aps/asym_patch_sampling.py at main · visresearch/aps · GitHub

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
import numpy as np


def get_overlap_region_box_x1y1x2y2(box1, box2):
    x1_1, y1_1, x2_1, y2_1 = box1
    x1_2, y1_2, x2_2, y2_2 = box2

    x1 = max(x1_1, x1_2)
    y1 = max(y1_1, y1_2)

    x2 = min(x2_1, x2_2)
    y2 = min(y2_1, y2_2)

    is_overlap = (x1 < x2) and (y1 < y2)

    if is_overlap:
        return x1, y1, x2, y2
    else:
        return None


def get_overlap_ratio_in_box1_x1y1x2y2(box1, box2):

    box_overlap = get_overlap_region_box_x1y1x2y2(box1, box2)

    ratio = 0.0

    if box_overlap is not None:
        x1_1, y1_1, x2_1, y2_1 = box1
        x1, y1, x2, y2 = box_overlap

        area_b1 = (x2_1 - x1_1) * (y2_1 - y1_1)
        area_overlap = (x2 - x1) * (y2 - y1)

        ratio = area_overlap / area_b1

    return ratio


def get_boxs_non_overlap_ratio_x1y1x2y2(boxs, height, width, ngrid):
    ygrid_size = height / ngrid
    xgrid_size = width / ngrid

    num_grid = ngrid * ngrid
    ratios = np.ones((num_grid, ))

    for box in boxs:
        x1, y1, x2, y2 = box
        ids = []

        r1 = int(y1 / ygrid_size)
        r2 = int(y2 / ygrid_size)

        r1 = max(0, r1)
        r2 = min(r2, ngrid - 1)

        c1 = int(x1 / xgrid_size)
        c2 = int(x2 / xgrid_size)

        c1 = max(0, c1)
        c2 = min(c2, ngrid - 1)

        for r in range(r1, r2 + 1):
            tmp = r * ngrid
            y1 = r * ygrid_size
            y2 = y1 + ygrid_size
            for c in range(c1, c2 + 1):
                _id = tmp + c

                x1 = c * xgrid_size

                box_grid = (x1, y1, x1 + xgrid_size , y2)

                ratio = get_overlap_ratio_in_box1_x1y1x2y2(box_grid, box)

                ratios[_id] -= ratio
                if ratios[_id] < 1e-6:
                    ratios[_id] = 0.0

    return ratios


def get_boxs_non_overlap_x1y1x2y2(boxs, height, width, ngrid):
    ygrid_size = height / ngrid
    xgrid_size = width / ngrid

    num_grid = ngrid * ngrid
    ratios = np.ones((num_grid, ))

    for box in boxs:
        x1, y1, x2, y2 = box
        ids = []

        r1 = int(y1 / ygrid_size)
        r2 = int(y2 / ygrid_size)

        r1 = max(0, r1)
        r2 = min(r2, ngrid - 1)

        c1 = int(x1 / xgrid_size)
        c2 = int(x2 / xgrid_size)

        c1 = max(0, c1)
        c2 = min(c2, ngrid - 1)

        for r in range(r1, r2 + 1):
            tmp = r * ngrid
            for c in range(c1, c2 + 1):
                _id = tmp + c
                ratios[_id] = 0.0

    return ratios


def get_grid_boxs(height, width, ngrid, x_offset=0., y_offset=0.):

    ygrid_size = height / ngrid
    xgrid_size = width / ngrid

    boxs = []

    for i in range(ngrid):

        y1 = i * ygrid_size + y_offset
        y2 = y1 + ygrid_size

        for j in range(ngrid):
            x1 = j * xgrid_size + x_offset
            x2 = x1 + xgrid_size
            boxs.append((x1, y1, x2, y2))

    return boxs


def sample_overlap_less_patches_with_box1(box1, box2, ngrid, idxs_patch_box1 , power):
    # box1, box2: (i, j, h, w)
    y1, x1, h1, w1 = box1
    y2, x2, h2, w2 = box2

    n = len(idxs_patch_box1)

    boxs_patch = get_grid_boxs(h1, w1, ngrid, x1 - x2, y1 - y2)

    boxs_patch_select = [boxs_patch[i] for i in idxs_patch_box1]

    ratios = get_boxs_non_overlap_ratio_x1y1x2y2(boxs_patch_select, h2, w2, ngrid)
    ratios = ratios**power
    probs = ratios / np.sum(ratios)

    idxs_patch_box2 = np.random.choice(ngrid*ngrid, n, replace=False, p=probs)


    return idxs_patch_box2


def get_random_patch_sequence_index_pair(boxs1, boxs2, ngrid, sampling_ratio, time=4, power=3):
    npatch = ngrid * ngrid
    npatch_select = int(npatch * sampling_ratio)


    idx_patches = np.arange(npatch)
    idxs1=[]
    idxs2=[]
    np.random.shuffle(idx_patches)
    for i in range(time):
        idx1 = idx_patches[npatch_select*i:npatch_select*(i+1)]
        idx2 = sample_overlap_less_patches_with_box1(boxs1, boxs2, ngrid, idx1, power)
        idxs1.append(idx1)
        idxs2.append(idx2)
    return idxs1, idxs2