DPM(Defomable Parts Model) 源码分析-检测（二）

DPM(Defomable Parts Model)原理
首先声明此版本为V3.1。因为和论文最相符。V4增加了模型数由2个增加为6个，V5提取了语义特征。源码太长纯代码应该在2K+,只选取了核心部分代码
demo.m

function demo()

test('000034.jpg', 'car');
test('000061.jpg', 'person');
test('000084.jpg', 'bicycle');

function test(name, cls)
% load and display image
im=imread(name);
clf;
image(im);
axis equal;
axis on;
disp('input image');
disp('press any key to continue'); pause;

% load and display model
load(['VOC2007/' cls '_final']); %加载模型
visualizemodel(model);
disp([cls ' model']);
disp('press any key to continue'); pause;

% detect objects
boxes = detect(im, model, 0); %model为mat中的结构体
top = nms(boxes, 0.5);  %Non-maximum suppression.
showboxes(im, top);
%print(gcf, '-djpeg90', '-r0', [cls '.jpg']);
disp('detections');
disp('press any key to continue'); pause;

% get bounding boxes
bbox = getboxes(model, boxes);	%根据检测到的root，parts，预测bounding
top = nms(bbox, 0.5);
bbox = clipboxes(im, top); %预测出来的bounding，可能会超过图像原始尺寸，所以要减掉
showboxes(im, bbox);
disp('bounding boxes');
disp('press any key to continue'); pause;

detect.m

function [boxes] = detect(input, model, thresh, bbox, ...
overlap, label, fid, id, maxsize)
% 论文 fig.4

% boxes = detect(input, model, thresh, bbox, overlap, label, fid, id, maxsize)
% Detect objects in input using a model and a score threshold.
% Higher threshold leads to fewer detections.
% boxes = [rx1 ry1 rx2 ry2 | px1 py1 px2 py2 ...| componetindex | score ]
% The function returns a matrix with one row per detected object.  The
% last column of each row gives the score of the detection.  The
% column before last specifies the component used for the detection.
% The first 4 columns specify the bounding box for the root filter and
% subsequent columns specify the bounding boxes of each part.
%
% If bbox is not empty, we pick best detection with significant overlap.
% If label and fid are included, we write feature vectors to a data file.

%phase 2： im, model, 0, bbox, overlap, 1, fid, 2*i-1
% trian boxex : detect(im, model, 0, bbox, overlap)
if nargin > 3 && ~isempty(bbox)
latent = true;
else
latent = false;
end

if nargin > 6 && fid ~= 0
write = true;
else
write = false;
end

if nargin < 9
maxsize = inf;
end

% we assume color images
input = color(input);	%如果是灰度图，扩充为三通道 R=G=B=Gray

% prepare model for convolutions
rootfilters = [];
for i = 1:length(model.rootfilters) %
rootfilters{i} = model.rootfilters{i}.w;% r*w*31维向量，9（方向范围 0~180） +18（方向范围 0-360）+4（cell熵和）
end
partfilters = [];
for i = 1:length(model.partfilters)
partfilters{i} = model.partfilters{i}.w;
end

% cache some data 获取所有 root，part的所有信息
for c = 1:model.numcomponents	% releas3.1 一种对象，只有2个模型，releas5 有3*2个模型
ridx{c} = model.components{c}.rootindex; % m1=1,m2=2
oidx{c} = model.components{c}.offsetindex; %o1=1,o2=2
root{c} = model.rootfilters{ridx{c}}.w;
rsize{c} = [size(root{c},1) size(root{c},2)]; %root size,单位为 sbin*sbin的block块，相当于原始HOG中的一个cell
numparts{c} = length(model.components{c}.parts); %目前为固定值6个,但是有些part是 fake
for j = 1:numparts{c}
pidx{c,j} = model.components{c}.parts{j}.partindex; %part是在该对象的所有component的part下连续编号
didx{c,j} = model.components{c}.parts{j}.defindex;  % 在 rootfiter中的 anchor location
part{c,j} = model.partfilters{pidx{c,j}}.w; % 6*6*31
psize{c,j} = [size(part{c,j},1) size(part{c,j},2)];	%
% reverse map from partfilter index to (component, part#)
rpidx{pidx{c,j}} = [c j];
end
end

