-
Notifications
You must be signed in to change notification settings - Fork 9
/
fastmatch.pyx
180 lines (153 loc) · 7.92 KB
/
fastmatch.pyx
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
# encoding: utf-8
# cython: profile=True
# filename: fastmatch.pyx
"""
Fast matching algorithm for image matching
Jonas Toft Arnfred, 2013-05-05
"""
####################################
# #
# Imports #
# #
####################################
from matchutil import get_features
from cache import Grid_Cache, Metric_Cache
import cv2
from imaging import get_thumbnail, get_size
import numpy
import itertools
cimport numpy
from cache cimport Metric_Cache, Grid_Cache
####################################
# #
# Functions #
# #
####################################
def match(Metric_Cache query_cache, numpy.ndarray[numpy.uint8_t, ndim=3] target_img, object options = {}) :
# Declare datatypes
cdef int thumb_x, thumb_y, grid_x, grid_y, grid_margin, radius
cdef double thumb_tau
cdef numpy.ndarray thumb_positions, thumb_ratios, positions_iter
# Get parameters
thumb_x, thumb_y = options.get("thumb_size", (400, 400))
grid_x, grid_y = options.get("grid_size", (50, 50))
thumb_strategy = options.get("thumb_strategy", lambda n : n)
log = options.get("log", None)
grid_margin = options.get("grid_margin", 25)
radius = options.get("radius", 100)
# Create target cache
cdef Grid_Cache target_cache = Grid_Cache(target_img, (grid_x, grid_y), get_features, margin = grid_margin)
thumb_positions, thumb_ratios = match_thumbs(target_img, query_cache, thumb_x = thumb_x, thumb_y = thumb_y)
# Create a function where we can wary tau to get different results
def get_matches(double tau) :
thumb_tau = thumb_strategy(tau)
positions_iter = itertools.chain(thumb_positions[thumb_ratios<thumb_tau])
matches = do_iter(positions_iter, query_cache, target_cache, tau = tau, thumb_tau = thumb_tau, radius = radius, log = log)
return matches
return get_matches
cdef object do_iter(object positions, Metric_Cache cache, Grid_Cache target_grid, double tau, double thumb_tau, int radius, object log = None) :
# Declare datatypes
cdef numpy.ndarray[numpy.double_t] query_pos, target_pos
cdef numpy.ndarray[numpy.double_t, ndim=2] p
cdef numpy.ndarray result_pos, ratios
cdef int col, row, query_col, query_row
cdef double r
matches = []
has_matched = {}
found_matches = {}
while True :
try :
query_pos, target_pos = positions.next()
col, row = target_grid.block(target_pos[0], target_pos[1])
query_col, query_row = target_grid.block(query_pos[0], query_pos[1])
if not has_matched.get((col, row, query_col, query_row), False) :
has_matched[(col, row, query_col, query_row)] = True
result_pos, ratios, query_idx = match_position((query_pos, target_pos), cache, target_grid, radius = radius)
# For each match we don't discard, we might want to examine the neighbor field
neighbors = get_neighbors(target_pos, result_pos[ratios<tau], target_grid)
if len(neighbors) > 0 :
positions = itertools.chain(neighbors, positions)
# Log if we have to
if log != None :
log.append(log_round(query_pos, target_pos, result_pos, target_grid, ratios, tau, radius))
# Yield result if it hasn't been yielded already
for p, r, index in zip(result_pos[ratios<tau], ratios[ratios<tau], query_idx[ratios<tau]) :
p_touple = map(int, (p[0,0], p[0,1], p[1,0], p[1,1]))
if p_touple not in found_matches.get(r,[]) :
found_matches[r] = found_matches.get(r,[]) + [p_touple]
matches.append((index, { "positions" : p, "ratio" : r }))
except StopIteration :
break
return matches
cdef object get_neighbors(numpy.ndarray[numpy.double_t] target_pos,
numpy.ndarray result_pos,
Grid_Cache target_grid) :
