def test_evalution_iou_multiclass_with_images():
classes = ['bkg', 'kite', 'person']
mask_shape = (480, 640)
gt_bbox_1 = BBox(10, 10, 10, 10, 1)
mask_bin_GT_1 = utils.bin_mask_from_bb(mask_shape, gt_bbox_1)
pred_bbox_1 = BBox(10, 10, 10, 10, 1)
mask_bin_pred_1 = utils.bin_mask_from_bb(mask_shape, pred_bbox_1)
gt_bbox_2 = BBox(110, 110, 320, 280, 2)
mask_bin_GT_2 = utils.bin_mask_from_bb(mask_shape, gt_bbox_2)
pred_bbox_2 = BBox(70, 50, 240, 220, 2)
mask_bin_pred_2 = utils.bin_mask_from_bb(mask_shape, pred_bbox_2)
mask_multi_GT = np.maximum(mask_bin_GT_1, mask_bin_GT_2 * 2)
mask_multi_pred = np.maximum(mask_bin_pred_1, mask_bin_pred_2 * 2)
metrics = evaluate_segmentation([mask_multi_GT], [mask_multi_pred],
classes)
assert metrics.count == 480 * 640
assert metrics.by_class[1].iou == 1.0
assert round(metrics.by_class[2].iou, 3) == .290
assert round(metrics.avg_iou_no_bkg, 3) == .645
def evaluate_detection_example():
# Class names
classes = ['kite', 'person']
# Image shape
mask_shape = (480, 640)
# Creating fake data
gt_bbox_1 = BBox(10, 10, 10, 10, 0, 1)
pred_bbox_1 = BBox(10, 10, 10, 10, 0, 1)
gt_bbox_2 = BBox(110, 110, 320, 280, 1, 1)
pred_bbox_2 = BBox(70, 50, 240, 220, 1, 1)
# Creating data suplier iterator
# It is not necessary here, but it's useful if you want to yield data
# from the disk i.e. from a Pytorch DataLoader
it_gt_masks = identity_iterator([gt_bbox_1, gt_bbox_2])
it_pred_masks = identity_iterator([pred_bbox_1, pred_bbox_2])
# Calculates and shows metrics
metrics = evaluate_detection(it_gt_masks, it_pred_masks, classes)
# Shows confusion matrix and returns its Figure and Axes
fig, axes = metrics.show_confusion_matrix()
# Shows confusion matrix for class `a`
metrics.by_class[0].show_confusion_matrix()
def show_detection_example():
# Class names
classes = ['kite', 'person', 'car']
# Image shape
img_shape = (480, 640)
# Bboxes creation
gt_bbox_1 = BBox(10, 10, 100, 100, 0)
pred_bbox_1 = BBox(10, 10, 100, 100, 0, .8212158464)
gt_bbox_2 = BBox(110, 110, 320, 280, 1)
pred_bbox_2 = BBox(70, 50, 240, 220, 1, .34844545)
gt_bbox_3 = BBox(300, 300, 100, 100, 2)
# Convert to results
results = [
Result(random_image(img_shape[0], img_shape[1]),
[gt_bbox_1, gt_bbox_2, gt_bbox_3], [pred_bbox_1, pred_bbox_2])
]
# GT only result
results = [
Result(random_image(img_shape[0], img_shape[1]),
[gt_bbox_1, gt_bbox_2, gt_bbox_3])
] + results
# Shows results and returns its Figure and Axes
fig, axes = show_detections(results, classes, show_bbox_label=False)
def test_center_point_box():
b0 = BBox(0, 0, 3, 3, 0)
assert b0.center_point == (1, 1)
b1 = BBox(100, 100, 100, 100, 0)
assert b1.center_point == (149, 149)
b2 = BBox(500, 500, 100, 100, 0)
assert b2.center_point == (549, 549)
b3 = BBox(100, 200, 76, 76, 0)
assert b3.center_point == (137, 237)
def test_evaluation_detection_iou_metric():
classes = ['kite']
gt_bbox = BBox(110, 110, 320, 280, 0)
pred_bbox = BBox(70, 50, 240, 220, 0, .5)
metrics = evaluate_detection([[gt_bbox]], [[pred_bbox]], classes)
assert round(metrics.avg_iou, 3) == .290
assert round(metrics.by_class[0].iou, 3) == .290
def evaluate_segmentation_example():
# Class names - Background must be at 0 index
classes = ['bkg', 'kite', 'person']
# Image shape
mask_shape = (480, 640)
# Creating fake data
# Creates a rectangle of 1s in a 0s array
gt_bbox_1 = BBox(10, 10, 10, 10, 1)
mask_bin_GT_1 = draw_rectangle(np.zeros(mask_shape, np.int),
gt_bbox_1.upper_left_point,
gt_bbox_1.bottom_right_point, 1)
pred_bbox_1 = BBox(10, 10, 10, 10, 1)
mask_bin_pred_1 = draw_rectangle(np.zeros(mask_shape, np.int),
pred_bbox_1.upper_left_point,
pred_bbox_1.bottom_right_point, 1)
# Creates a rectangle of 2s in a 0s array
gt_bbox_2 = BBox(110, 110, 320, 280, 2)
mask_bin_GT_2 = draw_rectangle(np.zeros(mask_shape, np.int),
gt_bbox_2.upper_left_point,
gt_bbox_2.bottom_right_point, 2)
pred_bbox_2 = BBox(70, 50, 240, 220, 2)
mask_bin_pred_2 = draw_rectangle(np.zeros(mask_shape, np.int),