% we pad the feature maps to detect partially visible objects
padx = ceil(model.maxsize(2)/2+1); % 7/2+1 = 5
pady = ceil(model.maxsize(1)/2+1); % 11/2+1 = 7

% the feature pyramid
interval = model.interval;	%10
%--------------------------------特征金字塔---------------------------------------------------------
% feat的尺寸为 img.rows/sbin,img.cols/sbin
% scales:缩放了多少
[feat, scales] = featpyramid(input, model.sbin, interval); % 8，10

% detect at each scale
best = -inf;
ex = [];
boxes = [];
%---------------------逐层检测目标-----------------------------------------------------------%
for level = interval+1:length(feat) %注意是从第二层开始
scale = model.sbin/scales(level);  % 1/缩小了多少
if size(feat{level}, 1)+2*pady < model.maxsize(1) || ... %扩展后还是未能达到 能同时计算两个component的得分
size(feat{level}, 2)+2*padx < model.maxsize(2) || ...
(write && ftell(fid) >= maxsize) %已经没有空间保存样本了
continue;
end

if latent	%训练时使用，检测时跳过
skip = true;
for c = 1:model.numcomponents
root_area = (rsize{c}(1)*scale) * (rsize{c}(2)*scale);% rootfilter
box_area = (bbox(3)-bbox(1)+1) * (bbox(4)-bbox(2)+1); % bbox该class 所有 rootfilter 的交集即minsize
if (root_area/box_area) >= overlap && (box_area/root_area) >= overlap %这句话真纠结，a>=0.7b,b>=0.7a -> a>=0.7b>=0.49a
skip = false;
end
end
if skip
continue;
end
end

% -----------convolve feature maps with filters -----------
%rootmatch,partmatch ,得分图root的尺度总是part的一半，
%rootmatch尺寸是partmatch的一半
featr = padarray(feat{level}, [pady padx 0], 0);	% 上下各补充 pady 行0，左右各补充padx行 0
%C = fconv(A, cell of B, start, end);
rootmatch = fconv(featr, rootfilters, 1, length(rootfilters));
if length(partfilters) > 0
featp = padarray(feat{level-interval}, [2*pady 2*padx 0], 0);
partmatch = fconv(featp, partfilters, 1, length(partfilters));
end
%-------------------逐component检测-----------------------------------
% 参见论文 Fig 4
% 最终得到  综合得分图   score
for c = 1:model.numcomponents
% root score + offset
score = rootmatch{ridx{c}} + model.offsets{oidx{c}}.w;
% add in parts
for j = 1:numparts{c}
def = model.defs{didx{c,j}}.w;
anchor = model.defs{didx{c,j}}.anchor;
% the anchor position is shifted to account for misalignment
% between features at different resolutions
ax{c,j} = anchor(1) + 1; %
ay{c,j} = anchor(2) + 1;
match = partmatch{pidx{c,j}};
[M, Ix{c,j}, Iy{c,j}] = dt(-match, def(1), def(2), def(3), def(4)); % dx,dy,dx^2,dy^2的偏移惩罚系数
% M part的综合匹配得分图，与part尺寸一致。Ix{c,j}, Iy{c,j} 即part实际的最佳位置（相对于root）
% 参见论文公式 9
score = score - M(ay{c,j}:2:ay{c,j}+2*(size(score,1)-1), ...
ax{c,j}:2:ax{c,j}+2*(size(score,2)-1));
end

