-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathrescore.py
116 lines (73 loc) · 4.43 KB
/
rescore.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
import sys, Image
from numpy import *
import scipy.ndimage
def score_from_autocorr(img0, img1, corres):
# Code by Philippe Weinzaepfel
# Compute autocorrelation
# parameters
sigma_image = 0.8 # for the gaussian filter applied to images before computing derivatives
sigma_matrix = 3.0 # for the integration gaussian filter
derivfilter = array([-0.5,0,0.5]) # function to compute the derivatives
# smooth_images
tmp = scipy.ndimage.filters.gaussian_filter1d(img0.astype(float32), sigma_image, axis=0, order=0, mode='nearest')
img0_smooth = scipy.ndimage.filters.gaussian_filter1d(tmp, sigma_image, axis=1, order=0, mode='nearest')
# compute the derivatives
img0_dx = scipy.ndimage.filters.convolve1d(img0_smooth, derivfilter, axis=0, mode='nearest')
img0_dy = scipy.ndimage.filters.convolve1d(img0_smooth, derivfilter, axis=1, mode='nearest')
# compute the auto correlation matrix
dx2 = sum(img0_dx*img0_dx,axis=2)
dxy = sum(img0_dx*img0_dy,axis=2)
dy2 = sum(img0_dy*img0_dy,axis=2)
# integrate it
tmp = scipy.ndimage.filters.gaussian_filter1d(dx2, sigma_matrix, axis=0, order=0, mode='nearest')
dx2_smooth = scipy.ndimage.filters.gaussian_filter1d(tmp, sigma_matrix, axis=1, order=0, mode='nearest')
tmp = scipy.ndimage.filters.gaussian_filter1d(dxy, sigma_matrix, axis=0, order=0, mode='nearest')
dxy_smooth = scipy.ndimage.filters.gaussian_filter1d(tmp, sigma_matrix, axis=1, order=0, mode='nearest')
tmp = scipy.ndimage.filters.gaussian_filter1d(dy2, sigma_matrix, axis=0, order=0, mode='nearest')
dy2_smooth = scipy.ndimage.filters.gaussian_filter1d(tmp, sigma_matrix, axis=1, order=0, mode='nearest')
# compute minimal eigenvalues: it is done by computing (dx2+dy2)/2 - sqrt( ((dx2+dy2)/2)^2 + (dxy)^2 - dx^2*dy^2)
tmp = 0.5*(dx2_smooth+dy2_smooth)
small_eigen = tmp - sqrt( maximum(0,tmp*tmp + dxy_smooth*dxy_smooth - dx2_smooth*dy2_smooth)) # the numbers can be negative in practice due to rounding errors
large_eigen = tmp + sqrt( maximum(0,tmp*tmp + dxy_smooth*dxy_smooth - dx2_smooth*dy2_smooth))
# Compute weight as flow score: preparing variable
#parameters
sigma_image = 0.8 # gaussian applied to images
derivfilter = array([1.0,-8.0,0.0,8.0,-1.0])/12.0 # filter to compute the derivatives
sigma_score = 50.0 # gaussian to convert dist to score
mul_coef = 10.0 # multiplicative coefficients
# smooth images
tmp = scipy.ndimage.filters.gaussian_filter1d(img0.astype(float32), sigma_image, axis=0, order=0, mode='nearest')
img0_smooth = scipy.ndimage.filters.gaussian_filter1d(tmp, sigma_image, axis=1, order=0, mode='nearest')
tmp = scipy.ndimage.filters.gaussian_filter1d(img1.astype(float32), sigma_image, axis=0, order=0, mode='nearest')
img1_smooth = scipy.ndimage.filters.gaussian_filter1d(tmp, sigma_image, axis=1, order=0, mode='nearest')
# compute derivatives
img0_dx = scipy.ndimage.filters.convolve1d(img0_smooth, derivfilter, axis=0, mode='nearest')
img0_dy = scipy.ndimage.filters.convolve1d(img0_smooth, derivfilter, axis=1, mode='nearest')
img1_dx = scipy.ndimage.filters.convolve1d(img1_smooth, derivfilter, axis=0, mode='nearest')
img1_dy = scipy.ndimage.filters.convolve1d(img1_smooth, derivfilter, axis=1, mode='nearest')
# compute it
res = []
for pos0, pos1, score in corres:
p0, p1 = tuple(pos0)[::-1], tuple(pos1)[::-1] # numpy coordinates
dist = sum( abs(img0_smooth[p0]-img1_smooth[p1]) + abs(img0_dx[p0]-img1_dx[p1]) + abs(img0_dy[p0]-img1_dy[p1]) )
score = mul_coef * sqrt( max(0,small_eigen[p0])) / (sigma_score*sqrt(2*pi))*exp(-0.5*square(dist/sigma_score));
res.append((pos0,pos1,score))
return res
if __name__=='__main__':
args = sys.argv[1:]
img0 = array(Image.open(args[0]).convert('RGB'))
img1 = array(Image.open(args[1]).convert('RGB'))
out = open(args[2]) if len(args)>=3 else sys.stdout
ty0, tx0 = img0.shape[:2]
ty1, tx1 = img1.shape[:2]
rint = lambda s: int(0.5+float(s))
retained_matches = []
for line in sys.stdin:
line = line.split()
if not line or len(line)!=6 or not line[0][0].isdigit(): continue
x0, y0, x1, y1, score, index = line
retained_matches.append(((min(tx0-1,rint(x0)),min(ty0-1,rint(y0))),
(min(tx1-1,rint(x1)),min(ty1-1,rint(y1))),0))
assert retained_matches, 'error: no matches piped to this program'
for p0, p1, score in score_from_autocorr(img0, img1, retained_matches):
print >>out, '%d %d %d %d %f' %(p0[0],p0[1],p1[0],p1[1],score)