摘要:本文主要描述ORB算法原理以及opencv中ORB算法的实现。
关键字:ORB,FAST,BRIEF
ORB算法ICCV论文:ORB:an efficient alternative to sift or surf。
1 ORB算法简介
ORB(Oriented Fast and Rotated BRIEF)是一种快速局部特征点提取和快速计算局部特征点描述子算法。该算法分为两个部分:特征点提取和特征点描述子提取。其中特征掉提取是根据FAST改进而来的oFAST;而特征点描述子提取是根据BRIEF(Binary Robust IndependentElementary Feature)特征描述算法改进而来。
O R B = O r i e n t e d F a s t ( k e y p o i n t ) + R o t a t e d B R I E F ( d e s c r i p t o r ) ORB=Oriented\quad Fast(keypoint)+Rotated\quad BRIEF(descriptor)ORB =O r i e n t e d F a s t (k ey p o in t )+R o t a t e d BR I EF (d escr i pt or )
2 算法原理
2.1 oFAST: FAST Keypoint Orientation
FAST
FAST算法计算性能比较好但是不具备尺度不变性,旋转不变性以及检测到的特征点没有旋转方向。oFAST正是针对这些缺点改进FAST而来。
FAST:Features from Accelerated Segment Test
FAST在进行特征点提取是根据当前点邻域内的点的差值作为特征点的筛选标准,它假定如果一个点与周围邻域内足够多的点的差值够大则认为是一个特征点:
- 选择像素p p p,该像素的像素值为I p I_{p}I p ,确定一个筛选的阈值T T T(测试集参考值20%);
- 计算以像素p p p为圆心3为半径确定16个像素点的灰度值和圆心p p p的灰度值I p I_{p}I p 的差值,如果存在连续n n n个点(算法的第一个版本的n取值为12)满足I x − I p > ∣ t ∣ I_x-I_p>|t|I x −I p >∣t ∣(I x I_x I x 表示以p p p为圆心的点的灰度值,t t t为根据T T T计算出的偏移量),则认为点p p p可以作为一个候选点,否则剔除;
- 对于16个点都计算差值时间复杂度过高,因此FAST采用即特征点过滤的方式:先判断圆上1,5,9,13号4个点中如果至少3个点满足特征点初选的要求在进行逐个计算,否则终止计算;
- 遍历图像中每一个像素,重复上述操作,直到遍历结束;
上面的筛选方式是FAST初版的筛选方式,但是该方式有一些基本的缺陷:
- 邻域内点数量n n n过小,容易导致筛选的点过多,降低计算效率;
- 筛选的点完全取决于图像中点的分布情况;
- 筛选出的点部分点集中在一起。
问题1,2可以通过机器学习的方式来解决,问题3通过极大值抑制解决。
机器学习的解决方式是针对应用场景训练一个决策树,利用该决策树进行筛选:
- 确定训练集;
- 对于训练集中每一张图像运行FAST算法筛选出图像中的特征点构成集合P P P;
- 对于筛选出的每一个特征点,将其邻域内的16个像素存储为一个一维向量;
- 对于每一个一维的邻域向量中的像素值I p → x , x ∈ 1 , . . . , 16 I_{p\rightarrow x},x\in{1,…,16}I p →x ,x ∈1 ,…,16将其通过下面的规则映射到3中状态 darker than I p I_p I p , brighter then I p I_p I p 或者和I p I_p I p 相似;
S p → x { d , I p → x ≤ I p − t ( d a r k e r ) s , I p − t < I p → x < I p + t ( s i m i l a r ) b , I p + t ≤ I p → x ( b r i g h t e r ) \begin{equation} S_{p\rightarrow x}\left{ \begin{array}{ll} d,I_{p\rightarrow x}&\le I_p – t\quad(darker)\ s,I_p-t&\lt I_{p\rightarrow x}\lt I_p+t\quad(similar)\ b,I_p+t&\le I_{p\rightarrow x}\quad (brighter) \end{array}\right. \end{equation}S p →x ⎩⎨⎧d ,I p →x s ,I p −t b ,I p +t ≤I p −t (d a r k er )<I p →x <I p +t (s imi l a r )≤I p →x (b r i g h t er ) - 对于给定的x x x可以将集合P P P分为三类P d , P s , P b P_d,P_s,P_b P d ,P s ,P b ,即P b = { p ∈ P : S p → x = b } P_b={p\in P:S_{p\rightarrow x=b}}P b ={p ∈P :S p →x =b };
- 定义一个布尔变量K p K_p K p ,表示如果p p p为角点则真,否则为假(由于有训练集所以GT我们都是已知的);
- 使用ID3决策树分类器以K p K_p K p 查询每个子集,以训练出正确特征点的分类器;