%-------阈值淘汰------------------------
if ~latent
% get all good matches
% ---thresh  在 分类时为0，在 找 hard exmaple 时是 -1.05--
I = find(score > thresh);	%返回的是从上到下从左到右的索引
[Y, X] = ind2sub(size(score), I);  %还原为 行，列坐标
tmp = zeros(length(I), 4*(1+numparts{c})+2);	%一个目标的root，part，score信息，见程序开头说明
for i = 1:length(I)
x = X(i);
y = Y(i);
[x1, y1, x2, y2] = rootbox(x, y, scale, padx, pady, rsize{c});
b = [x1 y1 x2 y2];
if write
rblocklabel = model.rootfilters{ridx{c}}.blocklabel;
oblocklabel = model.offsets{oidx{c}}.blocklabel;
f = featr(y:y+rsize{c}(1)-1, x:x+rsize{c}(2)-1, :);
xc = round(x + rsize{c}(2)/2 - padx); %
yc = round(y + rsize{c}(1)/2 - pady);
ex = [];
ex.header = [label; id; level; xc; yc; ...
model.components{c}.numblocks; ...
model.components{c}.dim];
ex.offset.bl = oblocklabel;
ex.offset.w = 1;
ex.root.bl = rblocklabel;
width1 = ceil(rsize{c}(2)/2);
width2 = floor(rsize{c}(2)/2);
f(:,1:width2,:) = f(:,1:width2,:) + flipfeat(f(:,width1+1:end,:));
ex.root.w = f(:,1:width1,:);
ex.part = [];
end
for j = 1:numparts{c}
[probex, probey, px, py, px1, py1, px2, py2] = ...
partbox(x, y, ax{c,j}, ay{c,j}, scale, padx, pady, ...
psize{c,j}, Ix{c,j}, Iy{c,j});
b = [b px1 py1 px2 py2];
if write
if model.partfilters{pidx{c,j}}.fake
continue;
end
pblocklabel = model.partfilters{pidx{c,j}}.blocklabel;
dblocklabel = model.defs{didx{c,j}}.blocklabel;
f = featp(py:py+psize{c,j}(1)-1,px:px+psize{c,j}(2)-1,:);
def = -[(probex-px)^2; probex-px; (probey-py)^2; probey-py];
partner = model.partfilters{pidx{c,j}}.partner;
if partner > 0
k = rpidx{partner}(2);
[kprobex, kprobey, kpx, kpy, kpx1, kpy1, kpx2, kpy2] = ...
partbox(x, y, ax{c,k}, ay{c,k}, scale, padx, pady, ...
psize{c,k}, Ix{c,k}, Iy{c,k});
kf = featp(kpy:kpy+psize{c,k}(1)-1,kpx:kpx+psize{c,k}(2)-1,:);
% flip linear term in horizontal deformation model
kdef = -[(kprobex-kpx)^2; kpx-kprobex; ...
(kprobey-kpy)^2; kprobey-kpy];
f = f + flipfeat(kf);
def = def + kdef;
else
width1 = ceil(psize{c,j}(2)/2);
width2 = floor(psize{c,j}(2)/2);
f(:,1:width2,:) = f(:,1:width2,:) + flipfeat(f(:,width1+1:end,:));
f = f(:,1:width1,:);
end
ex.part(j).bl = pblocklabel;
ex.part(j).w = f;
ex.def(j).bl = dblocklabel;
ex.def(j).w = def;
end
end
if write
exwrite(fid, ex); % 写入负样本
end
tmp(i,:) = [b c score(I(i))];
end
boxes = [boxes; tmp];
end

