-
Notifications
You must be signed in to change notification settings - Fork 0
/
setup_imdb_voc11inst.m
155 lines (129 loc) · 4.67 KB
/
setup_imdb_voc11inst.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
function imdb = setup_imdb_voc11inst(varargin)
% SETUP_IMDB_VOC11INST: Set up the imdb.
opts.clusterPath = '';
opts = vl_argparse(opts, varargin) ;
% load meta-data
% ------
imdb.classes.name = {'aeroplane', 'bicycle', 'bird', 'boat', 'bottle',...
'bus','car', 'cat', 'chair', 'cow',...
'diningtable', 'dog', 'horse', 'motorbike', 'person',...
'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor'};
imdb.imageDir = 'data/VOCdevkit/VOC-SBD/img/' ;
imdb.segPath = strrep(imdb.imageDir, 'img', 'cls-png', '%s.png');
imdb.instPath = strrep(imdb.segPath, 'cls-png', 'inst-png');
imdb.clusters.path = opts.clusterPath;
clusters = load(imdb.clusters.path, 'means', 'assign');
assign = clusters.assign;
clusters = clusters.means;
imdb.clusters.means = clusters;
C = numel(clusters);
nCluster = zeros(1, C);
for c = 1 : C
nCluster(c) = numel(clusters{c});
end
imdb.clusters.num = nCluster;
% images
% ------
k = 0 ;
for thisSet = {'train', 'val'}
thisSet = char(thisSet) ;
fprintf('Loading PASCAL VOC %s set\n', thisSet) ;
gtids = textread(sprintf('data/VOCdevkit/VOC-SBD/%s.txt',thisSet),'%s');
k = k + 1 ;
imdb_.images.name{k} = strcat(gtids,'.jpg');
N = numel(imdb_.images.name{k});
sizes = zeros(N, 2);
gtbox = cell(N, 1);
gtlabel = cell(N, 1);
gtdist = cell(N, 1);
gtdistflip = cell(N, 1);
% Load ground truth objects
start = tic;
% for i=1:length(gtids)
parfor i=1:length(gtids)
% Read annotation
[boxes_inst, labels_inst, size_inst, ~, ~, iou_cluster, iou_cluster_flip] = ...
read_record(imdb, gtids{i}, clusters); % boxes: Mx4
sizes(i, :) = size_inst;
gtbox{i} = boxes_inst;
gtlabel{i} = labels_inst;
gtdist{i} = iou_cluster;
gtdistflip{i} = iou_cluster_flip;
assert(numel(labels_inst) == size(iou_cluster, 1));
assert(isequal(size(iou_cluster), size(iou_cluster_flip)));
if mod(i-1, 100) == 0, fprintf('[%s %.1f sec] %d/%d.\n', thisSet, toc(start), i, length(gtids)); end
end
imdb_.images.set{k} = k * ones(1, N);
imdb_.images.size{k} = sizes;
imdb_.boxes.gtbox{k} = gtbox;
imdb_.boxes.gtlabel{k} = gtlabel;
imdb_.boxes.gtdist{k} = gtdist;
imdb_.boxes.gtdistflip{k} = gtdistflip;
end
imdb.images.name = vertcat(imdb_.images.name{:}) ;
imdb.images.size = vertcat(imdb_.images.size{:}) ;
imdb.images.set = horzcat(imdb_.images.set{:}) ;
imdb.boxes.gtbox = vertcat(imdb_.boxes.gtbox{:}) ;
imdb.boxes.gtlabel = vertcat(imdb_.boxes.gtlabel{:}) ;
imdb.boxes.gtdist = vertcat(imdb_.boxes.gtdist{:});
imdb.boxes.gtdistflip = vertcat(imdb_.boxes.gtdistflip{:});
% flip
% ------
imdb.boxes.flip = zeros(size(imdb.images.name));
% Add flipped
train = (imdb.images.set == 1) ;
imdb.images.name = vertcat(imdb.images.name, imdb.images.name(train)) ;
imdb.images.set = horzcat(imdb.images.set, imdb.images.set(train)) ;
imdb.images.size = vertcat(imdb.images.size, imdb.images.size(train,:)) ;
imdb.boxes.flip = vertcat(imdb.boxes.flip, ones(sum(train),1)) ; % [non-flip, flip]
imdb.boxes.gtbox = vertcat(imdb.boxes.gtbox , imdb.boxes.gtbox(train)) ;
imdb.boxes.gtlabel = vertcat(imdb.boxes.gtlabel, imdb.boxes.gtlabel(train)) ;
imdb.boxes.gtdist = vertcat(imdb.boxes.gtdist, imdb.boxes.gtdistflip(train));
imdb.boxes = rmfield(imdb.boxes, 'gtdistflip');
% when flipped, the sublabel has changed
for i=1:numel(imdb.boxes.gtbox)
if imdb.boxes.flip(i)
width = imdb.images.size(i, 1);
gtbox = imdb.boxes.gtbox{i} ;
assert(all(gtbox(:,1)<=width));
assert(all(gtbox(:,3)<=width));
gtbox(:,1) = width - gtbox(:,3) + 1;
gtbox(:,3) = width - imdb.boxes.gtbox{i}(:,1) + 1;
imdb.boxes.gtbox{i} = gtbox;
end
end
% gtsublabel: [N, 1]
imdb.boxes.gtsublabel = cell(size(imdb.boxes.gtlabel));
% train: find sub-label according to cluster
train = find(imdb.images.set == 1);
offset = ones(1, C);
for i = 1 : numel(train)
idx = train(i);
gtlabel = imdb.boxes.gtlabel{idx};
M = numel(gtlabel);
gtsublabel = zeros(M, 1);
for j = 1 : M
cls = gtlabel(j);
gtsublabel(j) = assign{cls}(offset(cls))';
offset(cls) = offset(cls) + 1;
end
imdb.boxes.gtsublabel{idx} = gtsublabel;
end
for c = 1 : C
assert(offset(c) == numel(assign{c}) + 1);
end
% val: find sublabels according to max iou
val = find(imdb.images.set == 2);
gtsublabels = cell(1, numel(val));
% for i = 1 : numel(val)
parfor i = 1 : numel(val)
idx = val(i);
[~, ~, ~, sublabel, ~, ~, ~] = read_record(imdb, imdb.images.name{idx}(1:end-4), clusters);
gtsublabels{i} = sublabel;
if mod(i-1, 100) == 0, fprintf('[val %.1f sec] %d/%d.\n', toc(start), i, numel(val)); end
end
% parfor necessity
for i = 1 : numel(val)
idx = val(i);
imdb.boxes.gtsublabel{idx} = gtsublabels{i};
end