-
Notifications
You must be signed in to change notification settings - Fork 8
/
bilateral_solver.py
194 lines (165 loc) · 7.1 KB
/
bilateral_solver.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
"""
The Fast Bilateral Solver for Optical Flow
https://github.com/poolio/bilateral_solver
"""
import os
import numpy as np
from scipy.sparse import csr_matrix
from scipy.sparse import diags
from scipy.sparse.linalg import cg
RGB_TO_YUV = np.array([
[ 0.299, 0.587, 0.114],
[-0.168736, -0.331264, 0.5],
[ 0.5, -0.418688, -0.081312]])
YUV_TO_RGB = np.array([
[1.0, 0.0, 1.402],
[1.0, -0.34414, -0.71414],
[1.0, 1.772, 0.0]])
YUV_OFFSET = np.array([0, 128.0, 128.0]).reshape(1, 1, -1)
MAX_VAL = 255.0
def rgb2yuv(im):
return np.tensordot(im, RGB_TO_YUV, ([2], [1])) + YUV_OFFSET
def yuv2rgb(im):
return np.tensordot(im.astype(float) - YUV_OFFSET, YUV_TO_RGB, ([2], [1]))
def get_valid_idx(valid, candidates):
"""Find which values are present in a list and where they are located"""
locs = np.searchsorted(valid, candidates)
# Handle edge case where the candidate is larger than all valid values
locs = np.clip(locs, 0, len(valid) - 1)
# Identify which values are actually present
valid_idx = np.flatnonzero(valid[locs] == candidates)
locs = locs[valid_idx]
return valid_idx, locs
class BilateralGrid(object):
def __init__(self, im, sigma_spatial=12, sigma_luma=4, sigma_chroma=4):
im_yuv = rgb2yuv(im)
# Compute 5-dimensional XYLUV bilateral-space coordinates
Iy, Ix = np.mgrid[:im.shape[0], :im.shape[1]]
x_coords = (Ix / sigma_spatial).astype(int)
y_coords = (Iy / sigma_spatial).astype(int)
luma_coords = (im_yuv[..., 0] /sigma_luma).astype(int)
chroma_coords = (im_yuv[..., 1:] / sigma_chroma).astype(int)
coords = np.dstack((x_coords, y_coords, luma_coords, chroma_coords))
coords_flat = coords.reshape(-1, coords.shape[-1])
self.npixels, self.dim = coords_flat.shape
# Hacky "hash vector" for coordinates,
# Requires all scaled coordinates be < MAX_VAL
self.hash_vec = (MAX_VAL**np.arange(self.dim))
# Construct S and B matrix
self._compute_factorization(coords_flat)
def _compute_factorization(self, coords_flat):
# Hash each coordinate in grid to a unique value
hashed_coords = self._hash_coords(coords_flat)
unique_hashes, unique_idx, idx = np.unique(hashed_coords, return_index=True, return_inverse=True)
# Identify unique set of vertices
unique_coords = coords_flat[unique_idx]
self.nvertices = len(unique_coords)
# Construct sparse splat matrix that maps from pixels to vertices
self.S = csr_matrix((np.ones(self.npixels), (idx, np.arange(self.npixels))))
# Construct sparse blur matrices.
# Note that these represent [1 0 1] blurs, excluding the central element
self.blurs = []
for d in xrange(self.dim):
blur = 0.0
for offset in (-1, 1):
offset_vec = np.zeros((1, self.dim))
offset_vec[:, d] = offset
neighbor_hash = self._hash_coords(unique_coords + offset_vec)
valid_coord, idx = get_valid_idx(unique_hashes, neighbor_hash)
blur = blur + csr_matrix((np.ones((len(valid_coord),)),
(valid_coord, idx)),
shape=(self.nvertices, self.nvertices))
self.blurs.append(blur)
def _hash_coords(self, coord):
"""Hacky function to turn a coordinate into a unique value"""
return np.dot(coord.reshape(-1, self.dim), self.hash_vec)
def splat(self, x):
return self.S.dot(x)
def slice(self, y):
return self.S.T.dot(y)
def blur(self, x):
"""Blur a bilateral-space vector with a 1 2 1 kernel in each dimension"""
assert x.shape[0] == self.nvertices
out = 2 * self.dim * x
for blur in self.blurs:
out = out + blur.dot(x)
return out
def filter(self, x):
"""Apply bilateral filter to an input x"""
return self.slice(self.blur(self.splat(x))) / self.slice(self.blur(self.splat(np.ones_like(x))))
def bistochastize(grid, maxiter=10):
"""Compute diagonal matrices to bistochastize a bilateral grid"""
m = grid.splat(np.ones(grid.npixels))
n = np.ones(grid.nvertices)
for i in xrange(maxiter):
n = np.sqrt(n * m / grid.blur(n))
