def masked_transfer_color(f0, f1, dest, flow, imask, row, col, t=0.5):
""" Find and set the transfer color for pixel [row][cow] """
# for now, just defauly to using the value at f1:
h = dest.shape[0]
w = dest.shape[1]
u = flow[row][col]
ux = u[0]/w
uy = -u[1]/h
(x, y) = pix.image_to_cartesian(flow, col, row)
xp0 = x - t*ux
yp0 = y - t*uy
xp1 = x + t*ux
yp1 = y + t*uy
(xi0, yi0) = pix.cartesian_to_index(flow, xp0, yp0)
(xi1, yi1) = pix.cartesian_to_index(flow, xp1, yp1)
if (pix.check_indices(flow, xi0, yi0) and
pix.check_indices(flow, xi1, yi1)):
f0vis = imask[row][col][0]
f1vis = imask[row][col][1]
if f0vis >= 1 and f1vis >= 1:
dest[row][col] = (f0[yi0][xi0])/2 + (f1[yi1][xi1])/2
elif f1vis >= 1:
dest[row][col] = f1[yi1][xi1]
else:
dest[row][col] = f0[yi0][xi0]
def find_occlusion(forward, backward, interp, imask, row, col, t=0.5):
h = interp.shape[0]
w = interp.shape[1]
im = interp[row][col]
ux = im[0]/w
uy = -im[1]/h
(x, y) = pix.image_to_cartesian(interp, col, row)
xp0 = x - t*ux
yp0 = y - t*uy
xp1 = x + t*ux
yp1 = y + t*uy
# Get each motion vector
(xi0, yi0) = pix.cartesian_to_index(forward, xp0, yp0)
(xi1, yi1) = pix.cartesian_to_index(backward, xp1, yp1)
if (pix.check_indices(forward, xi0, yi0) and
pix.check_indices(backward, xi1, yi1)):
fvec = forward[yi0][xi0]
bvec = -1*backward[yi1][xi1]
d0 = pix.flowdist(fvec, im)
d1 = pix.flowdist(bvec, im)
diff = abs(d0-d1)
if d0 <= d1:
imask[row][col][0] = 1
if diff < 0.1:
imask[row][col][1] = 1
else:
imask[row][col][1] = 1
if diff < 0.1:
imask[row][col][0] = 1
else:
# Default to using first frame
imask[row][col][0] = 1
def find_occlusion(forward, backward, interp, imask, row, col, t=0.5):
h = interp.shape[0]
w = interp.shape[1]
im = interp[row][col]
ux = im[0] / w
uy = -im[1] / h
(x, y) = pix.image_to_cartesian(interp, col, row)
xp0 = x - t * ux
yp0 = y - t * uy
xp1 = x + t * ux
yp1 = y + t * uy
# Get each motion vector
(xi0, yi0) = pix.cartesian_to_index(forward, xp0, yp0)
(xi1, yi1) = pix.cartesian_to_index(backward, xp1, yp1)
if (pix.check_indices(forward, xi0, yi0)
and pix.check_indices(backward, xi1, yi1)):
fvec = forward[yi0][xi0]
bvec = -1 * backward[yi1][xi1]
d0 = pix.flowdist(fvec, im)
d1 = pix.flowdist(bvec, im)
diff = abs(d0 - d1)
if d0 <= d1:
imask[row][col][0] = 1
if diff < 0.1:
imask[row][col][1] = 1
else:
imask[row][col][1] = 1
if diff < 0.1:
imask[row][col][0] = 1
else:
# Default to using first frame
imask[row][col][0] = 1
def masked_transfer_color(f0, f1, dest, flow, imask, row, col, t=0.5):
""" Find and set the transfer color for pixel [row][cow] """
# for now, just defauly to using the value at f1:
h = dest.shape[0]
w = dest.shape[1]
u = flow[row][col]
ux = u[0] / w
uy = -u[1] / h
(x, y) = pix.image_to_cartesian(flow, col, row)
xp0 = x - t * ux
yp0 = y - t * uy
xp1 = x + t * ux
yp1 = y + t * uy
(xi0, yi0) = pix.cartesian_to_index(flow, xp0, yp0)
(xi1, yi1) = pix.cartesian_to_index(flow, xp1, yp1)
if (pix.check_indices(flow, xi0, yi0)
and pix.check_indices(flow, xi1, yi1)):
f0vis = imask[row][col][0]
f1vis = imask[row][col][1]
if f0vis >= 1 and f1vis >= 1:
dest[row][col] = (f0[yi0][xi0]) / 2 + (f1[yi1][xi1]) / 2
elif f1vis >= 1:
dest[row][col] = f1[yi1][xi1]
else:
dest[row][col] = f0[yi0][xi0]
def find_occlusion(forward, backward, interp, imask, row, col, t=0.5):
im = interp[row][col]
xp0 = col - t * im[0] + 0.5
yp0 = row - t * im[1] + 0.5
xp1 = im[0] + xp0
yp1 = im[1] + yp0
# Get each motion vector
xi0 = int(xp0)
yi0 = int(yp0)
xi1 = int(xp1)
yi1 = int(yp1)
if (pix.check_indices(forward, xi0, yi0)
and pix.check_indices(backward, xi1, yi1)):
fvec = forward[yi0][xi0]
bvec = -backward[yi1][xi1]
d0 = pix.flowdist(fvec, im)
d1 = pix.flowdist(bvec, im)
diff = abs(d0 - d1)
if diff < 0.1:
imask[row][col][:] = 1
else:
if d0 <= d1:
imask[row][col][0] = 1
else:
imask[row][col][1] = 1
else:
# Default to using first frame
imask[row][col][0] = 1
def masked_transfer_color(f0, f1, forward, backward, interp, t=0.5):
""" Find and set the transfer color for pixel [row][cow] """
# for now, just defauly to using the value at f1:
dest = np.copy(f0)
h = dest.shape[0]
w = dest.shape[1]
for row in tqdm(range(h), position=True, desc="color transfer"):
for col in range(w):
u = interp[row][col]
xp0 = col - t * u[0] + 0.5
yp0 = row - t * u[1] + 0.5
xp1 = u[0] + xp0
yp1 = u[1] + yp0
xi0 = int(xp0)
yi0 = int(yp0)
xi1 = int(xp1)
yi1 = int(yp1)
if (pix.check_indices(interp, xi0, yi0)
and pix.check_indices(interp, xi1, yi1)):
fvec = forward[yi0][xi0]
bvec = -backward[yi1][xi1]
d0 = pix.flowdist(fvec, u)
d1 = pix.flowdist(bvec, u)
diff = abs(d0 - d1)
if diff < 0.1:
dest[row][col] = (f0[yi0][xi0] + f1[yi1][xi1]) * 0.5
elif d0 > d1:
dest[row][col] = f1[yi1][xi1]
else:
dest[row][col] = f0[yi0][xi0]
return dest
def color_transfer_weighted(frame, flow0, flowt, dest,
row, col, ws, t=0.5):
""" Transfer coloring using the Shiratori et al.
