-
Notifications
You must be signed in to change notification settings - Fork 1
/
process_images.m
240 lines (173 loc) · 7.17 KB
/
process_images.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
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
% PROCESS_IMAGES
%
% Demonstrates (windowed) feature extraction and pooling.
%
% It can be slow to extract features (depending on parameters and
% pooling type). Therefore, this script generates all the
% features and pools them; other scripts are used for analysis
% (e.g. of classificatin performance).
%
% Before running this script the first time, call set_path.m
% You also need to download the data set (see scripts in ./data).
% mjp, oct 2017
rng(1066);
%% PARAMETERS (change as desired for your experiment)
FIG_DIR = './Figures';
p.dataset = 'caltech-101-lean-iv3-layer1';
p.feature_type = 'raw';
if 0
% These were for 128x128 images we process locally
%p.window_size = [64,64]; p.stride = 32; % 74-65
%p.window_size = [32,32]; p.stride = 32; % 77-75 (non-monotonic)
p.window_size = [28,28]; p.stride = 28; % 79-76
%p.window_size = [20,20]; p.stride = 20; % 79-78 nearly equal perf. across alpha
else
p.window_size = [21,21]; p.stride = 21; % note: features from IV3 are 63x63
end
p.maxfun_supp = [2,6];
switch(lower(p.feature_type))
case 'raw'
% No local pre-processing.
% This is not expected to work well if the input file is raw images.
% However, we also use this when the input file contains some
% exogenously processed data (e.g. via a CNN)
featurize = @(x) x;
standardize = false;
case 'gabor'
f_xform = @(x) gabor_feature(double(x), 8, 8); % XXX: these may require tuning
featurize = @(x) apply_transform(x, f_xform);
standardize = false;
case 'gabor-edge'
% Note: Weilin recommends S=3
f_xform = @(x) gabor_edge_feature(double(x), 3, 16); % XXX: these may require tuning
featurize = @(x) apply_transform(x, f_xform);
standardize = false;
case 'dyadic-edge'
% Note: Weilin recommends S=3
f_xform = @(x) dyadic_edge_feature(double(x), 3, 16); % XXX: these may require tuning
featurize = @(x) apply_transform(x, f_xform);
standardize = false;
end
%% Load data set
switch (p.dataset)
case 'cifar-10'
batch_id = 1;
data = load_cifar_10('./data/cifar-10-batches-mat', batch_id);
% optional - reduce data set size
fprintf('[%s]: WARNING - reducing data set size temporarily (for speed)\n', mfilename);
data.X = data.X(:,:,:,1:1000);
data.y = data.y(1:1000);
case 'caltech-101-lean'
% it is a little slow to load this data set so cache it after loading for first time
cached_fn = 'caltech_101_lean.mat';
if exist(cached_fn)
load(cached_fn);
else
data = load_caltech101_lean('./data/101_ObjectCategories', 128);
save(cached_fn, 'data', '-v7.3');
end
% UPDATE: added some CNN features
case 'caltech-101-lean-iv3-layer1'
load('caltech_101_lean_iv3_layer1.mat');
data.X = permute(X_iv3, [2,3,4,1]); % back to matlab-preferred order
data.y = y(:)'; % force to be a row vector
clear X_f y;
otherwise
error('unknown dataset!');
end
n_images = size(data.X,4);
if ~exist(FIG_DIR)
mkdir(FIG_DIR);
end
%%
view_dataset(data.X, data.y, FIG_DIR);
if standardize
% optional: normalize data
% here we zero mean and unit variance along each pixel and channel, ie.