if latent
% get best match
for x = 1:size(score,2)
for y = 1:size(score,1)
if score(y, x) > best
% 以该(y，x)为left-top点的rootfilter的范围在原图像中的位置
[x1, y1, x2, y2] = rootbox(x, y, scale, padx, pady, rsize{c});
% intesection with bbox
xx1 = max(x1, bbox(1));
yy1 = max(y1, bbox(2));
xx2 = min(x2, bbox(3));
yy2 = min(y2, bbox(4));
w = (xx2-xx1+1);
h = (yy2-yy1+1);
if w > 0 && h > 0
% check overlap with bbox
inter = w*h;
a = (x2-x1+1) * (y2-y1+1); % rootfilter 的面积
b = (bbox(3)-bbox(1)+1) * (bbox(4)-bbox(2)+1); % bbox的面积
% 计算很很独特，如果只是 inter / b 那么 如果a很大，只是一部分与 bounding box重合,那就不可靠了，人再怎么标注错误，也不会这么大
% 所以，a越大，要求的重合率越高才好，所以分母+a,是个不错的选择，但是这样减小的太多了，所以减去 inter
o = inter / (a+b-inter);
if (o >= overlap)
%
best = score(y, x);
boxes = [x1 y1 x2 y2];
% 这一部分一直被覆盖，最后保留的是 best样本
if write
f = featr(y:y+rsize{c}(1)-1, x:x+rsize{c}(2)-1, :);
rblocklabel = model.rootfilters{ridx{c}}.blocklabel;
oblocklabel = model.offsets{oidx{c}}.blocklabel;
xc = round(x + rsize{c}(2)/2 - padx);
yc = round(y + rsize{c}(1)/2 - pady);
ex = [];
% label; id; level; xc; yc,正样本的重要信息！
% xc,yc,居然是相对于剪切后的图片
ex.header = [label; id; level; xc; yc; ...
model.components{c}.numblocks; ...
model.components{c}.dim];
ex.offset.bl = oblocklabel;
ex.offset.w = 1;
ex.root.bl = rblocklabel;
width1 = ceil(rsize{c}(2)/2);
width2 = floor(rsize{c}(2)/2);
f(:,1:width2,:) = f(:,1:width2,:) + flipfeat(f(:,width1+1:end,:));
ex.root.w = f(:,1:width1,:); %样本特征
ex.part = [];
end
for j = 1:numparts{c}
%probex，probey综合得分最高的位置，相对于featp
%px1，py1，px2，py2 转化成相对于featr
[probex, probey, px, py, px1, py1, px2, py2] = ...
partbox(x, y, ax{c,j}, ay{c,j}, scale, ...
padx, pady, psize{c,j}, Ix{c,j}, Iy{c,j});
boxes = [boxes px1 py1 px2 py2];
if write
if model.partfilters{pidx{c,j}}.fake
continue;
end
p = featp(py:py+psize{c,j}(1)-1, ...
px:px+psize{c,j}(2)-1, :);
def = -[(probex-px)^2; probex-px; (probey-py)^2; probey-py];
pblocklabel = model.partfilters{pidx{c,j}}.blocklabel;
dblocklabel = model.defs{didx{c,j}}.blocklabel;
partner = model.partfilters{pidx{c,j}}.partner;
if partner > 0
k = rpidx{partner}(2);
[kprobex, kprobey, kpx, kpy, kpx1, kpy1, kpx2, kpy2] = ...
partbox(x, y, ax{c,k}, ay{c,k}, scale, padx, pady, ...
psize{c,k}, Ix{c,k}, Iy{c,k});
kp = featp(kpy:kpy+psize{c,k}(1)-1, ...
kpx:kpx+psize{c,k}(2)-1, :);
% flip linear term in horizontal deformation model
kdef = -[(kprobex-kpx)^2; kpx-kprobex; ...
(kprobey-kpy)^2; kprobey-kpy];
p = p + flipfeat(kp);
def = def + kdef;
else
width1 = ceil(psize{c,j}(2)/2);
width2 = floor(psize{c,j}(2)/2);
p(:,1:width2,:) = p(:,1:width2,:) + ...
flipfeat(p(:,width1+1:end,:));
p = p(:,1:width1,:);
end
ex.part(j).bl = pblocklabel;
ex.part(j).w = p;
ex.def(j).bl = dblocklabel;
ex.def(j).w = def;
end
end
boxes = [boxes c best];
end
end
end
end
end
end
end
end

if latent && write && ~isempty(ex)
exwrite(fid, ex); %datfile
end

% The functions below compute a bounding box for a root or part
% template placed in the feature hierarchy.
%
% coordinates need to be transformed to take into account:
% 1. padding from convolution
% 2. scaling due to sbin & image subsampling
% 3. offset from feature computation
%

function [x1, y1, x2, y2] = rootbox(x, y, scale, padx, pady, rsize)
x1 = (x-padx)*scale+1;	%图像是先缩放（构造金字塔时）再打补丁
y1 = (y-pady)*scale+1;
x2 = x1 + rsize(2)*scale - 1; % 宽度也要缩放
y2 = y1 + rsize(1)*scale - 1;

function [probex, probey, px, py, px1, py1, px2, py2] = ...
partbox(x, y, ax, ay, scale, padx, pady, psize, Ix, Iy)
probex = (x-1)*2+ax; %最优位置
probey = (y-1)*2+ay;
px = double(Ix(probey, probex)); %综合得分最高的位置
py = double(Iy(probey, probex));
px1 = ((px-2)/2+1-padx)*scale+1; % pading是root的两倍
py1 = ((py-2)/2+1-pady)*scale+1;
px2 = px1 + psize(2)*scale/2 - 1;
py2 = py1 + psize(1)*scale/2 - 1;

