fhog-python/function_fDSST.py at master · lawpdas/fhog-python · 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
171
172
173
174
175
176
177
178
179
180
181
182
def get_scale_subwindow(img, pos, base_target_sz, scaleFactors, scale_model_sz):
    nScales = len(scaleFactors)

    for s in range(nScales):
        patch_sz = np.floor(base_target_sz * scaleFactors[s])

        xs = np.floor(pos[1]) + range(1,int(patch_sz[1])+1) - np.floor(patch_sz[1]//2)
        ys = np.floor(pos[0]) + range(1,int(patch_sz[0])+1) - np.floor(patch_sz[0]//2)

        xs[xs < 1] = 1
        ys[ys < 1] = 1
        xs[xs > img.shape[1]] = img.shape[1]
        ys[ys > img.shape[0]] = img.shape[0]

        im_patch = img[int(ys[0])-1:int(ys[-1]), int(xs[0])-1:int(xs[-1]), :]
        im_patch_resized = cv2.resize(im_patch,tuple(scale_model_sz[::-1].astype(int)))

        M = np.zeros(im_patch_resized.shape[:2], dtype='float32')
        O = np.zeros(im_patch_resized.shape[:2], dtype='float32')
        fhog.gradientMag(im_patch_resized.astype(np.float32),M,O)
        H = np.zeros([im_patch_resized.shape[0]//4, im_patch_resized.shape[1]//4, 32], dtype='float32')
        fhog.gradientHist(M,O,H)

        if s == 0:
            dim_scale = H.shape[0]*H.shape[1]*31
            out_pca = np.zeros([dim_scale, nScales], dtype='float32')

        out_pca[:,s] = H[:,:,:31].reshape(-1, order="F")

    return out_pca

def feature_projection_scale(xs_pca, projection_matrix, cos_window):
    tmp = np.matrix(projection_matrix)*np.matrix(xs_pca)
    tmp = np.array(tmp)
    return tmp*np.tile(cos_window, (tmp.shape[0], 1))


def resizeDFT(inputdft, desiredLen):

    len_ = len(inputdft)
    minsz = min(len_, desiredLen)

    scaling = desiredLen // len_


    resizeddft = np.zeros([1,int(desiredLen)], dtype='float32')+np.zeros([1,int(desiredLen)], dtype='float32')*1j

    mids = np.ceil(minsz/2.0)
    mide = np.floor((minsz-1)/2.0) - 1

    resizeddft[0,:int(mids)] = scaling * inputdft[:int(mids)]
    resizeddft[0,-1-int(mide):] = scaling * inputdft[-1-int(mide):]

    return  resizeddft


"""first frame"""
lambda_ = 1e-2
padding = 2.0
nScales = 17.0
scale_step = 1.02
nScalesInterp = 33.0
interp_factor = 0.025
scale_model_factor = 1.0
scale_sigma_factor = 1/16.0
scale_model_max_area = 512.0

pos_s = (bbox[:2] + bbox[2:]/2.0)[::-1]
target_sz = bbox[2:][::-1]
init_target_sz = bbox[2:][::-1]

currentScaleFactor = 1.0

base_target_sz = target_sz / currentScaleFactor

sz = np.floor( base_target_sz * (1 + padding ))


scale_sigma = nScalesInterp * scale_sigma_factor

scale_exp = np.array(range(-np.floor((nScales-1)/2).astype(int),np.ceil((nScales-1)/2).astype(int)+1)) * nScalesInterp/nScales
scale_exp_shift = np.roll(scale_exp, -np.floor((nScales-1)/2).astype(int))

interp_scale_exp = np.array(range(-np.floor((nScalesInterp-1)/2).astype(int), np.ceil((nScalesInterp-1)/2).astype(int)+1))
interp_scale_exp_shift = np.roll(interp_scale_exp, -np.floor((nScalesInterp-1)/2).astype(int))

scaleSizeFactors = np.power(scale_step,scale_exp)
interpScaleFactors = np.power(scale_step,interp_scale_exp_shift)

