github/DeepFaceLab
1
0
Fork 0
mirror of https://github.com/iperov/DeepFaceLab.git synced 2025-07-05 20:42:11 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

_

Browse source
This commit is contained in:
iperov 2021-08-12 14:28:36 +04:00
parent c8e6f23a31
commit a299324166
1 changed files with 39 additions and 70 deletions
Download patch file Download diff file Expand all files Collapse all files

109
core/imagelib/warp.py
View file

@ -3,97 +3,66 @@ import numpy.linalg as npla
import cv2
from core import randomex
def mls_rigid_deformation(vy, vx, src_pts, dst_pts, alpha=1.0, eps=1e-8):
dst_pts = dst_pts[..., ::-1].astype(np.int16)
src_pts = src_pts[..., ::-1].astype(np.int16)
def mls_rigid_deformation(vy, vx, p, q, alpha=1.0, eps=1e-8):
""" Rigid deformation
Parameters
----------
vx, vy: ndarray
coordinate grid, generated by np.meshgrid(gridX, gridY)
p: ndarray
an array with size [n, 2], original control points
q: ndarray
an array with size [n, 2], final control points
alpha: float
parameter used by weights
eps: float
epsilon
Return
------
A deformed image.
"""
# Change (x, y) to (row, col)
q = np.ascontiguousarray(q[:, [1, 0]].astype(np.int16))
p = np.ascontiguousarray(p[:, [1, 0]].astype(np.int16))
src_pts, dst_pts = dst_pts, src_pts
# Exchange p and q and hence we transform destination pixels to the corresponding source pixels.
p, q = q, p
grow = vx.shape[0]
gcol = vx.shape[1]
ctrls = src_pts.shape[0]
grow = vx.shape[0] # grid rows
gcol = vx.shape[1] # grid cols
ctrls = p.shape[0] # control points
reshaped_p = src_pts.reshape(ctrls, 2, 1, 1)
reshaped_v = np.vstack((vx.reshape(1, grow, gcol), vy.reshape(1, grow, gcol)))
# Compute
reshaped_p = p.reshape(ctrls, 2, 1, 1) # [ctrls, 2, 1, 1]
reshaped_v = np.vstack((vx.reshape(1, grow, gcol), vy.reshape(1, grow, gcol))) # [2, grow, gcol]
w = 1.0 / (np.sum((reshaped_p - reshaped_v).astype(np.float32) ** 2, axis=1) + eps) ** alpha # [ctrls, grow, gcol]
w /= np.sum(w, axis=0, keepdims=True) # [ctrls, grow, gcol]
w = 1.0 / (np.sum((reshaped_p - reshaped_v).astype(np.float32) ** 2, axis=1) + eps) ** alpha
w /= np.sum(w, axis=0, keepdims=True)
pstar = np.zeros((2, grow, gcol), np.float32)
for i in range(ctrls):
pstar += w[i] * reshaped_p[i] # [2, grow, gcol]
pstar += w[i] * reshaped_p[i]
vpstar = reshaped_v - pstar # [2, grow, gcol]
reshaped_vpstar = vpstar.reshape(2, 1, grow, gcol) # [2, 1, grow, gcol]
neg_vpstar_verti = vpstar[[1, 0],...] # [2, grow, gcol]
neg_vpstar_verti[1,...] = -neg_vpstar_verti[1,...]
reshaped_neg_vpstar_verti = neg_vpstar_verti.reshape(2, 1, grow, gcol) # [2, 1, grow, gcol]
mul_right = np.concatenate((reshaped_vpstar, reshaped_neg_vpstar_verti), axis=1) # [2, 2, grow, gcol]
reshaped_mul_right = mul_right.reshape(2, 2, grow, gcol) # [2, 2, grow, gcol]
vpstar = reshaped_v - pstar
# Calculate q
reshaped_q = q.reshape((ctrls, 2, 1, 1)) # [ctrls, 2, 1, 1]
reshaped_mul_right = np.concatenate((vpstar[:,None,...],
np.concatenate((vpstar[1:2,None,...],-vpstar[0:1,None,...]), 0)
), axis=1).transpose(2, 3, 0, 1)
reshaped_q = dst_pts.reshape((ctrls, 2, 1, 1))
qstar = np.zeros((2, grow, gcol), np.float32)
for i in range(ctrls):
qstar += w[i] * reshaped_q[i] # [2, grow, gcol]
qstar += w[i] * reshaped_q[i]
temp = np.zeros((grow, gcol, 2), np.float32)
for i in range(ctrls):
phat = reshaped_p[i] - pstar # [2, grow, gcol]
reshaped_phat = phat.reshape(1, 2, grow, gcol) # [1, 2, grow, gcol]
reshaped_w = w[i].reshape(1, 1, grow, gcol) # [1, 1, grow, gcol]
neg_phat_verti = phat[[1, 0]] # [2, grow, gcol]
neg_phat_verti[1] = -neg_phat_verti[1]
reshaped_neg_phat_verti = neg_phat_verti.reshape(1, 2, grow, gcol) # [1, 2, grow, gcol]
mul_left = np.concatenate((reshaped_phat, reshaped_neg_phat_verti), axis=0) # [2, 2, grow, gcol]
A = np.matmul((reshaped_w * mul_left).transpose(2, 3, 0, 1),
reshaped_mul_right.transpose(2, 3, 0, 1)) # [grow, gcol, 2, 2]
phat = reshaped_p[i] - pstar
qhat = reshaped_q[i] - qstar
qhat = reshaped_q[i] - qstar # [2, grow, gcol]
reshaped_qhat = qhat.reshape(1, 2, grow, gcol).transpose(2, 3, 0, 1) # [grow, gcol, 1, 2]
temp += np.matmul(qhat.reshape(1, 2, grow, gcol).transpose(2, 3, 0, 1),
np.matmul( ( w[None, i:i+1,...] *
np.concatenate((phat.reshape(1, 2, grow, gcol),
np.concatenate( (phat[None,1:2], -phat[None,0:1]), 1 )), 0)
).transpose(2, 3, 0, 1), reshaped_mul_right
)
).reshape(grow, gcol, 2)
# Get final image transfomer -- 3-D array
temp += np.matmul(reshaped_qhat, A).reshape(grow, gcol, 2) # [grow, gcol, 2]
temp = temp.transpose(2, 0, 1)
temp = temp.transpose(2, 0, 1) # [2, grow, gcol]
normed_temp = np.linalg.norm(temp, axis=0, keepdims=True) # [1, grow, gcol]
normed_vpstar = np.linalg.norm(vpstar, axis=0, keepdims=True) # [1, grow, gcol]
normed_temp = np.linalg.norm(temp, axis=0, keepdims=True)
normed_vpstar = np.linalg.norm(vpstar, axis=0, keepdims=True)
nan_mask = normed_temp[0]==0
transformers = np.true_divide(temp, normed_temp, out=np.zeros_like(temp), where= ~nan_mask) * normed_vpstar + qstar
# fix nan values
transformers = np.true_divide(temp, normed_temp, out=np.zeros_like(temp), where= ~nan_mask) * normed_vpstar + qstar
nan_mask_flat = np.flatnonzero(nan_mask)
nan_mask_anti_flat = np.flatnonzero(~nan_mask)
transformers[0][nan_mask] = np.interp(nan_mask_flat, nan_mask_anti_flat, transformers[0][~nan_mask])
transformers[1][nan_mask] = np.interp(nan_mask_flat, nan_mask_anti_flat, transformers[1][~nan_mask])
return transformers
return transformers
def gen_pts(W, H, rnd_state=None):