cdef int col, row
cdef numpy.ndarray[numpy.int_t] neighbor_pos
col, row = target_grid.block(target_pos[0], target_pos[1])
neighbors = []
for p_query, p_target in result_pos :
neighbor_pos = target_grid.get_neighbor(col, row, p_target[0], p_target[1])
if neighbor_pos[0] != -1 :
neighbors.append(numpy.array((p_query, neighbor_pos)))
return neighbors
cdef object match_thumbs(numpy.ndarray[numpy.uint8_t, ndim=3] img, Metric_Cache query_cache, int thumb_x = 400, int thumb_y = 400) :
cdef int t_orig_x, t_orig_y
cdef double t_ratio_x, t_ratio_y, q_ratio_x, q_ratio_y
cdef numpy.ndarray t_pos, q_pos, q_distances
cdef numpy.ndarray ratios, q_ratio, indices, pos_scaled
# Load target and find descriptors and size
cdef numpy.ndarray[numpy.uint8_t, ndim=3] target = get_thumbnail(img, (thumb_x, thumb_y))
t_orig_x, t_orig_y = get_size(img)
t_keypoints, t_descriptors = get_features(target)
# Similar for query
q_descriptors = query_cache.thumb["descriptors"]
q_distances = query_cache.thumb["distances"]
q_pos = query_cache.thumb["positions"]
# match thumbnails and find ratio
matcher = cv2.BFMatcher(cv2.NORM_L2, crossCheck=True)
matches = [m for m in matcher.knnMatch(q_descriptors, t_descriptors, k=1) if len(m) > 0]
ratios = numpy.array([m[0].distance / q_distances[m[0].queryIdx] for m in matches])
# Get positions of points and their scaling factor
t_pos = numpy.array([t_keypoints[m[0].trainIdx].pt for m in matches])
t_ratio_x = t_orig_x/float(target.shape[1])
t_ratio_y = t_orig_y/float(target.shape[0])
t_ratio = numpy.array([t_ratio_x, t_ratio_y])
q_pos = numpy.array([query_cache.thumb["positions"][m[0].queryIdx] for m in matches])
q_ratio_x = query_cache.original["size"][0]/float(query_cache.thumb["size"][0])
q_ratio_y = query_cache.original["size"][1]/float(query_cache.thumb["size"][1])
q_ratio = numpy.array([q_ratio_x, q_ratio_y])
# Sort ratios and scale positions
indices = numpy.argsort(ratios)
pos_scaled = numpy.array([(q_p * q_ratio, t_p * t_ratio) for q_p, t_p in zip(q_pos, t_pos)])
return pos_scaled[indices], ratios[indices]
# Match point strategy #1:
cdef match_position(pos, Metric_Cache query_cache, Grid_Cache target, int radius = 100) :
cdef numpy.ndarray query_cache_pos, query_cache_dis, ratios, positions, ratio_indices
cdef int r, offset_x, offset_y, target_x, target_y, query_x, query_y
# Find positions
query_x, query_y = pos[0]
target_x, target_y = pos[1]
query_ds, query_pos, query_dis, query_idx = query_cache.get(query_x, query_y, radius)
target_kp, target_ds = target.get(target_x, target_y)
if target_ds == None :
return numpy.array([]), numpy.array([]), numpy.array([])
offset_x, offset_y = target.offset(target_x, target_y)
target_pos = [numpy.array([k.pt[0]+offset_x, k.pt[1]+offset_y]) for k in target_kp]
# Match descriptors using bf
matcher = cv2.BFMatcher(cv2.NORM_L2, crossCheck=True)
matches = [m for m in matcher.knnMatch(query_ds, target_ds, k=1) if len(m) > 0]
# Distances to nearest neighbor and positions
ratios = numpy.array([m[0].distance / query_dis[m[0].queryIdx] for m in matches])
positions = numpy.array([(query_pos[m[0].queryIdx], target_pos[m[0].trainIdx]) for m in matches])
indices = numpy.array([query_idx[m[0].queryIdx] for m in matches])
return positions, ratios, indices
cdef log_round(numpy.ndarray query_pos, numpy.ndarray target_pos, numpy.ndarray result_pos, Grid_Cache target_grid, numpy.ndarray ratios, double tau, int radius) :
return {
"query_pos" : query_pos,
"target_pos" : target_pos,
"target_grid" : target_grid.last,
"matches" : result_pos[ratios<tau],
"radius" : radius,
"ratios" : ratios[ratios<tau],
"margin" : target_grid.margin }