ys = np.exp(-0.5*np.power(scale_exp_shift,2) / scale_sigma**2)
ysf = fft(ys)
scale_window = np.hanning(len(ysf))


if scale_model_factor**2 * np.prod(init_target_sz) > scale_model_max_area:
    scale_model_factor = np.sqrt(scale_model_max_area/np.prod(init_target_sz))

scale_model_sz = np.floor(init_target_sz * scale_model_factor)


min_scale_factor = np.power( scale_step, np.ceil(np.log(np.max(5/sz)) / np.log(scale_step)) )
max_scale_factor = np.power( scale_step, np.floor(np.log(np.min(img.shape[:2] / base_target_sz)) / np.log(scale_step)) )

max_scale_dim = 1
s_num_compressed_dim = len(scaleSizeFactors)

##############
xs_pca = get_scale_subwindow(img, pos_s, base_target_sz, currentScaleFactor*scaleSizeFactors, scale_model_sz)

s_num = xs_pca.copy()
bigY = s_num.copy()
bigY_den = xs_pca.copy()

scale_basis,_ = np.linalg.qr(bigY)
scale_basis_den,_ = np.linalg.qr(bigY_den)

scale_basis = scale_basis.T

#create the filter update coefficients
tmp = feature_projection_scale(s_num,scale_basis,scale_window)
sf_proj = fft(tmp,tmp.shape[1],1)
sf_num = np.tile(ysf, (tmp.shape[0], 1))*conj(sf_proj)

xs = feature_projection_scale(xs_pca,scale_basis_den.T,scale_window)
xsf = fft(xs,xs.shape[1],1)
new_sf_den = np.sum(xsf * conj(xsf),0)

sf_den = new_sf_den.copy()


"""estimate scale"""
pos_s[0] = pos[1]
pos_s[1] = pos[0]
xs_pca = get_scale_subwindow(img, pos_s, base_target_sz, currentScaleFactor*scaleSizeFactors, scale_model_sz)

xs = feature_projection_scale(xs_pca,scale_basis,scale_window)
xsf = fft(xs,xs.shape[1],1)

scale_responsef = np.sum(sf_num * xsf, 0) / (sf_den + lambda_)


interp_scale_response = real(ifft( resizeDFT(scale_responsef, nScalesInterp)[0]))

recovered_scale_index = np.where(interp_scale_response == np.max(interp_scale_response))[0][0]

currentScaleFactor = currentScaleFactor * interpScaleFactors[recovered_scale_index]

if currentScaleFactor < min_scale_factor:
    currentScaleFactor = min_scale_factor
elif currentScaleFactor > max_scale_factor:
    currentScaleFactor = max_scale_factor
print(currentScaleFactor)
# update

xs_pca = get_scale_subwindow(img, pos_s, base_target_sz, currentScaleFactor*scaleSizeFactors, scale_model_sz)

s_num = (1 - interp_factor) * s_num + interp_factor * xs_pca.copy()
bigY = s_num.copy()
bigY_den = xs_pca.copy()

scale_basis,_ = np.linalg.qr(bigY)
scale_basis_den,_ = np.linalg.qr(bigY_den)

scale_basis = scale_basis.T

#create the filter update coefficients
tmp = feature_projection_scale(s_num,scale_basis,scale_window)
sf_proj = fft(tmp,tmp.shape[1],1)
sf_num = np.tile(ysf, (tmp.shape[0], 1))*conj(sf_proj)


xs = feature_projection_scale(xs_pca,scale_basis_den.T,scale_window)
xsf = fft(xs,xs.shape[1],1)

new_sf_den = np.sum(xsf * conj(xsf),0)

sf_den = (1 - interp_factor) * sf_den + interp_factor * new_sf_den.copy()

target_sz = np.floor(base_target_sz * currentScaleFactor)

bbox_est = np.array([pos[0]-target_sz[1]/2.0, pos[1]-target_sz[0]/2.0, target_sz[1], target_sz[0]])
bbox_est = np.round(bbox_est).astype(int)