forked from scott89/FCNT
-
Notifications
You must be signed in to change notification settings - Fork 0
/
cnn2_pf_tracker.m
269 lines (216 loc) · 9.22 KB
/
cnn2_pf_tracker.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
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
function position = cnn2_pf_tracker(tracker_param)
randn('state',0);
load_tracker_param;
caffe('presolve_gnet');
caffe('presolve_snet');
%% read images
im1_name = sprintf([data_path 'img/%04d.jpg'], im1_id);
im1 = double(imread(im1_name));
if size(im1,3)~=3
im1(:,:,2) = im1(:,:,1);
im1(:,:,3) = im1(:,:,1);
end
%% extract roi
roi1 = ext_roi(im1, location, [0, 0], roi_size, s1);
%% preprocess roi
roi1 = impreprocess(roi1);
fea1 = caffe('forward', {single(roi1)});
%
fea_sz = size(fea1{1});
lfea1 = fea1{1};
gfea1 = imresize(fea1{2}, fea_sz(1:2));
%% ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
max_iter_select = 100;
max_iter = 50;
map1 = GetMap(size(im1), fea_sz, roi_size, location, [0, 0], s1, 'gaussian');
conf_store = 0;
%% select
caffe('set_phase_train');
for i=1:max_iter_select
s_pre_map = caffe('forward_snet', {lfea1});
s_pre_map = s_pre_map{1};
figure(1011); subplot(1,2,1); imagesc(permute(s_pre_map,[2,1,3]));
s_diff = s_pre_map-permute(single(map1), [2,1,3]);
caffe('backward_snet', {single(s_diff)});
caffe('update_snet');
g_pre_map = caffe('forward_gnet', {gfea1});
g_pre_map = g_pre_map{1};
figure(1011); subplot(1,2,2); imagesc(permute(g_pre_map,[2,1,3]));
g_diff = g_pre_map-permute(single(map1), [2,1,3]);
caffe('backward_gnet', {single(g_diff)});
caffe('update_gnet');
fprintf('Iteration %03d/%03d, Local Loss %f, Global Loss %f\n', i, max_iter, sum(abs(s_diff(:))), sum(abs(g_diff(:))));
end
[~, lid] = compute_saliency({lfea1}, map1, 'ssolver');
[~, gid] = compute_saliency({gfea1}, map1, 'gsolver');
lid = lid(1:ch_num);
gid = gid(1:ch_num);
lfea1 = lfea1(:,:,lid);
gfea1 = gfea1(:,:,gid);
fea2_store = lfea1;
map2_store = map1;
%% train
caffe('init_ssolver', snet_solver_def_file);
caffe('init_gsolver', gnet_solver_def_file);
caffe('set_phase_train');
caffe('presolve_gnet');
caffe('presolve_snet');
for i=1:max_iter
s_pre_map = caffe('forward_snet', {lfea1});
s_pre_map = s_pre_map{1};
figure(1011); subplot(1,2,1); imagesc(permute(s_pre_map,[2,1,3]));
g_pre_map = caffe('forward_gnet', {gfea1});
g_pre_map = g_pre_map{1};
figure(1011); subplot(1,2,2); imagesc(permute(g_pre_map,[2,1,3]));
s_diff = s_pre_map-permute(single(map1), [2,1,3]);
input_diff = caffe('backward_snet', {single(s_diff)});
caffe('update_snet');
g_diff = g_pre_map-permute(single(map1), [2,1,3]);
input_diff = caffe('backward_gnet', {single(g_diff)});
caffe('update_gnet');
fprintf('Iteration %03d/%03d, Local Loss %f, Global Loss %f\n', i, max_iter, sum(abs(s_diff(:))), sum(abs(g_diff(:))));
end
%% ================================================================
close all;