% write an example to the data file
function exwrite(fid, ex)
fwrite(fid, ex.header, 'int32');
buf = [ex.offset.bl; ex.offset.w(:); ...
ex.root.bl; ex.root.w(:)];
fwrite(fid, buf, 'single');
for j = 1:length(ex.part)
if ~isempty(ex.part(j).w)
buf = [ex.part(j).bl; ex.part(j).w(:); ...
ex.def(j).bl; ex.def(j).w(:)];
fwrite(fid, buf, 'single');
end
end

features.cc

#include <math.h>
#include "mex.h"

// small value, used to avoid division by zero
#define eps 0.0001

#define bzero(a, b) memset(a, 0, b)
int round(float a) { float tmp = a - (int)a; if( tmp >= 0.5 ) return (int)a + 1; else return (int)a; }
// unit vectors used to compute gradient orientation
// cos(20*i)
double uu[9] = {1.0000,
0.9397,
0.7660,
0.500,
0.1736,
-0.1736,
-0.5000,
-0.7660,
-0.9397};
//sin(20*i)
double vv[9] = {0.0000,
0.3420,
0.6428,
0.8660,
0.9848,
0.9848,
0.8660,
0.6428,
0.3420};

static inline double min(double x, double y) { return (x <= y ? x : y); }
static inline double max(double x, double y) { return (x <= y ? y : x); }

static inline int min(int x, int y) { return (x <= y ? x : y); }
static inline int max(int x, int y) { return (x <= y ? y : x); }

// main function:
// takes a double color image and a bin size
// returns HOG features
mxArray *process(const mxArray *mximage, const mxArray *mxsbin) {
double *im = (double *)mxGetPr(mximage);
const int *dims = mxGetDimensions(mximage);
if (mxGetNumberOfDimensions(mximage) != 3 ||
dims[2] != 3 ||
mxGetClassID(mximage) != mxDOUBLE_CLASS)
mexErrMsgTxt("Invalid input");

int sbin = (int)mxGetScalar(mxsbin);

// memory for caching orientation histograms & their norms
int blocks[2];
blocks[0] = (int)round((double)dims[0]/(double)sbin);//行
blocks[1] = (int)round((double)dims[1]/(double)sbin);//列
double *hist = (double *)mxCalloc(blocks[0]*blocks[1]*18, sizeof(double));//只需要计算18bin，9bin的推
double *norm = (double *)mxCalloc(blocks[0]*blocks[1], sizeof(double));

// memory for HOG features
int out[3];//size
out[0] = max(blocks[0]-2, 0);//减去2干嘛？？
out[1] = max(blocks[1]-2, 0);
out[2] = 27+4;
mxArray *mxfeat = mxCreateNumericArray(3, out, mxDOUBLE_CLASS, mxREAL);//特征,size=out
double *feat = (double *)mxGetPr(mxfeat);

int visible[2];
visible[0] = blocks[0]*sbin;
visible[1] = blocks[1]*sbin;
//先列再行
for (int x = 1; x < visible[1]-1; x++) {
for (int y = 1; y < visible[0]-1; y++) {
// first color channel
double *s = im + min(x, dims[1]-2)*dims[0] + min(y, dims[0]-2);//在im中的位置
double dy = *(s+1) - *(s-1);
double dx = *(s+dims[0]) - *(s-dims[0]); //坐标系是一样的，c和matlab的存储顺序不一样
double v = dx*dx + dy*dy;

// second color channel
s += dims[0]*dims[1];
double dy2 = *(s+1) - *(s-1);
double dx2 = *(s+dims[0]) - *(s-dims[0]);
double v2 = dx2*dx2 + dy2*dy2;

// third color channel
s += dims[0]*dims[1];
double dy3 = *(s+1) - *(s-1);
double dx3 = *(s+dims[0]) - *(s-dims[0]);
double v3 = dx3*dx3 + dy3*dy3;

// pick channel with strongest gradient，计算v
if (v2 > v) {
v = v2;
dx = dx2;
dy = dy2;
}
if (v3 > v) {
v = v3;
dx = dx3;
dy = dy3;
}

// snap to one of 18 orientations，就算角度best_o
double best_dot = 0;
int best_o = 0;
for (int o = 0; o < 9; o++) {
// (sinθ)^2+(cosθ)^2 =1
// max cosθ*dx+ sinθ*dy 对其求导，可得极大值 θ = arctan dy/dx
double dot = uu[o]*dx + vv[o]*dy;
if (dot > best_dot) {
best_dot = dot;
best_o = o;
} else if (-dot > best_dot) {
best_dot = -dot;
best_o = o+9;
}
}

