def gen_noise_samples(self, imgwh, noise_type, num, **kwargs):
center_xy = self.xy + self.wh * 0.5
mean_wh = sum(self.wh) / 2.0
gaussian_translation_f = kwargs.get('gaussian_translation_f', 0.1)
uniform_translation_f = kwargs.get('uniform_translation_f', 1)
uniform_scale_f = kwargs.get('uniform_scale_f', 10)
samples = []
if noise_type == 'whole':
#print('{}, {}, {}, {}'.format(imgwh.x, imgwh.y, self.wh.x, self.wh.y))
grid_x = range(self.wh.x // 2, imgwh.x - self.wh.x // 2,
self.wh.x // 5)
grid_y = range(self.wh.y // 2, imgwh.y - self.wh.y // 2,
self.wh.y // 5)
samples_tmp = []
#print(grid_x, ":", grid_y)
for dx, dy, ds in itertools.product(grid_x, grid_y,
range(-5, 5, 1)):
box = BoundingBox(dx, dy, self.wh.x * (1.05**ds),
self.wh.y * (1.05**ds))
box.fit_image(imgwh)
samples_tmp.append(box)
#print('len samples_temp: ', len(samples_tmp))
for _ in range(num):
samples.append(random.choice(samples_tmp))
else:
for _ in range(num):
if noise_type == 'gaussian':
dx = gaussian_translation_f * mean_wh * minmax(
0.5 * random.normalvariate(0, 1), -1, 1)
dy = gaussian_translation_f * mean_wh * minmax(
0.5 * random.normalvariate(0, 1), -1, 1)
dwh = 1.05**(
3 * minmax(0.5 * random.normalvariate(0, 1), -1, 1))
elif noise_type == 'uniform':
dx = uniform_translation_f * mean_wh * random.uniform(
-1.0, 1.0)
dy = uniform_translation_f * mean_wh * random.uniform(
-1.0, 1.0)
dwh = 1.05**(uniform_scale_f * random.uniform(-1.0, 1.0))
else:
raise
new_cxy = center_xy + (dx, dy)
new_wh = self.wh * dwh
box = BoundingBox(new_cxy.x - new_wh.x / 2.0,
new_cxy.y - new_wh.y / 2.0, new_wh.x,
new_wh.y)
box.fit_image(imgwh)
samples.append(box)
return samples
def fit_image(self, imgwh, margin=0):
# Make sure the bbox is within the imgwh
# margin is the max padding size
if not isinstance(imgwh, Coordinate):
imgwh = Coordinate(*imgwh)
x = minmax(self.xy.x, -margin/2, imgwh.x-margin/2)
y = minmax(self.xy.y, -margin/2, imgwh.y-margin/2)
w = minmax(self.wh.x, margin, imgwh.x+margin/2-x)
h = minmax(self.wh.y, margin, imgwh.y+margin/2-y)
return BoundingBox(x,y,w,h)