Skip to content

Commit 4d76745

Browse files
alanlukezicalanlukezic
committed
Initial commit
1 parent 4038687 commit 4d76745

32 files changed

+4411
-0
lines changed

‎compile.m

+51
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,51 @@
1+
% compile hog and segmetation mex files
2+
% before compiling set the following variables to the correct paths:
3+
% opencv_include and opencv_libpath
4+
5+
current_folder = pwd;
6+
mkdir('mex');
7+
8+
cd(['mex_src' filesep 'hog']);
9+
mex gradientMex.cpp
10+
movefile('*.mex*', [current_folder filesep 'mex'])
11+
cd(current_folder);
12+
13+
if ispc % Windows machine
14+
% set opencv include path
15+
opencv_include = 'E:\development\opencv-2.4.12\opencv\build\include\';
16+
% set opencv lib path
17+
opencv_libpath = 'E:\development\opencv-2.4.12\opencv\build\x64\vc11\lib\';
18+
19+
files = dir([opencv_libpath '*opencv*.lib']);
20+
lib = [];
21+
for i = 1:length(files),
22+
lib = [lib ' -l' files(i).name(1:end-4)];
23+
end
24+
25+
cd(['mex_src' filesep 'segmentation']);
26+
eval(['mex mex_extractforeground.cpp src\segment.cpp -Isrc\ -I' opencv_include ' -L' opencv_libpath ' ' lib]);
27+
eval(['mex mex_extractbackground.cpp src\segment.cpp -Isrc\ -I' opencv_include ' -L' opencv_libpath ' ' lib]);
28+
eval(['mex mex_segment.cpp src\segment.cpp -Isrc\ -I' opencv_include ' -L' opencv_libpath ' ' lib]);
29+
movefile('*.mex*', [current_folder filesep 'mex'])
30+
cd(current_folder);
31+
32+
elseif isunix % Unix machine
33+
% set opencv include path
34+
opencv_include = '/usr/local/include/';
35+
% set opencv lib path
36+
opencv_libpath = '/usr/lib/x86_64-linux-gnu/';
37+
38+
lib = [];
39+
files = dir([opencv_libpath '*opencv*.so']);
40+
for i = 1:length(files)
41+
lib = [lib ' -l' files(i).name(4:end-3)];
42+
end
43+
44+
cd(['mex_src' filesep 'segmentation']);
45+
eval(['mex mex_extractforeground.cpp src/segment.cpp -Isrc/ -I' opencv_include ' -L' opencv_libpath ' ' lib]);
46+
eval(['mex mex_extractbackground.cpp src/segment.cpp -Isrc/ -I' opencv_include ' -L' opencv_libpath ' ' lib]);
47+
eval(['mex mex_segment.cpp src/segment.cpp -Isrc/ -I' opencv_include ' -L' opencv_libpath ' ' lib]);
48+
movefile('*.mex*', [current_folder filesep 'mex'])
49+
cd(current_folder);
50+
51+
end