pred_bbox_2.upper_left_point,
pred_bbox_2.bottom_right_point, 2)
# Merging masks
mask_GT = np.maximum(mask_bin_GT_1, mask_bin_GT_2)
mask_pred = np.maximum(mask_bin_pred_1, mask_bin_pred_2)
# Creating data suplier iterator
# It is not necessary here, but it's useful if you want to yield data
# from the disk i.e. from a Pytorch DataLoader
it_gt_masks = identity_iterator(mask_GT)
it_pred_masks = identity_iterator(mask_pred)
# Calculates and shows metrics
metrics = evaluate_segmentation(it_gt_masks, it_pred_masks, classes)
# Shows confusion matrix and returns its Figure and Axes
fig, axes = metrics.show_confusion_matrix()
# Shows confusion matrix for class `a`
metrics.by_class[0].show_confusion_matrix()
def test_evalution_iou_with_image():
classes = ['bkg', 'kite']
mask_shape = (480, 640)
gt_bbox = BBox(110, 110, 320, 280, 1)
mask_bin_GT = utils.bin_mask_from_bb(mask_shape, gt_bbox)
pred_bbox = BBox(70, 50, 240, 220, 1)
mask_bin_pred = utils.bin_mask_from_bb(mask_shape, pred_bbox)
metrics = evaluate_segmentation([mask_bin_GT], [mask_bin_pred], classes)
assert metrics.count == 480 * 640
assert round(metrics.by_class[1].iou, 3) == .290
assert round(metrics.avg_iou_no_bkg, 3) == .290
assert round(metrics.avg_iou, 3) == .502
def show_segmentation_example():
# Class names
classes = ['bkg', 'kite', 'person', 'car']
# Image/mask shape
mask_shape = (480, 640)
# Multiclass mask creation
gt_bbox_1 = BBox(10, 10, 100, 100, 1)
mask_bin_GT_1 = draw_bboxes(mask_shape, [gt_bbox_1])
pred_bbox_1 = BBox(10, 10, 100, 100, 1)
mask_bin_pred_1 = draw_bboxes(mask_shape, [pred_bbox_1])
gt_bbox_2 = BBox(110, 110, 320, 280, 2)
mask_bin_GT_2 = draw_bboxes(mask_shape, [gt_bbox_2])
pred_bbox_2 = BBox(70, 50, 240, 220, 2)
mask_bin_pred_2 = draw_bboxes(mask_shape, [pred_bbox_2])
gt_bbox_3 = BBox(300, 300, 100, 100, 3)
mask_bin_GT_3 = draw_bboxes(mask_shape, [gt_bbox_3])
mask_multi_GT = np.maximum(np.maximum(mask_bin_GT_1, mask_bin_GT_2 * 2),
mask_bin_GT_3 * 3)
mask_multi_pred = np.maximum(mask_bin_pred_1, mask_bin_pred_2 * 2)
# Convert to results
results = [
Result(random_image(mask_shape[0], mask_shape[1]), mask_multi_GT,
mask_multi_pred)
]
# GT only result
results = [
Result(random_image(mask_shape[0], mask_shape[1]), mask_multi_GT)
] + results
# Custom color map
cmap = ListedColormap(['#ff0000', '#00ff00', '#0000ff', '#ffffff'])
# Shows results and returns its Figure and Axes
fig, axes = show_segmentations(results, classes, cmap=cmap)
def test_evaluation_detection_no_pred():
classes = ['kite', 'person']
gt_bbox = BBox(110, 110, 320, 280, 0)
metrics = evaluate_detection([[gt_bbox]], [[]], classes)
assert round(metrics.avg_iou, 3) == 0
assert round(metrics.by_class[0].iou, 3) == 0
assert metrics._by_class[0].FN == 1
def test_evaluation_detection_no_gt():
classes = ['kite', 'person']
pred_bbox = BBox(70, 50, 240, 220, 0)
metrics = evaluate_detection([[]], [[pred_bbox]], classes)
assert round(metrics.avg_iou, 3) == 0
assert round(metrics.by_class[0].iou, 3) == 0
assert metrics._by_class[0].FP == 1
def test_calculate_pairwise_bbox_ious():
gt_bboxes = [
BBox(39, 63, 203 - 39 + 1, 112 - 63 + 1, 0),
BBox(49, 75, 203 - 49 + 1, 125 - 75 + 1, 0),
BBox(31, 69, 201 - 31 + 1, 125 - 69 + 1, 0),
BBox(50, 72, 197 - 50 + 1, 121 - 72 + 1, 0),
BBox(35, 51, 196 - 35 + 1, 110 - 51 + 1, 0),
BBox(35, 51, 196, 110, 0),
BBox(0, 0, 5, 5, 0),
]
pred_bboxes = [
BBox(54, 66, 198 - 54 + 1, 114 - 66 + 1, 0),
BBox(42, 78, 186 - 42 + 1, 126 - 78 + 1, 0),
BBox(18, 63, 235 - 18 + 1, 135 - 63 + 1, 0),
BBox(54, 72, 198 - 54 + 1, 120 - 72 + 1, 0),
BBox(36, 60, 180 - 36 + 1, 108 - 60 + 1, 0),
BBox(35, 51, 196, 110, 0),
BBox(6, 6, 5, 5, 0),
]
ious = calculate_pairwise_bbox_ious(gt_bboxes, pred_bboxes)
assert round(ious[0, 0], 3) == .798
assert round(ious[1, 1], 3) == .79
assert round(ious[2, 2], 3) == .612
assert round(ious[3, 3], 3) == .947
assert round(ious[4, 4], 3) == .731
assert ious[5, 5] == 1
assert ious[6, 6] == 0