- 决策树使用熵最小化来逼近,类似交叉熵。
H ( P ) = ( c + c ‾ ) l o g 2 ( c + c ‾ ) − c l o g 2 c − c ‾ l o g 2 c ‾ c = ∣ { p ∣ K p i s t r u e } ∣ (角点数量) c ‾ = ∣ { p ∣ K p i s f a l s e } ∣ (非角点的数量) \begin{equation} \begin{aligned} H(P)&=(c+\overline{c})log_2(c+\overline{c})-clog_2c-\overline{c}log_2\overline{c}\ c&=|{p|K_p\quad is\quad true}|(角点数量)\ \overline{c}&=|{p|K_p\quad is\quad false}|(非角点的数量) \end{aligned} \end{equation}H (P )c c =(c +c )l o g 2 (c +c )−c l o g 2 c −c l o g 2 c =∣{p ∣K p i s t r u e }∣(角点数量)=∣{p ∣K p i s f a l se }∣(非角点的数量) - 递归应用熵最小化来处理所有的集合,直到熵值为0终止计算,得到的决策树就可以用来进行特征点筛选。
问题3可以用非极大值抑制来解决:
- 针对每个点计算打分函数V V V,打分函数的输出是通过周围16个像素的差分和
- 去除v v v值较低的点即可。
V = m a x { ∑ ( I s − I p ) , i f ( I s − I p ) > t ∑ ( I p − I s ) , i f ( I p − I s ) > t \begin{equation} V=max\left{ \begin{array}{ll} \sum(I_s-I_p),if\quad(I_s-I_p)\gt t\ \sum(I_p-I_s),if\quad(I_p-I_s)\gt t \end{array}\right. \end{equation}V =ma x {∑(I s −I p ),i f (I s −I p )>t ∑(I p −I s ),i f (I p −I s )>t
oFAST
oFAST采用的是FAST-9,邻域半径为9。
oFAST使用多尺度金字塔解决FAST不具备尺度不变性的问题,使用矩来确定特征点的方向,来解决其不具备旋转不变形的问题。
多尺度金字塔就是将给定的图像以一个给定的尺度进行缩放,来生成多层不同尺度的图像I 1 , I 2 , . . . , I n I_1,I_2,…,I_n I 1 ,I 2 ,…,I n 。这些图像每张图像的尺寸不同但是相邻两层图像间的宽高比例相同。然后针对多个尺度的图像应用FAST算法筛选特征点。
oFAST使用强度质心(intensity centroid,强度质心假设角的强度偏离其中心,并且该向量可用于估算方向)来确定特征点的方向,即计算特征点以r r r为半径范围内的质心,特征点坐标到质心形成一个向量作为特征点的方向。矩的定义
m p q = ∑ x , y ∈ r x p y q I ( x , y ) m_{pq}=\sum_{x,y\in r}x^p y^q I(x,y)m pq =x ,y ∈r ∑x p y q I (x ,y )
矩的质心为(这里计算是限定的r r r的邻域内):
C = ( m 10 m 00 , m 01 m 00 ) C=(\frac{m_{10}}{m_{00}},\frac{m_{01}}{m_{00}})C =(m 00 m 10 ,m 00 m 01 )
特征点的方向为:
θ = a t a n 2 ( m 01 , m 10 ) \theta=atan2(m_{01},m_{10})θ=a t an 2 (m 01 ,m 10 )
oFast会使用Harris角点检测对筛选出的特征点进行角点度量,然后根据度量值将特征点排序,最后取出top-N得到目标特征点。
; 2.2 rBRIEF
BRIEF
BRIEF是一种对已经检测到的特征点进行描述的算法,该算法生成一种二进制描述子来描述已知的特征点。这些描述子可以用来进行特征点匹配等操作。BRIEF摒弃了利用区域直方图描述特征点的传统方法,采用二进制、位异或运算处理,大大的加快了特征点描述符建立的速度,同时也极大的降低了特征点匹配的时间,是一种非常快速的特征点描述子算法。
BRIEF的目标是得到一个二进制串,该串描述了特征点的特性。描述子生成的方式:
- 为了减少噪声干扰,首先对图像进行高斯滤波(方差为2,高斯窗口为9 × 9 9\times 9 9 ×9);
- 然后以特征点为中心,取S × S S\times S S ×S大小的邻域窗口,在窗口内以一定方式选取一对点p , q p,q p ,q,比较两个像素点的大小:
a. 如果I p > I q I_p>I_q I p >I q ,则当前位二进制值为1;
b. 否则为0; - 不断循环,直到生成目标长度的二进制串(ORB采用256长度的二进制串)。
上面第二步,选取点的采样方式可以为:
- p , q p,q p ,q都符合[ − S 2 , S 2 ] [-\frac{S}{2},\frac{S}{2}][−2 S ,2 S ]的均匀采样;
- p , q p,q p ,q都符合各向同性的高斯分布[ 0 , 1 25 S 2 ] [0,\frac{1}{25}S^2][0 ,25 1 S 2 ]采样;