‎create_csr_tracker.m

+257
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,257 @@
1+
function tracker = create_csr_tracker(img, init_bbox, init_params)
2+
3+
if nargin < 3
4+
init_params = read_default_csr_parameters();
5+
end
6+
7+
% transform polygon to axis-aligned bbox
8+
if numel(init_bbox) > 4,
9+
bb8 = round(init_bbox(:));
10+
x1 = round(min(bb8(1:2:end)));
11+
x2 = round(max(bb8(1:2:end)));
12+
y1 = round(min(bb8(2:2:end)));
13+
y2 = round(max(bb8(2:2:end)));
14+
bb = round([x1, y1, x2 - x1, y2 - y1]);
15+
16+
init_mask = poly2mask(bb8(1:2:end)-bb(1), bb8(2:2:end)-bb(2), bb(4), bb(3));
17+
else
18+
bb = round(init_bbox);
19+
init_mask = ones(bb(4), bb(3));
20+
end
21+
22+
23+
% filter parameters
24+
padding = init_params.padding; % padding
25+
learning_rate = init_params.learning_rate; % learning rate for updating filter
26+
feature_type = init_params.feature_type;
27+
28+
% load and store pre-computed lookup table for colornames
29+
w2c = [];
30+
if sum(strcmp(feature_type, 'cn'))
31+
w2c = load('w2crs.mat');
32+
w2c = w2c.w2crs;
33+
end
34+
35+
% segmentation parameters
36+
hist_lr = init_params.hist_lr;
37+
nbins = init_params.nbins; % N bins for segmentation
38+
seg_colorspace = init_params.seg_colorspace; % 'rgb' or 'hsv'
39+
use_segmentation = init_params.use_segmentation; % false to disable use of segmentation
40+
mask_diletation_type = init_params.mask_diletation_type; % for function strel (square, disk, ...)
41+
mask_diletation_sz = init_params.mask_diletation_sz;
42+
43+
% check if grayscale image (only 1 channel) or
44+
% check if grayscale image (3 the same channels)
45+
img0 = bsxfun(@minus, double(img), mean(img,3));
46+
if size(img,3) < 3 || sum(abs(img0(:))) < 10
47+
use_segmentation = false;
48+
% also do not use colornames
49+
[isused_cn, cn_idx] = ismember('cn', feature_type);
50+
if isused_cn
51+
feature_type(cn_idx) = [];
52+
end
53+
end
54+
55+
% features parameters
56+
cell_size = 1.0;
57+
if sum(strcmp(feature_type, 'hog'))
58+
cell_size = min(4, max(1, ceil((bb(3)*bb(4))/400)));
59+
end
60+
61+
% size parameters
62+
% reference target size: [width, height]
63+
base_target_sz = [bb(3), bb(4)];
64+
% reference template size: [w, h], does not change during tracking
65+
template_size = floor(base_target_sz + padding*sqrt(prod(base_target_sz)));
66+
template_size = mean(template_size);
67+
template_size = [template_size, template_size];
68+
69+
% rescale template after extracting to have fixed area
70+
rescale_ratio = sqrt((200^2) / (template_size(1) * template_size(2)));
71+
if rescale_ratio > 1 % if already smaller - do not rescale
72+
rescale_ratio = 1;
73+
end
74+
75+
rescale_template_size = floor(rescale_ratio * template_size);
76+
77+
% position of the target center
78+
c = bb([1,2]) + base_target_sz/2;
79+
80+
% create gaussian shaped labels
81+
sigma = init_params.y_sigma;
82+
Y = fft2(gaussian_shaped_labels(1,sigma, floor(rescale_template_size([2,1]) / cell_size)));
83+
84+
%store pre-computed cosine window
85+
cos_win = hann(size(Y,1)) * hann(size(Y,2))';
86+
87+
% scale adaptation parameters (from DSST)
88+
currentScaleFactor = init_params.currentScaleFactor;
89+
n_scales = init_params.n_scales;
90+
scale_model_factor = init_params.scale_model_factor;
91+
scale_sigma_factor = init_params.scale_sigma_factor;
92+
scale_step = init_params.scale_step;
93+
scale_model_max_area = init_params.scale_model_max_area;
94+
scale_sigma = sqrt(n_scales) * scale_sigma_factor;
95+
scale_lr = init_params.scale_lr; % learning rate parameter
96+
97+
%label function for the scales
98+
ss = (1:n_scales) - ceil(n_scales/2);
99+
ys = exp(-0.5 * (ss.^2) / scale_sigma^2);
100+
ysf = single(fft(ys));
101+
102+
if mod(n_scales,2) == 0
103+
scale_window = single(hann(n_scales+1));
104+
scale_window = scale_window(2:end);
105+
else
106+
scale_window = single(hann(n_scales));
107+
end
108+
109+
ss = 1:n_scales;
110+
scaleFactors = scale_step.^(ceil(n_scales/2) - ss);
111+
112+
template_size_ = template_size;
113+
if scale_model_factor^2 * prod(template_size_) > scale_model_max_area
114+
scale_model_factor = sqrt(scale_model_max_area/prod(template_size_));
115+
end
116+
117+
scale_model_sz = floor(template_size_ * scale_model_factor);
118+
scaleSizeFactors = scaleFactors;
119+
min_scale_factor = scale_step ^ ceil(log(max(5 ./ template_size_)) / log(scale_step));
120+
max_scale_factor = scale_step ^ floor(log(min([size(img,1) size(img,2)] ./ base_target_sz)) / log(scale_step));
121+
122+
% create dummy mask (approximation for segmentation)
123+
% size of the object in feature space
124+
obj_size = floor(rescale_ratio * (base_target_sz/cell_size));
125+
x0 = floor((size(Y,2)-obj_size(1))/2);
126+
y0 = floor((size(Y,1)-obj_size(2))/2);
127+
x1 = x0 + obj_size(1);
128+
y1 = y0 + obj_size(2);
129+
target_dummy_mask = zeros(size(Y));
