File size: 7,836 Bytes
d4035c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
function [sift, SIFTparam] = LMdenseSift(D, HOMEIMAGES, SIFTparam, HOMESIFT)
%
% Computes dense SIFT features.
% The SIFT grid will be defined by the parameters:
%    SIFTparam.grid_spacing = 1; % distance between grid centers
%    SIFTparam.patch_size = 16;  % size of patch from which to compute SIFT
%    descriptor (it has to be a factor of 4)
%
% Run demoSIFT.m to see an example of how it works.
%
% The SIFT descriptor at each location has 128 dimensions.
%
% This function can be called as:
%
% [sift, param] = LMdenseSift(D(n), HOMEIMAGES, param);
% [sift, param] = LMdenseSift(filename, HOMEIMAGES, param);
% [sift, param] = LMdenseSift(filename, HOMEIMAGES, param, HOMESIFT);
% LMdenseSift(D, HOMEIMAGES, param, HOMESIFT);
%
% 'sift' corresponds to the features of the last image. So, call it passing
% just one image. But you can precompute the SIFT features for a set of
% images: When calling LMdenseSift with a fourth argument it will store the sift descriptors in a
% new folder structure mirroring the folder structure of the images. Then,
% when called again, if the sift files already exist, it will just read
% them without recomputing them.
%
% Antonio Torralba, 2008


if nargin==4
    precomputed = 1;
    % get list of folders and create non-existing ones
    %listoffolders = {D(:).annotation.folder};
else
    precomputed = 0;
    HOMESIFT = '';
end

if nargin<3
    % Default parameters
    SIFTparam.grid_spacing = 1; % distance between grid centers
    SIFTparam.patch_size = 16; % size of patch from which to compute SIFT descriptor (it has to be a factor of 4)
end
SIFTparam.w = SIFTparam.patch_size/2; % boundary
Nfeatures = 128;

if isstruct(D)
    % [gist, param] = LMdenseSift(D, HOMEIMAGES, param);
    Nscenes = length(D);
    typeD = 1;
end

if iscell(D)
    % [gist, param] = LMdenseSift(filename, HOMEIMAGES, param);
    Nscenes = length(D);
    typeD = 2;
end

if isnumeric(D)
    % [gist, param] = LMdenseSift(img, HOMEIMAGES, param);
    Nscenes = size(D,4);
    typeD = 3;
end

if Nscenes >1
    fig = figure;
end

% Loop: Compute SIFT features for all scenes
sift = zeros([Nscenes Nfeatures], 'single');
for n = 1:Nscenes
    g = [];
    todo = 1;
    
    % if SIFT has already been computed, just read the file
    if precomputed==1
        filesift = fullfile(HOMESIFT, D(n).annotation.folder, [D(n).annotation.filename(1:end-4) '.mat']);
        if exist(filesift, 'file')
            load(filesift, 'sift', 'SIFTparam');
            todo = 0;
        end
    end
    
    % otherwise compute SIFT
    if todo==1
        disp([n Nscenes])

        % load image
        try
            switch typeD
                case 1
                    img = LMimread(D, n, HOMEIMAGES);
                case 2
                    img = imread(fullfile(HOMEIMAGES, D{n}));
                case 3
                    img = D(:,:,:,n);
            end
        catch
            disp(D(n).annotation.folder)
            disp(D(n).annotation.filename)
            rethrow(lasterror)
        end        
        
        % get SIFT descriptors
        [sift, SIFTparam.grid_x, SIFTparam.grid_y] = dense_sift(img, SIFTparam);
        
        if isfield(SIFTparam, 'edges')
            % 'dont-compute': default if field not present
            % 'siftrepeat'
            w = SIFTparam.w-1;
            switch lower(SIFTparam.edges)
                case 'siftrepeat'
                    sift = [repmat(sift(1,:,:),[w 1 1]); sift; repmat(sift(end,:,:),[w 1 1])];
                    sift = [repmat(sift(:,1,:),[1 w 1]), sift, repmat(sift(:,end,:),[1 w 1])];
                otherwise
                    error('Unknown edges method')
            end
        end
        
        % save SIFT if a HOMESIFT file is provided
        if precomputed
            mkdir(fullfile(HOMESIFT, D(n).annotation.folder))
            save (filesift, 'sift', 'SIFTparam')
        end

        if Nscenes >1
            figure(fig);
            subplot(121)
            imshow(uint8(img))
            subplot(122)
            showColorSIFT(sift)
        end
    end

    drawnow
end




function [sift_arr, grid_x, grid_y] = dense_sift(I, SIFTparam)
% Original script by Svetlana Lazebnick
% Antonio Torralba: modified using convolutions to speed up the
% computations.

grid_spacing = SIFTparam.grid_spacing;
patch_size = SIFTparam.patch_size;

I = double(I);
I = mean(I,3);
I = I /max(I(:));