position = zeros(6, fnum-im1_id+1);
best_geo_param = loc2affgeo(location, pf_param.p_sz);
for im2_id = im1_id:fnum
s_distractor = false;
g_distractor = false;
caffe('set_phase_test');
location_last = location;
fprintf('Processing Img: %d/%d, ', im2_id, fnum);
im2_name = sprintf([data_path 'img/%04d.jpg'], im2_id);
im2 = double(imread(im2_name));
if size(im2,3)~=3
im2(:,:,2) = im2(:,:,1);
im2(:,:,3) = im2(:,:,1);
end
%% extract roi and display
[roi2, roi_pos, padded_zero_map, pad] = ext_roi(im2, location, [0, 0], roi_size, s2);
%% draw particles
geo_param = drawparticals(best_geo_param, pf_param);
%% preprocess roiroi
roi2 = impreprocess(roi2);
fea2 = caffe('forward', {single(roi2)});
lfea2 = fea2{1};
lfea2 = lfea2(:,:,lid);
gfea2 = imresize(fea2{2}, fea_sz(1:2));
gfea2 = gfea2(:,:,gid);
%% compute confidence map
s_pre_map = caffe('forward_snet', {lfea2});
s_pre_map = permute(s_pre_map{1}, [2,1,3])/(max(s_pre_map{1}(:))+eps);
g_pre_map = caffe('forward_gnet', {gfea2});
g_pre_map = permute(g_pre_map{1}, [2,1,3])/(max(g_pre_map{1}(:))+eps);
%% compute global confidence
g_roi_map = imresize(g_pre_map, roi_pos(4:-1:3));
g_im_map = padded_zero_map;
g_im_map(roi_pos(2):roi_pos(2)+roi_pos(4)-1, roi_pos(1):roi_pos(1)+roi_pos(3)-1) = g_roi_map;
g_im_map = g_im_map(pad+1:end-pad, pad+1:end-pad);
g_im_map = double(g_im_map>0.1).*g_im_map;
wmaps = warpimg(g_im_map, affparam2mat(geo_param), [pf_param.p_sz, pf_param.p_sz]);
g_conf = reshape(sum(sum(wmaps))/pf_param.p_sz^2, [], 1);
g_rank_conf = g_conf.*(pf_param.p_sz^2*geo_param(3,:)'.*geo_param(3,:)'.*geo_param(5,:)').^0.7;
[~, g_maxid] = max(g_rank_conf);
%% compute local confidence
s_roi_map = imresize(s_pre_map, roi_pos(4:-1:3));
s_im_map = padded_zero_map;
s_im_map(roi_pos(2):roi_pos(2)+roi_pos(4)-1, roi_pos(1):roi_pos(1)+roi_pos(3)-1) = s_roi_map;
s_im_map = s_im_map(pad+1:end-pad, pad+1:end-pad);
s_im_map = double(s_im_map>0.1).*s_im_map;
wmaps = warpimg(s_im_map, affparam2mat(geo_param), [pf_param.p_sz, pf_param.p_sz]);
s_conf = reshape(sum(sum(wmaps))/pf_param.p_sz^2, [], 1);
s_rank_conf = s_conf.*(pf_param.p_sz^2*geo_param(3,:)'.*geo_param(3,:)'.*geo_param(5,:)').^0.75;
[~, s_maxid] = max(s_rank_conf);
%% gnet detect distractor
potential_location = affgeo2loc(geo_param(:, g_maxid), pf_param.p_sz);
px1 = min(max(potential_location(1),1), size(im2, 2));
px2 = min(max(px1+potential_location(3)-1,1), size(im2, 2));
py1 = min(max(potential_location(2),1), size(im2, 1));
py2 = min(max(py1+potential_location(4)-1,1), size(im2, 1));
rectified_im_map = single(g_im_map>0.2).*single(g_im_map);
inside_conf = sum(sum(rectified_im_map(py1:py2, px1:px2)));
outside_conf = sum(sum(rectified_im_map)) - inside_conf;
if outside_conf >= 0.2*inside_conf
g_distractor = true;
end
%% snet detect distractor
rectified_im_map = single(s_im_map>0.01).*single(s_im_map);
inside_conf = sum(sum(rectified_im_map(py1:py2, px1:px2)));