130+
target_dummy_mask(y0:y1, x0:x1) = 1;
131+
target_dummy_mask = single(target_dummy_mask);
132+
133+
target_dummy_area = sum(target_dummy_mask(:));
134+
135+
if use_segmentation
136+
% convert image in desired colorspace
137+
if strcmp(seg_colorspace, 'rgb')
138+
seg_img = img;
139+
elseif strcmp(seg_colorspace, 'hsv')
140+
seg_img = rgb2hsv(img);
141+
seg_img = seg_img * 255;
142+
else
143+
error('Unknown colorspace parameter');
144+
end
145+
146+
% object rectangle region (to zero-based coordinates)
147+
obj_reg = [bb(1), bb(2), bb(1)+bb(3), bb(2)+bb(4)] - [1 1 1 1];
148+
149+
% extract histograms
150+
hist_fg = mex_extractforeground(seg_img, obj_reg, nbins);
151+
hist_bg = mex_extractbackground(seg_img, obj_reg, nbins);
152+
153+
% extract masked patch: mask out parts outside image
154+
[seg_patch, valid_pixels_mask] = get_patch(seg_img, c, currentScaleFactor, template_size);
155+
156+
% segmentation
157+
[fg_p, bg_p] = get_location_prior([1 1 size(seg_patch, 2) size(seg_patch, 1)], base_target_sz, [size(seg_patch,2), size(seg_patch, 1)]);
158+
[~, fg, ~] = mex_segment(seg_patch, hist_fg, hist_bg, nbins, fg_p, bg_p);
159+
160+
% cut out regions outside from image
161+
mask = single(fg).*single(valid_pixels_mask);
162+
mask = binarize_softmask(mask);
163+
164+
% use mask from init pose
165+
init_mask_padded = zeros(size(mask));
166+
pm_x0 = floor(size(init_mask_padded,2) / 2 - size(init_mask,2) / 2);
167+
pm_y0 = floor(size(init_mask_padded,1) / 2 - size(init_mask,1) / 2);
168+
init_mask_padded(pm_y0:pm_y0+size(init_mask,1)-1, pm_x0:pm_x0+size(init_mask,2)-1) = init_mask;
169+
mask = mask.*single(init_mask_padded);
170+
171+
% resize to filter size
172+
mask = imresize(mask, size(Y), 'nearest');
173+
174+
% check if mask is too small (probably segmentation is not ok then)
175+
if mask_normal(mask, target_dummy_area)
176+
if mask_diletation_sz > 0
177+
D = strel(mask_diletation_type, mask_diletation_sz);
178+
mask = imdilate(mask, D);
179+
end
180+
else
181+
mask = target_dummy_mask;
182+
end
183+
184+
else
185+
mask = target_dummy_mask;
186+
end
187+
188+
% extract features
189+
f = get_csr_features(img, c, currentScaleFactor, template_size, ...
190+
rescale_template_size, cos_win, feature_type, w2c, cell_size);
191+
192+
% create filter - using segmentation mask
193+
H = create_csr_filter(f, Y, single(mask));
194+
195+
% calculate per-channel feature weights
196+
response = real(ifft2(fft2(f).*conj(H)));
197+
chann_w = max(reshape(response, [size(response,1)*size(response,2), size(response,3)]), [], 1);
198+
% normalize: sum = 1
199+
chann_w = chann_w / sum(chann_w);
200+
201+
% make a scale search model aswell
202+
xs = get_scale_subwindow(img, c([2,1]), base_target_sz([2,1]), ...
203+
currentScaleFactor * scaleSizeFactors, scale_window, scale_model_sz([2,1]), []);
204+
% fft over the scale dim
205+
xsf = fft(xs,[],2);
206+
new_sf_num = bsxfun(@times, ysf, conj(xsf));
207+
new_sf_den = sum(xsf .* conj(xsf), 1);
208+
209+
% store all important const's and variables to the tracker structure
210+
tracker.feature_type = feature_type;
211+
tracker.padding = padding;
212+
tracker.learning_rate = learning_rate; % filter learning rate
213+
tracker.cell_size = cell_size;
214+
tracker.H = H;
215+
tracker.weight_lr = init_params.channels_weight_lr;
216+
tracker.use_channel_weights = init_params.use_channel_weights;
217+
tracker.chann_w = chann_w;
218+
tracker.Y = Y;
219+
tracker.mask_diletation_type = mask_diletation_type;
220+
tracker.mask_diletation_sz = mask_diletation_sz;
221+
tracker.target_dummy_mask = target_dummy_mask;
222+
tracker.target_dummy_area = target_dummy_area;
223+
tracker.use_segmentation = use_segmentation;
224+
tracker.bb = bb;
225+
tracker.cos_win = cos_win;
226+
tracker.w2c = w2c;
227+
tracker.template_size = template_size;
228+
tracker.obj_size = obj_size;
229+
tracker.c = c;
230+
tracker.nbins = nbins;
231+
tracker.currentScaleFactor = currentScaleFactor;
232+
tracker.rescale_template_size = rescale_template_size;
233+
tracker.rescale_ratio = rescale_ratio;
234+
if use_segmentation
235+
tracker.hist_fg = hist_fg;
236+
tracker.hist_bg = hist_bg;
237+
tracker.hist_lr = hist_lr;
238+
tracker.seg_colorspace = seg_colorspace;
239+
end
240+
tracker.ysf = ysf;
241+
tracker.sf_num = new_sf_num;
242+
tracker.sf_den = new_sf_den;
243+
tracker.scale_lr = scale_lr;
244+
tracker.base_target_sz = base_target_sz;
245+
tracker.scaleSizeFactors = scaleSizeFactors;
246+
tracker.scale_window = scale_window;
247+
tracker.scale_model_sz = scale_model_sz;
248+
tracker.scaleFactors = scaleFactors;
249+
tracker.min_scale_factor = min_scale_factor;
250+
tracker.max_scale_factor = max_scale_factor;
251+
tracker.mask = mask;
252+
253+
tracker.H_prev = H;
254+
255+
tracker.img_prev = img;
256+
257+
end % endfunction

