def test_per_class_map():
annotations = [
Annotation.ground_truth("a", BBox(0, 5, 0, 5)),
Annotation.prediction("a", BBox(0, 5, 0, 5), 0.9),
Annotation.ground_truth("b", BBox(0, 5, 0, 5)),
Annotation.prediction("b", BBox(5, 6, 5, 6), 0.9),
]
metrics = get_metrics([annotations], iou_threshold=0.9)
assert metrics.per_class["a"].AP == 1
assert metrics.per_class["b"].AP == 0
assert metrics.mAP == 0.5
def test_iou():
bb_a = BBox(0, 2, 0, 2)
bb_b = BBox(3, 5, 0, 1)
bb_c = BBox(0, 1, 3, 4)
assert iou(bb_a, bb_b) == 0
assert iou(bb_a, bb_c) == 0
assert iou(bb_b, bb_c) == 0
assert iou(bb_a, bb_a) == 1
bb_d = BBox(1, 3, 1, 3)
assert iou(bb_a, bb_d) == get_intersection_area(bb_a, bb_d) / get_union_area(
bb_a, bb_d
)
def test_to_xywh():
bbox = BBox(xmin=5, xmax=6, ymin=10, ymax=30)
x, y, w, h = bbox.xywh()
assert x == 5
assert w == 1
assert y == 10
assert h == 20
def test_get_union_area():
bb_a = BBox(0, 2, 0, 2)
bb_b = BBox(3, 5, 0, 1)
bb_c = BBox(0, 1, 3, 4)
assert get_union_area(bb_a, bb_b) == bb_a.area + bb_b.area
assert get_union_area(bb_a, bb_c) == bb_a.area + bb_c.area
assert get_union_area(bb_b, bb_c) == bb_b.area + bb_c.area
assert get_union_area(bb_a, bb_a) == bb_a.area
bb_d = BBox(1, 3, 1, 3)
assert get_union_area(bb_a, bb_d) == bb_a.area + bb_d.area - get_intersection_area(
bb_a, bb_d
)
bb_e = BBox(0, 5, 0, 5)
assert get_union_area(bb_e, bb_b) == bb_e.area
def test_get_intersection_area():
bb_a = BBox(0, 2, 0, 2)
bb_b = BBox(3, 5, 0, 1)
bb_c = BBox(0, 1, 3, 4)
assert get_intersection_area(bb_a, bb_b) == 0
assert get_intersection_area(bb_a, bb_c) == 0
assert get_intersection_area(bb_b, bb_c) == 0
bb_d = BBox(1, 3, 1, 3)
assert get_intersection_area(bb_a, bb_d) == 1
assert get_intersection_area(bb_d, bb_a) == 1
assert get_intersection_area(bb_a, bb_a) == 4
bb_e = BBox(0, 5, 0, 5)
assert get_intersection_area(bb_e, bb_a) == 4
assert get_intersection_area(bb_e, bb_b) == 2
def test_metrics_perfect_prediction():
bb = BBox(0, 5, 0, 5).normalize(10, 10)
annotations = [
Annotation.ground_truth("a", bb),
Annotation.prediction("a", bb, 0.9),
]
metrics = get_metrics([annotations])
assert metrics.mAP == 1.0
def test_metrics_missing_gt():
bb = BBox(0, 5, 0, 5).normalize(10, 10)
annotations = [
Annotation.prediction("a", bb, 0.9),
Annotation.prediction("b", bb, 0.9),
Annotation.prediction("b", bb, 0.8),
]
metrics = get_metrics([annotations])
assert metrics.per_class["a"].total_FP == 1
assert metrics.per_class["b"].total_FP == 2
assert metrics.mAP == 0.0
def test_metrics_two_predictions_one_gt_2():
width, height = 10, 10
bbox = BBox(0, 2, 0, 2).normalize(width, height)
annotations = [
Annotation.ground_truth("a", bbox),
Annotation.prediction("a", bbox, 0.9),
Annotation.prediction("a", bbox.move(0.5, 0.5), 0.5),
]
metrics = get_metrics([annotations])
assert metrics.per_class["a"].total_FP == 1
assert metrics.per_class["a"].total_TP == 1
assert metrics.mAP == 1
def test_metrics_multiple_images_perfect_prediction():
width, height = 10, 10
bbox = BBox(0, 5, 0, 5).normalize(width, height)
image_a = [
Annotation.ground_truth("a", bbox),
Annotation.prediction("a", bbox, 0.9),
]
image_b = [
Annotation.ground_truth("a", bbox),
Annotation.prediction("a", bbox, 0.9),
]
metrics = get_metrics([image_a, image_b])
assert metrics.mAP == 1.0
def test_iou_threshold():
bbox = BBox(0, 5, 0, 5)
annotations = [
Annotation.ground_truth("a", bbox),
Annotation.prediction("a", bbox.move(2.5, 0), 0.9),
]
metrics = get_metrics([annotations], iou_threshold=0.9)
assert metrics.per_class["a"].total_FP == 1
assert metrics.per_class["a"].total_TP == 0
assert metrics.mAP == 0
metrics = get_metrics([annotations], iou_threshold=0.2)
assert metrics.per_class["a"].total_FP == 0
assert metrics.per_class["a"].total_TP == 1
assert metrics.mAP == 1
def test_bb_intersect():
bb_a = BBox(0, 2, 0, 2)
bb_b = BBox(3, 5, 0, 1)
bb_c = BBox(0, 1, 3, 4)
assert not boxes_intersect(bb_a, bb_b)
assert not boxes_intersect(bb_a, bb_c)
assert not boxes_intersect(bb_b, bb_c)
bb_d = BBox(2, 4, 2, 4)
assert not boxes_intersect(bb_d, bb_a)
bb_e = BBox(1, 4, 1, 4)
assert boxes_intersect(bb_e, bb_a)
assert not boxes_intersect(bb_e, bb_b)
assert not boxes_intersect(bb_e, bb_c)
assert boxes_intersect(bb_e, bb_d)
bb_f = BBox(0, 5, 0, 5)
assert boxes_intersect(bb_f, bb_a)
assert boxes_intersect(bb_f, bb_b)
assert boxes_intersect(bb_f, bb_c)
assert boxes_intersect(bb_f, bb_d)
assert boxes_intersect(bb_f, bb_e)
def test_invalid_bbox():
with pytest.raises(Exception):
BBox(xmin=0, xmax=0, ymin=0.5, ymax=0.6)
with pytest.raises(Exception):
BBox(xmin=1, xmax=0, ymin=0.3, ymax=0.5)
def test_normalize_bbox():
bbox = BBox(xmin=0, xmax=50, ymin=50, ymax=75)
bbox = bbox.normalize(100, 100)
assert bbox.as_tuple() == (0, 0.5, 0.5, 0.75)
def test_metrics_missing_prediction():
bb = BBox(0, 5, 0, 5).normalize(10, 10)
annotations = [Annotation.ground_truth("a", bb)]
metrics = get_metrics([annotations])
assert metrics.mAP == 0.0