weighted blending technique. In this case [row][col]
refers to the f0 pixel, not the output pixel."""
h = frame.shape[0]
w = frame.shape[1]
# Get the flow for previous timestwp
q_flow = flow0[row][col]
# Get the previous frame cartesian coords
(qx, qy) = pix.image_to_cartesian(frame, col, row)
ux = q_flow[0]/w
uy = -q_flow[1]/h
# These are the target coordinates in the interpolated
# image. We need to find all frames within 0.5 pixels
# of this spot
px = qx + t*ux
py = qy + t*uy
# These are the indices of all the target pixels. We will
# add to each one...
ipixels = pix.splat(dest, px, py)
if ipixels:
for p in ipixels:
if (pix.check_indices(dest, p[0], p[1])):
p_flow = flowt[p[1]][p[0]]
r = pix.flowdist(p_flow, q_flow)
add_weighted_color(frame[row][col], p, dest,
px, py, ws, r, t=0.5)
def transfer_color(f0, f1, dest, flow, row, col, t=0.5):
""" Find and set the transfer color for pixel [row][cow] """
# for now, just defauly to using the value at f1:
h = dest.shape[0]
w = dest.shape[1]
u = flow[row][col]
ux = u[0]
uy = u[1]
xp0 = col - t * ux
yp0 = row - t * uy
xp1 = col + (1 - t) * ux
yp1 = row + (1 - t) * uy
xi0 = int(round(xp0))
yi0 = int(round(yp0))
xi1 = int(round(xp1))
yi1 = int(round(yp1))
if (pix.check_indices(flow, xi0, yi0)
and pix.check_indices(flow, xi1, yi1)):
#dest[row][col] = (f0[yi0][xi0])/2 + (f1[yi1][xi1])/2
dest[row][col] = f0[yi0][xi0]
def find_occlusions(forward, backward, interp, t=0.5):
""" Generate the occlusion mask for the interpolated frame. """
h = interp.shape[0]
w = interp.shape[1]
imask = np.zeros((h, w, 2), dtype='uint8')
# Default to using first frame
imask[:, :, 0] = 1
for row in tqdm(range(h), position=True, desc="generating occlusion mask"):
for col in range(w):
im = interp[row][col]
xp0 = col - t * im[0] + 0.5
yp0 = row - t * im[1] + 0.5
xp1 = im[0] + xp0
yp1 = im[1] + yp0
# Get each motion vector
xi0 = int(xp0)
yi0 = int(yp0)
xi1 = int(xp1)
yi1 = int(yp1)
if (pix.check_indices(forward, xi0, yi0)
and pix.check_indices(backward, xi1, yi1)):
fvec = forward[yi0][xi0]
bvec = -backward[yi1][xi1]
d0 = pix.flowdist(fvec, im)
d1 = pix.flowdist(bvec, im)
diff = abs(d0 - d1)
if diff < 0.1:
imask[row][col][1] = 1
else:
if d0 > d1:
imask[row][col][0] = 0
imask[row][col][1] = 1
return imask
def color_transfer_weighted(frame, flow0, flowt, dest, row, col, ws, t=0.5):
""" Transfer coloring using the Shiratori et al.
weighted blending technique. In this case [row][col]
refers to the f0 pixel, not the output pixel."""
h = frame.shape[0]
w = frame.shape[1]
# Get the flow for previous timestwp
q_flow = flow0[row][col]
# Get the previous frame cartesian coords
(qx, qy) = pix.image_to_cartesian(frame, col, row)
ux = q_flow[0] / w
uy = -q_flow[1] / h
# These are the target coordinates in the interpolated
# image. We need to find all frames within 0.5 pixels
# of this spot
px = qx + t * ux
py = qy + t * uy
# These are the indices of all the target pixels. We will
# add to each one...
ipixels = pix.splat(dest, px, py)
if ipixels:
for p in ipixels:
if (pix.check_indices(dest, p[0], p[1])):
p_flow = flowt[p[1]][p[0]]
r = pix.flowdist(p_flow, q_flow)
add_weighted_color(frame[row][col],
p,
dest,
px,
py,
ws,
r,
t=0.5)