- p p p符合高斯分布[ 0 , 1 25 S 2 ] [0,\frac{1}{25}S^2][0 ,25 1 S 2 ]采样,q q q符合[ 0 , 1 100 S 2 ] [0,\frac{1}{100}S^2][0 ,100 1 S 2 ]采样,采样方式是首先在原点处为p p p采样,然后以p p p为中心为q q q采样;
- p , q p,q p ,q在空间量化极坐标下的离散位置处进行随机采样;
- p = ( 0 , 0 ) T , q p=(0,0)^T,q p =(0 ,0 )T ,q在空间量化极坐标下的离散位置处进行随机采样;
BRIEF存在不具备旋转不变性,不具备尺度不变性和对噪声敏感的问题。
steered BRIEF
rBRIEF通过将筛选的特征点旋转一定的角度再计算对应的描述子来得到旋转不变性。比如点集S S S:
S = ( x 1 , . . . , x n y 1 , . . . , y n ) \begin{equation} S=\left( \begin{array}{ll} x_1,…,x_n\ y_1,…,y_n \end{array}\right) \end{equation}S =(x 1 ,…,x n y 1 ,…,y n )
通过旋转矩阵R θ R_{\theta}R θ(这一角度并不是固定的而是预先以2π/30为增量构建的模式查找表获取的)旋转得到S θ S_{\theta}S θ然后再点集S θ S_{\theta}S θ上计算描述子:
S θ = R θ S S_{\theta}=R_{\theta}S S θ=R θS
上述方式虽然能够得到旋转不变性但是这种方式计算得到的描述子在不同特征点之间的区分度不是很大。因此为了解决该问题,进一步采用统计学习的方式来重新选择点的集合。
- 构建300K个特征点训练集(ORB采用从PASCAL 2006 构建训练集);
- 在300K个特征点的每个点的31 × 31 31\times 31 31 ×31邻域内按照M M M种方法取点对,比较点对的大小形成300 K × M 300K\times M 300 K ×M的二进制矩阵Q Q Q。矩阵的每一列表示对应点的二进制描述;
- 按距离平均值 0.5 对测试进行排序,形成向量T T T。
- 贪心搜索
a. 将第一个测试放入结果向量 R 并将其从 T 中删除;
b. 从 T 中获取下一个测试,并将其与 R 中的所有测试进行比较。如果其绝对相关性大于阈值,则丢弃它; 否则将其添加到 R;
c. 重复上一步,直到R R R中有 256 个测试。如果少于 256,则提高阈值并重试。
; 3 ORB opencv实现
Ptr<ORB> ORB::create(int nfeatures, float scaleFactor, int nlevels, int edgeThreshold, int firstLevel, int wta_k, int scoreType, int patchSize, int fastThreshold){
CV_Assert(firstLevel >= 0);
return makePtr<ORB_Impl>(nfeatures, scaleFactor, nlevels, edgeThreshold, firstLevel, wta_k, scoreType, patchSize, fastThreshold);
}
ORB的具体实现直接看 detectAndCompute
就可以, FeatureExtractor
也提供了 detect
和 compute
的接口,但是区别不大:
void ORB_Impl::detectAndCompute( InputArray _image, InputArray _mask, std::vector<KeyPoint>& keypoints, OutputArray _descriptors, bool useProvidedKeypoints )
下面的代码在 opencv中的modules/feature2d/orb.cpp
中,并且删除了部分不影响实际流程的宏处理比如OpenCL,以及部分Assert等相关宏。
void ORB_Impl::detectAndCompute( InputArray _image, InputArray _mask, std::vector<KeyPoint>& keypoints,
OutputArray _descriptors, bool useProvidedKeypoints ) {
CV_Assert(patchSize >= 2);
bool do_keypoints = !useProvidedKeypoints;
bool do_descriptors = _descriptors.needed();
if( (!do_keypoints && !do_descriptors) || _image.empty() )
return;
const int HARRIS_BLOCK_SIZE = 9;
int halfPatchSize = patchSize / 2;
int descPatchSize = cvCeil(halfPatchSize*sqrt(2.0));
int border = std::max(edgeThreshold, std::max(descPatchSize, HARRIS_BLOCK_SIZE/2))+1;
Mat image = _image.getMat(), mask = _mask.getMat();
if( image.type() != CV_8UC1 )