// add to 4 histograms around pixel using linear interpolation
double xp = ((double)x+0.5)/(double)sbin - 0.5;
double yp = ((double)y+0.5)/(double)sbin - 0.5;
int ixp = (int)floor(xp);
int iyp = (int)floor(yp);
double vx0 = xp-ixp;
double vy0 = yp-iyp;
double vx1 = 1.0-vx0;
double vy1 = 1.0-vy0;
v = sqrt(v);
//左上角
if (ixp >= 0 && iyp >= 0) {
*(hist + ixp*blocks[0] + iyp + best_o*blocks[0]*blocks[1]) +=
vx1*vy1*v;
}
//右上角
if (ixp+1 < blocks[1] && iyp >= 0) {
*(hist + (ixp+1)*blocks[0] + iyp + best_o*blocks[0]*blocks[1]) +=
vx0*vy1*v;
}
//左下角
if (ixp >= 0 && iyp+1 < blocks[0]) {
*(hist + ixp*blocks[0] + (iyp+1) + best_o*blocks[0]*blocks[1]) +=
vx1*vy0*v;
}
//右下角
if (ixp+1 < blocks[1] && iyp+1 < blocks[0]) {
*(hist + (ixp+1)*blocks[0] + (iyp+1) + best_o*blocks[0]*blocks[1]) +=
vx0*vy0*v;
}
}
}

// compute energy in each block by summing over orientations
//计算每一个cell的 sum( ( v(oi)+v(oi+9) )^2 ),oi=0..8
for (int o = 0; o < 9; o++) {
double *src1 = hist + o*blocks[0]*blocks[1];
double *src2 = hist + (o+9)*blocks[0]*blocks[1];
double *dst = norm;
double *end = norm + blocks[1]*blocks[0];
while (dst < end) {
*(dst++) += (*src1 + *src2) * (*src1 + *src2);
src1++;
src2++;
}
}

// compute features
for (int x = 0; x < out[1]; x++) {
for (int y = 0; y < out[0]; y++) {
double *dst = feat + x*out[0] + y;
double *src, *p, n1, n2, n3, n4;

p = norm + (x+1)*blocks[0] + y+1;//右下角的constrain insensitive sum
n1 = 1.0 / sqrt(*p + *(p+1) + *(p+blocks[0]) + *(p+blocks[0]+1) + eps);
p = norm + (x+1)*blocks[0] + y;//右边
n2 = 1.0 / sqrt(*p + *(p+1) + *(p+blocks[0]) + *(p+blocks[0]+1) + eps);
p = norm + x*blocks[0] + y+1;//下边
n3 = 1.0 / sqrt(*p + *(p+1) + *(p+blocks[0]) + *(p+blocks[0]+1) + eps);
p = norm + x*blocks[0] + y;//自己
n4 = 1.0 / sqrt(*p + *(p+1) + *(p+blocks[0]) + *(p+blocks[0]+1) + eps);

double t1 = 0;
double t2 = 0;
double t3 = 0;
double t4 = 0;

// contrast-sensitive features
src = hist + (x+1)*blocks[0] + (y+1);
for (int o = 0; o < 18; o++) {
double h1 = min(*src * n1, 0.2);//截短
double h2 = min(*src * n2, 0.2);
double h3 = min(*src * n3, 0.2);
double h4 = min(*src * n4, 0.2);
*dst = 0.5 * (h1 + h2 + h3 + h4);//求和
t1 += h1;
t2 += h2;
t3 += h3;
t4 += h4;
dst += out[0]*out[1];//下一个bin
src += blocks[0]*blocks[1];
}

// contrast-insensitive features
src = hist + (x+1)*blocks[0] + (y+1);
for (int o = 0; o < 9; o++) {
double sum = *src + *(src + 9*blocks[0]*blocks[1]);
double h1 = min(sum * n1, 0.2);
double h2 = min(sum * n2, 0.2);
double h3 = min(sum * n3, 0.2);
double h4 = min(sum * n4, 0.2);
*dst = 0.5 * (h1 + h2 + h3 + h4);
dst += out[0]*out[1];
src += blocks[0]*blocks[1];
}