# Correct m to satisfy the assumption of bistochastization regardless
# of how many iterations have been run.
m = n * grid.blur(n)
Dm = diags(m, 0)
Dn = diags(n, 0)
return Dn, Dm
class BilateralSolver(object):
def __init__(self, grid, params):
self.grid = grid
self.params = params
self.Dn, self.Dm = bistochastize(grid)
def solve(self, x, w):
# Check that w is a vector or a nx1 matrix
if w.ndim == 2:
assert(w.shape[1] == 1)
elif w.dim == 1:
w = w.reshape(w.shape[0], 1)
A_smooth = (self.Dm - self.Dn.dot(self.grid.blur(self.Dn)))
w_splat = self.grid.splat(w)
A_data = diags(w_splat[:,0], 0)
A = self.params["lam"] * A_smooth + A_data
xw = x * w
b = self.grid.splat(xw)
# Use simple Jacobi preconditioner
A_diag = np.maximum(A.diagonal(), self.params["A_diag_min"])
M = diags(1 / A_diag, 0)
# Flat initialization
y0 = self.grid.splat(xw) / w_splat
yhat = np.empty_like(y0)
for d in xrange(x.shape[-1]):
yhat[..., d], info = cg(A, b[..., d], x0=y0[..., d], M=M, maxiter=self.params["cg_maxiter"], tol=self.params["cg_tol"])
xhat = self.grid.slice(yhat)
return xhat
def bil_solv_flo(reference, flow, conf_x, conf_y, grid_params, bs_params):
"""Compute Bilateral Solver for each direction (x,y) in flow"""
# images tranform
reference = np.array(reference*255).astype(np.uint8)
# split directions
x = flow[:, :, 0]
y = flow[:, :, 1]
# move flow to positive and then normalize to [0,1]
min_x= abs(np.amin(x))
min_y= abs(np.amin(y))
max_x = np.amax(x+min_x)
max_y = np.amax(y+min_y)
target_0 = (x + min_x)/max_x
target_1 = (y + min_y)/max_y
# check shapes
im_shape = reference.shape[:2]
assert(im_shape[0] == flow.shape[0])
assert(im_shape[1] == flow.shape[1])
assert(im_shape[0] == conf_x.shape[0])
assert(im_shape[1] == conf_x.shape[1])
assert(conf_x.shape == conf_y.shape)
# calc bilateral grid
grid = BilateralGrid(reference, **grid_params)
# reshape to vector
t_0 = target_0.reshape(-1, 1).astype(np.double)
t_1 = target_1.reshape(-1, 1).astype(np.double)
conf_x[np.where(conf_x == 0)] = 0.0001
conf_y[np.where(conf_y == 0)] = 0.0001
cx = conf_x.reshape(-1, 1).astype(np.double)
cy = conf_y.reshape(-1, 1).astype(np.double)
#Apply a bilateral solver
solved_flow_x = BilateralSolver(grid, bs_params).solve(t_0, cx).reshape(im_shape)
solved_flow_y = BilateralSolver(grid, bs_params).solve(t_1, cy).reshape(im_shape)
solved_flow = np.zeros(flow.shape)
solved_flow[:, :, 0] = solved_flow_x*max_x - min_x
solved_flow[:, :, 1] = solved_flow_y*max_y - min_y
return solved_flow.astype(np.float32)