okooooooo/supervision
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'fix/mAP-area' of https://github.com/roboflow/supervision into fix/mAP-area

Browse Source
This commit is contained in:
Balthasar committed 2025-07-16 20:07:30 +02:00
2 parents a444b3bde8 4009f9d886
commit bf88636487
2 files changed
+83 -63
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

supervision/metrics/mean_average_precision.py
+1 -1
View File
@@ -1237,7 +1237,7 @@ class MeanAveragePrecision(Metric):
area = data.get("area") if data else None
if area is None:
area = image_targets.area[target_idx]
dict_annotation = {
"area": area,
"iscrowd": data.get("iscrowd", 0),
test/metrics/test_mean_average_precision_area.py
+82 -62
View File
@@ -13,40 +13,47 @@ class TestMeanAveragePrecisionArea:
"""Test that area is calculated from bbox when data is empty (normal case)."""
# Create detections with empty data (normal case)
gt = Detections(
xyxy=np.array([
[10, 10, 40, 40], # Small: 30x30 = 900
[100, 100, 200, 150], # Medium: 100x50 = 5000
[300, 300, 500, 400] # Large: 200x100 = 20000
], dtype=np.float32),
xyxy=np.array(
[
[10, 10, 40, 40], # Small: 30x30 = 900
[100, 100, 200, 150], # Medium: 100x50 = 5000
[300, 300, 500, 400], # Large: 200x100 = 20000
],
dtype=np.float32,
),
class_id=np.array([0, 0, 0]),
confidence=np.array([1.0, 1.0, 1.0])
confidence=np.array([1.0, 1.0, 1.0]),
)
pred = Detections(
xyxy=gt.xyxy.copy(),
class_id=gt.class_id.copy(),
confidence=np.array([0.9, 0.9, 0.9])
confidence=np.array([0.9, 0.9, 0.9]),
)
# Verify data is empty (normal case)
assert gt.data == {}
assert pred.data == {}
# Create mAP metric and test area calculation
map_metric = MeanAveragePrecision()
map_metric.update([pred], [gt])
# Check that areas were calculated correctly from bbox
prepared_targets = map_metric._prepare_targets(map_metric._targets_list)
areas = [ann["area"] for ann in prepared_targets["annotations"]]
expected_areas = [900.0, 5000.0, 20000.0] # width * height for each bbox
assert np.allclose(areas, expected_areas, rtol=1e-05, atol=1e-08), f"Expected {expected_areas}, got {areas}"
assert np.allclose(areas, expected_areas, rtol=1e-05, atol=1e-08), (
f"Expected {expected_areas}, got {areas}"
)
# Verify mAP works correctly (no -1.0 for medium/large objects)
result = map_metric.compute()
assert result.medium_objects.map50 >= 0.0, "Medium objects should have valid mAP"
assert result.medium_objects.map50 >= 0.0, (
"Medium objects should have valid mAP"
)
assert result.large_objects.map50 >= 0.0, "Large objects should have valid mAP"
def test_area_preserved_when_provided_in_data(self):
@@ -55,125 +62,138 @@ class TestMeanAveragePrecisionArea:
gt = Detections(
xyxy=np.array([[100, 100, 200, 150]], dtype=np.float32), # Would be 5000
class_id=np.array([0]),
confidence=np.array([1.0])
confidence=np.array([1.0]),
)
# Add custom area to data (different from calculated)
gt.data = {"area": np.array([3000.0])}
pred = Detections(
xyxy=gt.xyxy.copy(),
class_id=gt.class_id.copy(),
confidence=np.array([0.9])
xyxy=gt.xyxy.copy(), class_id=gt.class_id.copy(), confidence=np.array([0.9])
)
pred.data = {"area": np.array([3000.0])}
# Test area calculation
map_metric = MeanAveragePrecision()
map_metric.update([pred], [gt])
# Check that provided area is used (not calculated)
prepared_targets = map_metric._prepare_targets(map_metric._targets_list)
used_area = prepared_targets["annotations"][0]["area"]
assert np.allclose(used_area, 3000.0, rtol=1e-05, atol=1e-08), f"Should use provided area 3000.0, got {used_area}"
assert np.allclose(used_area, 3000.0, rtol=1e-05, atol=1e-08), (
f"Should use provided area 3000.0, got {used_area}"
)
# Verify it's different from what would be calculated
calculated_area = (200 - 100) * (150 - 100) # 100 * 50 = 5000
assert not np.allclose(used_area, calculated_area, rtol=1e-05, atol=1e-08), "Should use provided area, not calculated"
assert not np.allclose(used_area, calculated_area, rtol=1e-05, atol=1e-08), (
"Should use provided area, not calculated"
)
def test_mixed_area_sources(self):
"""Test mix of detections with and without area in data."""