% mean(data.X(i,j,k,:)) = 0 (approximately)
% std(data.X(i,j,k,:)) = 1 (approximately)
%
fprintf('[%s]: WARNING - standardizing data\n', mfilename);
mu = mean(data.X, 4);
sigma = std(data.X, 0, 4);
data.X = bsxfun(@minus, data.X, mu);
data.X = bsxfun(@rdivide, data.X, sigma);
end
%% Preallocate space for features
% lazy way of determining how many feature dimensions we will have
% for each image.
x_dummy = featurize(data.X(:,:,:,1));
dummy = extract_all_windows(x_dummy, p.window_size, p.stride);
n_feats = size(dummy,3);
fprintf('[%s]: windowed data will have %d features\n', mfilename, n_feats);
feats.maxpool = zeros(n_feats, n_images);
feats.avgpool = zeros(n_feats, n_images);
feats.probpool = zeros(n_feats, n_images);
feats.maxfun = zeros(n_feats, n_images);
feats.maxfun_oo = zeros(n_feats, n_images, p.maxfun_supp(2) - p.maxfun_supp(1) + 1); % MAXFUN "one window"
feats.maxfun_centered = zeros(n_feats, n_images);
feats.y = zeros(size(data.y));
% shuffle images (to remove correlation in labels)
feats.idx = randperm(n_images);
%% extract features
tic
last_chatter = -Inf;
w_maxfun = NaN*ones(n_feats,n_images); % track maxfun support size (for debugging/analysis)
for ii = 1:n_images
orig_idx = feats.idx(ii); % index into original data set order
% feature extraction
x_i = data.X(:,:,:,orig_idx); % raw image
x_f = featurize(x_i); % filtered image
x_f = abs(x_f); % NOTE: we always take modulus for now...
x_fw = extract_all_windows(x_f, p.window_size, p.stride); % filtered and windowed
% pooling baselines
feats.maxpool(:,ii) = max_pooling(x_fw);
feats.avgpool(:,ii) = avg_pooling(x_fw);
feats.probpool(:,ii) = prob_pooling(x_fw);
feats.y(ii) = data.y(orig_idx);
% MAXFUN
[feats.maxfun(:,ii), w_maxfun(:,ii), loc] = maxfun_pooling(x_fw, p.maxfun_supp(1), p.maxfun_supp(2));
% "approximate MAXFUN" (max of fixed spatial average)
for jj = 1:size(feats.maxfun_oo,3)
pool_size = p.maxfun_supp(1) + (jj-1);
feats.maxfun_oo(:,ii,jj) = maxfun_pooling(x_fw, pool_size, pool_size);
end
[feats.maxfun_centered(:,ii), tmp] = centered_maxfun_pooling(x_fw);
%% Postprocessing
% max should *never* be less than avg
assert(all(feats.maxpool(:,ii) >= feats.avgpool(:,ii)));
% report status periodically to user
runtime = toc;
if runtime - last_chatter > 30
fprintf('[%s] processed %d images (of %d) in %0.2f seconds (%s features)\n', ...
mfilename, ii, n_images, runtime, p.feature_type);
last_chatter = runtime;
end
% optional: investigate where maxfun prefers non-trivial pooling region
if 0 && any(w_maxfun(:,ii) > p.maxfun_supp(1))
h = findobj('type', 'figure'); if length(h) > 50, continue; end % don't open too many figures
idx = find(w_maxfun(:,ii) > p.maxfun_supp(1));
% to avoid too many figures, just look at one
maxfun_pooling(x_fw(:,:,idx(1)), p.maxfun_supp(1), p.maxfun_supp(2), true);
title('maxfun non-trivial pooling region example');
end
end
fprintf('[%s]: total runtime: %0.2f sec\n', mfilename, toc);
save(sprintf('feats_%s.mat', p.feature_type), 'feats', 'p', '-v7.3');
pct_maxfun_nontrivial = sum(w_maxfun(:) > p.maxfun_supp(1)) / numel(w_maxfun);
%% look at distribution of maxfun pooling cardinalities
figure;
histogram(w_maxfun(:));
title(sprintf('maxfun support sizes (%0.2f%% non-trivial)', 100*pct_maxfun_nontrivial));
xlabel('support size');
ylabel('frequency');
saveas(gca, fullfile(FIG_DIR, 'maxfun_supp.png'));
%%
figure('Position', [100, 100, 900, 300]);
subplot(1,3,1);
imagesc(x_i(:,:,1)); colorbar; title('input');
subplot(1,3,2);
imagesc(x_f(:,:,1)); colorbar; title(p.feature_type);
subplot(1,3,3);
imagesc(x_fw(:,:,1)); colorbar; title('first window');
saveas(gca, fullfile(FIG_DIR, 'sample_windowing.png'));