outside_conf = sum(sum(rectified_im_map)) - inside_conf;
if outside_conf >= 0.2*inside_conf
s_distractor = true;
end
if g_distractor %|| s_distractor
maxconf = s_conf(s_maxid);
maxid = s_maxid;
pre_map = s_roi_map;
else
maxconf = g_conf(g_maxid);
maxid = g_maxid;
pre_map = g_roi_map;
end
fprintf('lmaxconf = %f, gmaxconf = %f\n', s_conf(s_maxid), g_conf(g_maxid));
if maxconf>pf_param.mv_thr
location = affgeo2loc(geo_param(:, maxid), pf_param.p_sz);
best_geo_param = geo_param(:, maxid);
elseif s_conf(s_maxid)>pf_param.mv_thr
location = affgeo2loc(geo_param(:, s_maxid), pf_param.p_sz);
best_geo_param = geo_param(:, s_maxid);
maxconf = s_conf(s_maxid);
end
if maxconf< pf_param.up_thr
best_geo_param([3,5]) = position([3,5], im2_id-im1_id);
location = affgeo2loc( best_geo_param, pf_param.p_sz);
end
drawresult(im2_id, mat2gray(im2), [pf_param.p_sz, pf_param.p_sz], affparam2mat(best_geo_param));
position(:, im2_id- im1_id +1) = best_geo_param;
mask = mat2gray(imresize(pre_map, [roi_size, roi_size]));
if maxconf>conf_store && maxconf>pf_param.up_thr
l_off = location_last(1:2)-location(1:2);
map = GetMap(size(im2), fea_sz, roi_size, location, l_off, s2, 'gaussian');
fea2_store = lfea2;
map2_store = map;
conf_store = maxconf;
end
if conf_store>pf_param.up_thr && mod(im2_id-im1_id +1,20) == 0 % && ~l_distractor_store
caffe('set_phase_train');
caffe('reshape_input', 'ssolver', [0, 2, length(lid), fea_sz(2), fea_sz(1)]);
fea2_train{1}(:,:,:,1) = lfea1;
fea2_train{1}(:,:,:,2) = fea2_store;
s_pre_map = caffe('forward_snet', fea2_train);
diff{1}(:,:,:,1) = 0.5*(s_pre_map{1}(:,:,:,1)-permute(single(map1), [2,1,3]));
diff{1}(:,:,:,2) = 0.5*(s_pre_map{1}(:,:,:,2)-permute(single(map2_store), [2,1,3]));
caffe('backward_snet', diff);
caffe('update_snet');
conf_store = pf_param.up_thr;
caffe('reshape_input', 'ssolver', [0, 1, length(lid), fea_sz(2), fea_sz(1)]);
end
if s_distractor && maxconf> pf_param.up_thr
caffe('set_phase_train');
caffe('reshape_input', 'ssolver', [0, 2,length(lid), fea_sz(2), fea_sz(1)]);
lfea2_train(:,:,:,1) = fea2_store;
lfea2_train(:,:,:,2) = lfea2;
l_off = location_last(1:2)-location(1:2);
map = GetMap(size(im2), fea_sz, roi_size, location, l_off, s2, 'gaussian');
iter = 10;
% diff = cell(1);
for i=1:iter
s_pre_map = caffe('forward_snet', {lfea2_train});
diff{1}(:,:,:,1) = 0.5*(s_pre_map{1}(:,:,:,1)-permute(single(map2_store), [2,1,3]));
diff{1}(:,:,:,2) = 0.5*squeeze(s_pre_map{1}(:,:,:,2)-permute(single(map),[2,1,3])).*permute(single(map<=0), [2,1,3]);
caffe('backward_snet', diff);
caffe('update_snet');
end
caffe('reshape_input', 'ssolver', [0, 1, length(lid), fea_sz(2), fea_sz(1)]);
end
if pf_param.minconf > maxconf
pf_param.minconf = maxconf;
end
if im2_id == check_num
pf_param = reestimate_param(pf_param);
end
end
%% save results
% save([track_res '/position.mat'], 'position');
end