/PythonAPI/pycocotools/cocoeval.py#L523
         oks_sigma: float = 0.09,
+        use_image_space_for_point_group_comparisons: bool = False,
         compare_line_orientation: bool = False,
         line_torso_radius: float = 0.01,
         panoptic_comparison: bool = False,
@@ -968,6 +972,11 @@ def __init__(
         self.oks_sigma = oks_sigma
         "% of the shape area"
 
+        self.use_image_space_for_point_group_comparisons = (
+            use_image_space_for_point_group_comparisons
+        )
+        "Do not infer bbox for point group comparison, use the image space"
+
         self.compare_line_orientation = compare_line_orientation
         "Whether lines are oriented or not"
 
@@ -1293,13 +1302,18 @@ def _distance(a: dm.Points, b: dm.Points) -> float:
             else:
                 # Complex case: multiple points, grouped points, points with a bbox
                 # Try to align points and then return the metric
-                # match them in their bbox space
 
-                if dm.ops.bbox_iou(a_bbox, b_bbox) <= 0:
-                    return 0
+                if self.use_image_space_for_point_group_comparisons:
+                    scale = img_h * img_w
+                else:
+                    # match points in their bbox space
+
+                    if dm.ops.bbox_iou(a_bbox, b_bbox) <= 0:
+                        # this early exit may not work for points forming an axis-aligned line
+                        return 0
 
-                bbox = dm.ops.mean_bbox([a_bbox, b_bbox])
-                scale = bbox[2] * bbox[3]
+                    bbox = dm.ops.mean_bbox([a_bbox, b_bbox])
+                    scale = bbox[2] * bbox[3]
 
                 a_points = np.reshape(a.points, (-1, 2))
                 b_points = np.reshape(b.points, (-1, 2))
@@ -1525,6 +1539,9 @@ def __init__(self, categories: dm.CategoriesInfo, *, settings: ComparisonParamet
             panoptic_comparison=settings.panoptic_comparison,
             iou_threshold=settings.iou_threshold,
             oks_sigma=settings.oks_sigma,
+            use_image_space_for_point_group_comparisons=(
+                settings.use_image_space_for_point_group_comparisons
+            ),
             line_torso_radius=settings.line_thickness,
             compare_line_orientation=False,  # should not be taken from outside, handled differently
         )

cvat/apps/quality_control/serializers.py

@@ -81,6 +81,7 @@ class Meta:
             "max_validations_per_job",
             "iou_threshold",
             "oks_sigma",
+            "use_image_space_for_point_group_comparisons",
             "line_thickness",
             "low_overlap_threshold",
             "compare_line_orientation",
@@ -98,6 +99,9 @@ class Meta:
         )
 
         extra_kwargs = {k: {"required": False} for k in fields}
+        extra_kwargs.setdefault("use_image_space_for_point_group_comparisons", {}).setdefault(
+            "default", False
+        )
 
         for field_name, help_text in {
             "target_metric": "The primary metric used for quality estimation",
@@ -119,6 +123,11 @@ class Meta:
                 where the checked point is expected to be.
                 Read more: https://cocodataset.org/#keypoints-eval
             """,
+            "use_image_space_for_point_group_comparisons": """
+                Compare point groups in the image space, instead of using the point group bbox.
+                Useful if point groups may not represent a single object or grouped boxes
+                do not represent object boundaries.
+            """,
             "line_thickness": """
                 Thickness of polylines, relatively to the (image area) ^ 0.5.
                 The distance to the boundary around the GT line,