# Create detections where some have area in data, others don't
gt1 = Detections(
xyxy=np.array([[10, 10, 40, 40]], dtype=np.float32), # 900
class_id=np.array([0])
class_id=np.array([0]),
)
# No area in data - should be calculated
gt2 = Detections(
xyxy=np.array([[100, 100, 200, 150]], dtype=np.float32), # 5000
class_id=np.array([1])
class_id=np.array([1]),
)
# Add area in data - should be preserved
gt2.data = {"area": np.array([3000.0])}
pred1 = Detections(
xyxy=gt1.xyxy.copy(),
class_id=gt1.class_id.copy(),
confidence=np.array([0.9])
confidence=np.array([0.9]),
)
pred2 = Detections(
xyxy=gt2.xyxy.copy(),
class_id=gt2.class_id.copy(),
confidence=np.array([0.8])
confidence=np.array([0.8]),
)
pred2.data = {"area": np.array([3000.0])}
# Test area calculation for mixed sources
map_metric = MeanAveragePrecision()
map_metric.update([pred1, pred2], [gt1, gt2])
prepared_targets = map_metric._prepare_targets(map_metric._targets_list)
areas = [ann["area"] for ann in prepared_targets["annotations"]]
expected_areas = [900.0, 3000.0] # calculated, then provided
assert np.allclose(areas, expected_areas, rtol=1e-05, atol=1e-08), f"Expected {expected_areas}, got {areas}"
assert np.allclose(areas, expected_areas, rtol=1e-05, atol=1e-08), (
f"Expected {expected_areas}, got {areas}"
)
def test_size_specific_map_works_correctly(self):
"""Test that size-specific mAP works correctly with area fix."""
# Create detections with one object of each size
gt = Detections(
xyxy=np.array([
[10, 10, 40, 40], # Small: 30x30 = 900 < 1024
[100, 100, 200, 150], # Medium: 100x50 = 5000 (1024 <= x < 9216)
[300, 300, 500, 400] # Large: 200x100 = 20000 >= 9216
], dtype=np.float32),
class_id=np.array([0, 0, 0])
xyxy=np.array(
[
[10, 10, 40, 40], # Small: 30x30 = 900 < 1024
[100, 100, 200, 150], # Medium: 100x50 = 5000 (1024 <= x < 9216)
[300, 300, 500, 400], # Large: 200x100 = 20000 >= 9216
],
dtype=np.float32,
),
class_id=np.array([0, 0, 0]),
)
# Perfect predictions
pred = Detections(
xyxy=gt.xyxy.copy(),
class_id=gt.class_id.copy(),
confidence=np.array([0.9, 0.9, 0.9])
confidence=np.array([0.9, 0.9, 0.9]),
)
# Test mAP calculation
map_metric = MeanAveragePrecision()
map_metric.update([pred], [gt])
result = map_metric.compute()
# All size categories should have valid results (not -1.0)
assert result.small_objects.map50 >= 0.0, "Small objects should have valid mAP"
assert result.medium_objects.map50 >= 0.0, "Medium objects should have valid mAP"
assert result.medium_objects.map50 >= 0.0, (
"Medium objects should have valid mAP"
)
assert result.large_objects.map50 >= 0.0, "Large objects should have valid mAP"
# Perfect matches should yield high mAP for medium and large
assert result.medium_objects.map50 > 0.9, "Perfect medium matches should have high mAP"
assert result.large_objects.map50 > 0.9, "Perfect large matches should have high mAP"
assert result.medium_objects.map50 > 0.9, (
"Perfect medium matches should have high mAP"
)
assert result.large_objects.map50 > 0.9, (
"Perfect large matches should have high mAP"
)
def test_area_uses_detections_property(self):
"""Test that area calculation uses Detections.area property correctly."""
# Create detection
gt = Detections(
xyxy=np.array([[100, 100, 200, 150]], dtype=np.float32),
class_id=np.array([0])
class_id=np.array([0]),
)
pred = Detections(
xyxy=gt.xyxy.copy(),
class_id=gt.class_id.copy(),
confidence=np.array([0.9])
xyxy=gt.xyxy.copy(), class_id=gt.class_id.copy(), confidence=np.array([0.9])
)
# Test that internal calculation matches Detections.area property
map_metric = MeanAveragePrecision()
map_metric.update([pred], [gt])
prepared_targets = map_metric._prepare_targets(map_metric._targets_list)
used_area = prepared_targets["annotations"][0]["area"]
expected_area = gt.area[0]
assert np.allclose(used_area, expected_area, rtol=1e-05, atol=1e-08), f"Should use Detections.area property {expected_area}, got {used_area}"
assert np.allclose(used_area, expected_area, rtol=1e-05, atol=1e-08), (
f"Should use Detections.area property {expected_area}, got {used_area}"
)