def test_pytorch_grid():
c_dst = np.array([
[0., 0],
[1., 0],
[1, 1],
[0, 1],
], dtype=np.float32)
c_src = np.array([
[10., 10],
[20., 10],
[20, 20],
[10, 20],
],
dtype=np.float32) / 40.
theta = tps.tps_theta_from_points(c_src, c_dst)
theta_r = tps.tps_theta_from_points(c_src, c_dst, reduced=True)
np_grid = tps.tps_grid(theta, c_dst, (20, 20))
np_grid_r = tps.tps_grid(theta_r, c_dst, (20, 20))
pth_theta = torch.tensor(theta).unsqueeze(0)
pth_grid = tps.torch.tps_grid(pth_theta, torch.tensor(c_dst),
(1, 1, 20, 20)).squeeze().numpy()
pth_grid = (pth_grid + 1) / 2 # convert [-1,1] range to [0,1]
pth_theta_r = torch.tensor(theta_r).unsqueeze(0)
pth_grid_r = tps.torch.tps_grid(pth_theta_r, torch.tensor(c_dst),
(1, 1, 20, 20)).squeeze().numpy()
pth_grid_r = (pth_grid_r + 1) / 2 # convert [-1,1] range to [0,1]
assert_allclose(np_grid, pth_grid)
assert_allclose(np_grid_r, pth_grid_r)
assert_allclose(np_grid_r, np_grid)
def tps_transform(img, dshape=None):
while True:
point1 = round(random.uniform(0.3, 0.7), 2)
point2 = round(random.uniform(0.3, 0.7), 2)
range_1 = round(random.uniform(-0.25, 0.25), 2)
range_2 = round(random.uniform(-0.25, 0.25), 2)
if math.isclose(point1 + range_1, point2 + range_2):
continue
else:
break
c_src = np.array([
[0.0, 0.0],
[1., 0],
[1, 1],
[0, 1],
[point1, point1],
[point2, point2],
])
c_dst = np.array([
[0., 0],
[1., 0],
[1, 1],
[0, 1],
[point1 + range_1, point1 + range_1],
[point2 + range_2, point2 + range_2],
])
dshape = dshape or img.shape
theta = tps.tps_theta_from_points(c_src, c_dst, reduced=True)
grid = tps.tps_grid(theta, c_dst, dshape)
mapx, mapy = tps.tps_grid_to_remap(grid, img.shape)
return cv2.remap(img, mapx, mapy, cv2.INTER_CUBIC)
def warp_dual_cv(img, mask, c_src, c_dst):
dshape = img.shape
theta = tps.tps_theta_from_points(c_src, c_dst, reduced=True)
grid = tps.tps_grid(theta, c_dst, dshape)
mapx, mapy = tps.tps_grid_to_remap(grid, img.shape)
return cv2.remap(img, mapx, mapy,
cv2.INTER_LINEAR), cv2.remap(mask, mapx, mapy,
cv2.INTER_NEAREST)
def test_numpy_densegrid():
# enlarges a small rectangle to full view
import cv2
img = np.zeros((40, 40), dtype=np.uint8)
img[10:21, 10:21] = 255
c_dst = np.array([
[0., 0],
[1., 0],
[1, 1],
[0, 1],
])
c_src = np.array([
[10., 10],
[20., 10],
[20, 20],
[10, 20],
]) / 40.
theta = tps.tps_theta_from_points(c_src, c_dst)
theta_r = tps.tps_theta_from_points(c_src, c_dst, reduced=True)
grid = tps.tps_grid(theta, c_dst, (20, 20))
grid_r = tps.tps_grid(theta_r, c_dst, (20, 20))
mapx, mapy = tps.tps_grid_to_remap(grid, img.shape)
warped = cv2.remap(img, mapx, mapy, cv2.INTER_CUBIC)
assert img.min() == 0.
assert img.max() == 255.
assert warped.shape == (20, 20)
assert warped.min() == 255.
assert warped.max() == 255.
assert np.linalg.norm(grid.reshape(-1, 2) - grid_r.reshape(-1, 2)) < 1e-3
def get_tps_thetas(self, num_images):
c_src = np.array([
[0.0, 0.0],
[1., 0],
[1, 1],
[0, 1],
[0.2, 0.3],
[0.6, 0.7],
])
theta_list = []
dst_list = []
for i in range(num_images):
c_dst = c_src + np.random.uniform(-1, 1, (6, 2)) / 20
theta = tps.tps_theta_from_points(c_src, c_dst)
theta_list.append(torch.from_numpy(theta).unsqueeze(0))
dst_list.append(torch.from_numpy(c_dst).unsqueeze(0))
theta = torch.cat(theta_list, dim=0).float()
dst = torch.cat(dst_list, dim=0).float()
return theta, dst
def warp_image_cv(img, c_src, c_dst, dshape=None):
dshape = dshape or img.shape
theta = tps.tps_theta_from_points(c_src, c_dst, reduced=True)
grid = tps.tps_grid(theta, c_dst, dshape)
mapx, mapy = tps.tps_grid_to_remap(grid, img.shape)
return cv2.remap(img, mapx, mapy, cv2.INTER_CUBIC)