def test_random_overlapping(self):
"""Tests by creating a random set of boxes: one true one per class
and a number of random ones.
Don't start using this test until you've got all other,
more basic tests working."""
def random_in_range(a_min, a_max):
return a_min + random.random() * (a_max - a_min)
# Use today's date as a seed, so that we are reproducible all day
# today (while debugging:)
random.seed(str(date.today()))
MIN_COORD = 0
MAX_COORD = 1000
for run in range(20):
boxes = []
scores = []
classes = []
good_indexes = []
for c in range(99):
# For each class, create one real box, and a number
# of boxes with lower scores that overlap the real box
left = random.randrange(MAX_COORD - 1)
top = random.randrange(MAX_COORD - 1)
right = random.randrange(left + 1, MAX_COORD)
bottom = random.randrange(top + 1, MAX_COORD)
real_box = [left, top, right, bottom]
real_score = random_in_range(.3, .9)
boxes.append(real_box)
scores.append(real_score)
classes.append(c)
if real_score > 0.4:
good_indexes.append(len(boxes) - 1)
for fake_boxes in range(20):
h_factor = random_in_range(-.2, .2)
v_factor = random_in_range(-.2, .2)
delta_h = (right - left) * h_factor
delta_v = (bottom - top) * v_factor
fake_box = [
left + delta_h, top + delta_v, right + delta_h,
bottom + delta_v
]
boxes.append(fake_box)
fake_score = random_in_range(real_score - .2,
real_score - .001)
scores.append(fake_score)
classes.append(c)
result = nms(np.array(boxes), np.array(scores),
np.array(classes))
expected = np.array(good_indexes)
self.assertEqualAnyOrder(
np.array(good_indexes),
nms(np.array(boxes), np.array(scores), np.array(classes)))
def test_high_iou_different_class(self):
""" the boxes are different classes and the iou is high but since
they are different classes, it should return all the boxes """
boxes = np.asarray([(1,1,2,2),(1.01,1,2.01,2)])
classes = np.asarray([1,2])
scores = [0.6,0.5]
self.assertEqualAnyOrder(nms(boxes, scores, classes), np.array([0,1]))
def test_high_iou_same_class(self):
""" the boxes are the same classes but the iou is high so you should
the box index with the highest score """
boxes = np.asarray([(1,1,2,2),(1.01,1,2.01,2)])
classes = np.asarray([1,1])
scores = np.asarray([0.6,0.5])
self.assertEqualAnyOrder(nms(boxes, scores, classes), np.array([0]))
def test_low_iou_different_classes(self):
""" the iou is greater than 0.45 but the boxes are different classes, so
it should return the indices of all the boxes """
boxes = np.asarray([[1,2,3,4],[2,3,5,6]])
classes = np.asarray([1,2])
scores = np.asarray([0.5,0.55])
self.assertEqualAnyOrder(nms(boxes, scores, classes), np.asarray([0,1]))
def test_same_classes_lowscore(self):
""" the scores are less than 0.4 but the classes are the same, so it
should not return anything """
boxes = np.asarray([(1,2,3,4),(2,3,5,6)])
classes = np.asarray([1,1])
scores = np.asarray([0.2,0.3])
self.assertEqualAnyOrder(nms(boxes, scores, classes), np.array([]))
def test_same_class_multiple_overlap(self):
boxes = [[1, 5, 4, 10], [1, 5, 4, 10]]
boxes = np.asarray(boxes)
scores = [0.9, 0.8]
scores = np.asarray(scores)
classes = [2, 2]
classes = np.array(classes)
self.assertEqualAnyOrder(nms(boxes, scores, classes), np.asarray([0]))
def test_below_min_score(self):
boxes = [[1, 2, 3, 4], [2, 1, 3, 4], [1, 3, 4, 5]]
boxes = np.asarray(boxes)
scores = [0.3, 0.2, 0.1]
scores = np.asarray(scores)
classes = [0, 1, 2]
classes = np.asarray(classes)
self.assertEqualAnyOrder(nms(boxes, scores, classes), np.asarray([]))
def test_multiple_class_no_overlap(self):
boxes = [[1, 2, 3, 4], [2, 1, 3, 4], [5, 10, 7, 11]]
boxes = np.asarray(boxes)
scores = [0.9, 0.8, 0.7]
scores = np.asarray(scores)
classes = [0, 1, 2]
classes = np.asarray(classes)
self.assertEqualAnyOrder(nms(boxes, scores, classes),
np.asarray([0, 1, 2]))
def test_same_class_no_overlap(self):
boxes = [[10, 10, 20, 20], [30, 10, 40, 20]]
boxes = np.asarray(boxes)
scores = [0.9, 0.9]
scores = np.asarray(scores)
classes = [1, 1]
classes = np.asarray(classes)
self.assertEqualAnyOrder(nms(boxes, scores, classes),
np.asarray([0, 1]))
def test_no_boxes_returns_empty(self):
empty = np.array([])
self.assertEqual(0, nms(empty, empty, empty).size)
) # must run predict method with this flag to get a confidence score
score = -raw_prediction[1][0][
0] # a lower value of the returned score represents high confidence, so we invert the score to make it more intuitive & suitable for non-max supression
if score > SVM_SCORE_THRESHOLD: # If the classification produces a positive result that passes a certain confidence threshold, we add the rectangle's corner coordinates and confidence score to a list of positive detections
h, w = roi.shape
scale = gray_img.shape[0] / \
float(resized.shape[0])
pos_rects.append([
int(
x * scale
), # rescale the coorindate back to the original images' scale
int(y * scale),
int((x + w) * scale),
int((y + h) * scale),
score
]) #
# apply non-max suppression to get the highest-scoring rectangles among the ones that overlap due to sliding window
pos_rects = nms(np.array(pos_rects), NMS_OVERLAP_THRESHOLD)
# draw the rectangles and scores
for x0, y0, x1, y1, score in pos_rects:
cv2.rectangle(img, (int(x0), int(y0)), (int(x1), int(y1)),
(0, 255, 255), 2)
text = '%.2f' % score
cv2.putText(img, text, (int(x0), int(y0) - 20),
cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 255), 2)
cv2.imshow(test_img_path, img)
cv2.waitKey()
cv2.destroyAllWindows()
def test_same_boxes_same_classes_samescores(self): """ everything is the same so it should return either of the boxes """ boxes = np.asarray([(1,2,3,4),(1,2,3,4)]) classes = np.asarray([1,1]) scores = np.asarray([0.5,0.5]) self.assertEither(nms(boxes, scores, classes), np.array([(1,2,3,4)]))