cvtColor(_image, image, COLOR_BGR2GRAY);
int i, level, nLevels = this->nlevels, nkeypoints = (int)keypoints.size();
bool sortedByLevel = true;
std::vector<Rect> layerInfo(nLevels);
std::vector<int> layerOfs(nLevels);
std::vector<float> layerScale(nLevels);
Mat imagePyramid, maskPyramid;
float level0_inv_scale = 1.0f / getScale(0, firstLevel, scaleFactor);
size_t level0_width = (size_t)cvRound(image.cols * level0_inv_scale);
size_t level0_height = (size_t)cvRound(image.rows * level0_inv_scale);
Size bufSize((int)alignSize(level0_width + border*2, 16), 0);
int level_dy = (int)level0_height + border*2;
Point level_ofs(0, 0);
for( level = 0; level < nLevels; level++ )
{
float scale = getScale(level, firstLevel, scaleFactor);
layerScale[level] = scale;
float inv_scale = 1.0f / scale;
Size sz(cvRound(image.cols * inv_scale), cvRound(image.rows * inv_scale));
Size wholeSize(sz.width + border*2, sz.height + border*2);
if( level_ofs.x + wholeSize.width > bufSize.width )
{
level_ofs = Point(0, level_ofs.y + level_dy);
level_dy = wholeSize.height;
}
Rect linfo(level_ofs.x + border, level_ofs.y + border, sz.width, sz.height);
layerInfo[level] = linfo;
layerOfs[level] = linfo.y*bufSize.width + linfo.x;
level_ofs.x += wholeSize.width;
}
bufSize.height = level_ofs.y + level_dy;
imagePyramid.create(bufSize, CV_8U);
if( !mask.empty() )
maskPyramid.create(bufSize, CV_8U);
Mat prevImg = image, prevMask = mask;
for (level = 0; level < nLevels; ++level)
{
Rect linfo = layerInfo[level];
Size sz(linfo.width, linfo.height);
Size wholeSize(sz.width + border*2, sz.height + border*2);
Rect wholeLinfo = Rect(linfo.x - border, linfo.y - border, wholeSize.width, wholeSize.height);
Mat extImg = imagePyramid(wholeLinfo), extMask;
Mat currImg = extImg(Rect(border, border, sz.width, sz.height)), currMask;
if( !mask.empty() )
{
extMask = maskPyramid(wholeLinfo);
currMask = extMask(Rect(border, border, sz.width, sz.height));
}
if( level != firstLevel )
{
resize(prevImg, currImg, sz, 0, 0, INTER_LINEAR_EXACT);
if( !mask.empty() )
{
resize(prevMask, currMask, sz, 0, 0, INTER_LINEAR_EXACT);
if( level > firstLevel )
threshold(currMask, currMask, 254, 0, THRESH_TOZERO);
}
copyMakeBorder(currImg, extImg, border, border, border, border,
BORDER_REFLECT_101+BORDER_ISOLATED);
if (!mask.empty())
copyMakeBorder(currMask, extMask, border, border, border, border,
BORDER_CONSTANT+BORDER_ISOLATED);
}
else
{
