-
Notifications
You must be signed in to change notification settings - Fork 6
/
sdsift.m
140 lines (121 loc) · 3.79 KB
/
sdsift.m
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
function [descs mask] = siftseg4(im,seg,settings)
% [descs mask] = siftseg4(im,seg,settings)
%
% Compute segmentation-aware SIFT.
% INPUTS
% im: RGB/grayscale image
% seg: Soft segmentations (image height x image width x segm. layers)
% Leave empty ([]) to get plain SIFT features
% settings: See this file for details
%
% OUTPUTS
% descs: (Segmentation-aware) SIFT descriptors
% masks: 4x4 masks for every image pixel (optional)
%
% Copyright (C) 2012 Eduard Trulls
%
% This program is free software: you can redistribute it and/or modify
% it under the terms of the GNU General Public License as published by
% the Free Software Foundation, either version 3 of the License, or
% (at your option) any later version.
%
% This program is distributed in the hope that it will be useful,
% but WITHOUT ANY WARRANTY; without even the implied warranty of
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
% GNU General Public License for more details.
%
% You should have received a copy of the GNU General Public License
% along with this program. If not, see <http://www.gnu.org/licenses/>.
%
%% init
if nargin~=3
error('Bad parameters');
end
if isempty(seg)
use_seg = false;
if nargout==2
error('Only one output for this mode');
end
else
use_seg = true;
end
%% check
assert(settings.lambda>=0);
[h w c] = size(im);
if c>1
im = rgb2gray(im);
end
if use_seg
assert(h == size(seg,1) && w == size(seg,2));
num_seg = size(seg,3);
end
if round(settings.scale) ~= settings.scale || settings.scale < 1
error('The bin size must be an integer (to do: subsample the segmentation further)');
end
%% pad image
pad_size = 2*settings.scale;%+2;
im_padded = padarray(im,[pad_size pad_size],'symmetric');
if use_seg
seg_padded = padarray(seg,[pad_size pad_size],'symmetric');
end
frames = zeros(4,h*w);
count = 0;
[x y] = meshgrid(1:w,1:h);
frames = [x(:)'+pad_size; y(:)'+pad_size; repmat(settings.scale,1,numel(y)); repmat(0,1,numel(y)) ];
%% get descriptors
magnif = 1;
tic;
[dummy descs] = vl_sift(single(im_padded),'frames',frames,'magnif',1);
fprintf('Computed descriptors in %.02f\n',toc);
% convert from uint8
descs = double(descs);
% normalize
descs = descs ./ repmat(sqrt(sum(descs.^2,1)),[128,1]);
descs(find(isnan(descs))) = 0;
%% compute and apply masks
if use_seg
t_masks = tic;
SBP = magnif * settings.scale; % bin size
NBP = 4; % number of hor/vert bins
% precompute block values
% subsample to get bin centered at half pixels
seg_padded_ss = imresize(seg_padded,2);
f = ones(2*SBP) / (2*SBP)^2;
blocks = zeros(size(seg_padded_ss,1),size(seg_padded_ss,2),num_seg);
for k=1:num_seg
blocks(:,:,k) = conv2(seg_padded_ss(:,:,k),f,'same');
end
% get masks
[x y] = meshgrid([-3 -1 1 3]*SBP);
d = zeros(h*w,num_seg,16);
r = size(blocks);
for s=1:num_seg
for k=1:16
% we use the average over SBPxSBP around the pixel as a reference value
d(:,s,k) = blocks(2*frames(2,:)+y(k) + (2*frames(1,:)+x(k)-1)*r(1) + (s-1)*r(1)*r(2)) - blocks(2*frames(2,:) + (2*frames(1,:)-1)*r(1) + (s-1)*r(1)*r(2));
end
end
d = squeeze(sqrt( sum(d.^2,2) ));
mask = exp(-settings.lambda * d);
% re-normalize
if settings.norm_mask
mask = mask ./ repmat(sum(mask,2),[1 16]) * 16 ;
end
fprintf('Computed segmentation masks in %.02f\n',toc(t_masks));
% apply masks
% transpose because vlsift follows row-major order
mask = reshape(mask,[h*w 4 4]);
m = permute(mask,[1 3 2]);
m = reshape(m,h*w,16)';
m = repmat(m(:)',[8 1]);
descs = descs(:) .* m(:);
mask = reshape(mask,[h w 4 4]);
if ~settings.norm_mask
descs = reshape(descs,128,h*w);
% renormalize or the masks could get very small values
descs = descs ./ repmat(sqrt(sum(descs.^2,1)),[128,1]);
descs(find(isnan(descs))) = 0;
end
end
descs = permute(reshape(descs,[128,h,w]),[2 3 1]);