% parameters
num_angles = 8;
num_bins = 4;
num_samples = num_bins * num_bins;
alpha = 9; %% parameter for attenuation of angles (must be odd)

if nargin < 5
    sigma_edge = 1;
end

angle_step = 2 * pi / num_angles;
angles = 0:angle_step:2*pi;
angles(num_angles+1) = []; % bin centers

[hgt wid] = size(I);

[G_X,G_Y]=gen_dgauss(sigma_edge);

% add boundary:
I = [I(2:-1:1,:,:); I; I(end:-1:end-1,:,:)];
I = [I(:,2:-1:1,:) I I(:,end:-1:end-1,:)];

I = I-mean(I(:));
I_X = filter2(G_X, I, 'same'); % vertical edges
I_Y = filter2(G_Y, I, 'same'); % horizontal edges

I_X = I_X(3:end-2,3:end-2,:);
I_Y = I_Y(3:end-2,3:end-2,:);

I_mag = sqrt(I_X.^2 + I_Y.^2); % gradient magnitude
I_theta = atan2(I_Y,I_X);
I_theta(find(isnan(I_theta))) = 0; % necessary????

% grid 
grid_x = patch_size/2:grid_spacing:wid-patch_size/2+1;
grid_y = patch_size/2:grid_spacing:hgt-patch_size/2+1;

% make orientation images
I_orientation = zeros([hgt, wid, num_angles], 'single');

% for each histogram angle
cosI = cos(I_theta);
sinI = sin(I_theta);
for a=1:num_angles
    % compute each orientation channel
    tmp = (cosI*cos(angles(a))+sinI*sin(angles(a))).^alpha;
    tmp = tmp .* (tmp > 0);

    % weight by magnitude
    I_orientation(:,:,a) = tmp .* I_mag;
end

% Convolution formulation:
weight_kernel = zeros(patch_size,patch_size);
r = patch_size/2;
cx = r - 0.5;
sample_res = patch_size/num_bins;
weight_x = abs((1:patch_size) - cx)/sample_res;
weight_x = (1 - weight_x) .* (weight_x <= 1);

for a = 1:num_angles
    %I_orientation(:,:,a) = conv2(I_orientation(:,:,a), weight_kernel, 'same');
    I_orientation(:,:,a) = conv2(weight_x, weight_x', I_orientation(:,:,a), 'same');
end

% Sample SIFT bins at valid locations (without boundary artifacts)
% find coordinates of sample points (bin centers)
[sample_x, sample_y] = meshgrid(linspace(1,patch_size+1,num_bins+1));
sample_x = sample_x(1:num_bins,1:num_bins); sample_x = sample_x(:)-patch_size/2;
sample_y = sample_y(1:num_bins,1:num_bins); sample_y = sample_y(:)-patch_size/2;

sift_arr = zeros([length(grid_y) length(grid_x) num_angles*num_bins*num_bins], 'single');
b = 0;
for n = 1:num_bins*num_bins
    sift_arr(:,:,b+1:b+num_angles) = I_orientation(grid_y+sample_y(n), grid_x+sample_x(n), :);
    b = b+num_angles;
end
clear I_orientation


% Outputs:
[grid_x,grid_y] = meshgrid(grid_x, grid_y);
[nrows, ncols, cols] = size(sift_arr);

% normalize SIFT descriptors

%sift_arr = reshape(sift_arr, [nrows*ncols num_angles*num_bins*num_bins]);
%sift_arr = normalize_sift(sift_arr);
%sift_arr = reshape(sift_arr, [nrows ncols num_angles*num_bins*num_bins]);


ct = .1;
sift_arr = sift_arr + ct;
tmp = sqrt(sum(sift_arr.^2, 3));
sift_arr = sift_arr ./ repmat(tmp, [1 1 size(sift_arr,3)]);

function [GX,GY]=gen_dgauss(sigma)

% laplacian of size sigma
%f_wid = 4 * floor(sigma);
%G = normpdf(-f_wid:f_wid,0,sigma);
%G = G' * G;
G = gen_gauss(sigma);
[GX,GY] = gradient(G); 

GX = GX * 2 ./ sum(sum(abs(GX)));
GY = GY * 2 ./ sum(sum(abs(GY)));


function G=gen_gauss(sigma)

if all(size(sigma)==[1, 1])
    % isotropic gaussian
	f_wid = 4 * ceil(sigma) + 1;
    G = fspecial('gaussian', f_wid, sigma);
%	G = normpdf(-f_wid:f_wid,0,sigma);
%	G = G' * G;
else
    % anisotropic gaussian
    f_wid_x = 2 * ceil(sigma(1)) + 1;
    f_wid_y = 2 * ceil(sigma(2)) + 1;
    G_x = normpdf(-f_wid_x:f_wid_x,0,sigma(1));
    G_y = normpdf(-f_wid_y:f_wid_y,0,sigma(2));
    G = G_y' * G_x;
end