‎csr_wrapper.m

+54
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,54 @@
1+
function csr_wrapper(varargin)
2+
% csr_wrapper Wrapper function for VOT-like interface
3+
%
4+
% This function is a wrapper for the CSR-DCF tracker using the VOT interface for input/output
5+
% data loading (i.e. images, initial region) and storing (i.e. output.txt or trax output)
6+
%
7+
% Input:
8+
% - varargin[OutputDir] (string): path to dir where some output images/data should be stored
9+
10+
cleanup = onCleanup(@() exit() ); % Always call exit command at the end to terminate Matlab!
11+
RandStream.setGlobalStream(RandStream('mt19937ar', 'Seed', sum(clock))); % Set random seed to a different value every time as required by the VOT rules.
12+
13+
[handle, image_path, region] = vot('polygon'); % Obtain communication object
14+
image = imread(image_path); % Read first image from file
15+
16+
% matlab indexing: starts with (1,1); gt starts with (0,0)
17+
if numel(region) > 4
18+
% all x,y points shifted by 1
19+
region = region + 1;
20+
else
21+
% shift x,y by 1
22+
region(1:2) = region(1:2) + 1;
23+
end
24+
25+
tracker = create_csr_tracker(image, region);
26+
27+
while true
28+
29+
[handle, image_path] = handle.frame(handle); % Get the next frame
30+
31+
if isempty(image_path) % Are we done?
32+
break;
33+
end;
34+
35+
image = imread(image_path);
36+
37+
[tracker, region] = track_csr_tracker(tracker, image);
38+
39+
% matlab indexing: starts with (1,1); gt starts with (0,0)
40+
if numel(region) > 4
41+
% all x,y points shifted by 1
42+
region = region - 1;
43+
else
44+
% shift x,y by 1
45+
region(1:2) = region(1:2) - 1;
46+
end
47+
48+
handle = handle.report(handle, double(region)); % Report position for the given frame
49+
50+
end;
51+
52+
handle.quit(handle); % Output the results and clear the resources
53+
54+
end % endfunction

0 commit comments

Comments
 (0)