// texture features
*dst = 0.2357 * t1;
dst += out[0]*out[1];
*dst = 0.2357 * t2;
dst += out[0]*out[1];
*dst = 0.2357 * t3;
dst += out[0]*out[1];
*dst = 0.2357 * t4;
}
}

mxFree(hist);
mxFree(norm);
return mxfeat;
}

// matlab entry point
// F = features(image, bin)
// image should be color with double values
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {
if (nrhs != 2)
mexErrMsgTxt("Wrong number of inputs");
if (nlhs != 1)
mexErrMsgTxt("Wrong number of outputs");
plhs[0] = process(prhs[0], prhs[1]);
}

 
dt.cc

#include <math.h>
#include <sys/types.h>
#include "mex.h"

#define int32_t int
/*
* Generalized distance transforms.
* We use a simple nlog(n) divide and conquer algorithm instead of the
* theoretically faster linear method, for no particular reason except
* that this is a bit simpler and I wanted to test it out.
*
* The code is a bit convoluted because dt1d can operate either along
* a row or column of an array.
*/

static inline int square(int x) { return x*x; }

// dt helper function
void dt_helper(double *src, double *dst, int *ptr, int step,
int s1, int s2, int d1, int d2, double a, double b) {
if (d2 >= d1) {
int d = (d1+d2) >> 1;
int s = s1;
for (int p = s1+1; p <= s2; p++)
if (src[s*step] + a*square(d-s) + b*(d-s) >
src[p*step] + a*square(d-p) + b*(d-p))
s = p;
dst[d*step] = src[s*step] + a*square(d-s) + b*(d-s);
ptr[d*step] = s;
dt_helper(src, dst, ptr, step, s1, s, d1, d-1, a, b);
dt_helper(src, dst, ptr, step, s, s2, d+1, d2, a, b);
}
}

// dt of 1d array
void dt1d(double *src, double *dst, int *ptr, int step, int n,
double a, double b) {
dt_helper(src, dst, ptr, step, 0, n-1, 0, n-1, a, b);
}

// matlab entry point
// [M, Ix, Iy] = dt(vals, ax, bx, ay, by)
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[]) {
if (nrhs != 5)
mexErrMsgTxt("Wrong number of inputs");
if (nlhs != 3)
mexErrMsgTxt("Wrong number of outputs");
if (mxGetClassID(prhs[0]) != mxDOUBLE_CLASS)
mexErrMsgTxt("Invalid input");

const int *dims = mxGetDimensions(prhs[0]);
double *vals = (double *)mxGetPr(prhs[0]);
double ax = mxGetScalar(prhs[1]);
double bx = mxGetScalar(prhs[2]);
double ay = mxGetScalar(prhs[3]);
double by = mxGetScalar(prhs[4]);

mxArray *mxM = mxCreateNumericArray(2, dims, mxDOUBLE_CLASS, mxREAL);
mxArray *mxIx = mxCreateNumericArray(2, dims, mxINT32_CLASS, mxREAL);
mxArray *mxIy = mxCreateNumericArray(2, dims, mxINT32_CLASS, mxREAL);
double *M = (double *)mxGetPr(mxM);
int32_t *Ix = (int32_t *)mxGetPr(mxIx);
int32_t *Iy = (int32_t *)mxGetPr(mxIy);

double *tmpM = (double *)mxCalloc(dims[0]*dims[1], sizeof(double)); // part map
int32_t *tmpIx = (int32_t *)mxCalloc(dims[0]*dims[1], sizeof(int32_t));
int32_t *tmpIy = (int32_t *)mxCalloc(dims[0]*dims[1], sizeof(int32_t));

for (int x = 0; x < dims[1]; x++)
dt1d(vals+x*dims[0], tmpM+x*dims[0], tmpIy+x*dims[0], 1, dims[0], ay, by);

for (int y = 0; y < dims[0]; y++)
dt1d(tmpM+y, M+y, tmpIx+y, dims[0], dims[1], ax, bx);

// get argmins and adjust for matlab indexing from 1
for (int x = 0; x < dims[1]; x++) {
for (int y = 0; y < dims[0]; y++) {
int p = x*dims[0]+y;
Ix[p] = tmpIx[p]+1;
Iy[p] = tmpIy[tmpIx[p]*dims[0]+y]+1;
}
}

mxFree(tmpM);
mxFree(tmpIx);
mxFree(tmpIy);
plhs[0] = mxM;
plhs[1] = mxIx;
plhs[2] = mxIy;
}