copyMakeBorder(image, extImg, border, border, border, border,
BORDER_REFLECT_101);
if( !mask.empty() )
copyMakeBorder(mask, extMask, border, border, border, border,
BORDER_CONSTANT+BORDER_ISOLATED);
}
if (level > firstLevel)
{
prevImg = currImg;
prevMask = currMask;
}
}
computeKeyPoints(imagePyramid, uimagePyramid, maskPyramid,
layerInfo, ulayerInfo, layerScale, keypoints,
nfeatures, scaleFactor, edgeThreshold, patchSize, scoreType, useOCL, fastThreshold);
if( do_descriptors )
{
int dsize = descriptorSize();
nkeypoints = (int)keypoints.size();
if( nkeypoints == 0 )
{
_descriptors.release();
return;
}
_descriptors.create(nkeypoints, dsize, CV_8U);
std::vector<Point> pattern;
const int npoints = 512;
Point patternbuf[npoints];
const Point* pattern0 = (const Point*)bit_pattern_31_;
if( patchSize != 31 )
{
pattern0 = patternbuf;
makeRandomPattern(patchSize, patternbuf, npoints);
}
CV_Assert( wta_k == 2 || wta_k == 3 || wta_k == 4 );
if( wta_k == 2 )
std::copy(pattern0, pattern0 + npoints, std::back_inserter(pattern));
else
{
int ntuples = descriptorSize()*4;
initializeOrbPattern(pattern0, pattern, ntuples, wta_k, npoints);
}
for( level = 0; level < nLevels; level++ )
{
Mat workingMat = imagePyramid(layerInfo[level]);
GaussianBlur(workingMat, workingMat, Size(7, 7), 2, 2, BORDER_REFLECT_101);
}
{
Mat descriptors = _descriptors.getMat();
computeOrbDescriptors(imagePyramid, layerInfo, layerScale,
keypoints, descriptors, pattern, dsize, wta_k);
}
}
}
生成的金字塔是一整张大图,所有的图像会放到一张图像中如果几张小图能够放在同一行则是有限并排存放而不是总是按照顺序存放,如下:
static void computeKeyPoints(const Mat& imagePyramid, const UMat& uimagePyramid, const Mat& maskPyramid, const std::vector<Rect>& layerInfo, const UMat& ulayerInfo, const std::vector<float>& layerScale, std::vector<KeyPoint>& allKeypoints, int nfeatures, double scaleFactor, int edgeThreshold, int patchSize, int scoreType, bool useOCL, int fastThreshold ) {
int i, nkeypoints, level, nlevels = (int)layerInfo.size();
std::vector<int> nfeaturesPerLevel(nlevels);
float factor = (float)(1.0 / scaleFactor);
float ndesiredFeaturesPerScale = nfeatures*(1 - factor)/(1 - (float)std::pow((double)factor, (double)nlevels));
int sumFeatures = 0;
for( level = 0; level < nlevels-1; level++ ) {
nfeaturesPerLevel[level] = cvRound(ndesiredFeaturesPerScale);
sumFeatures += nfeaturesPerLevel[level];
ndesiredFeaturesPerScale *= factor;
}
nfeaturesPerLevel[nlevels-1] = std::max(nfeatures - sumFeatures, 0);
int halfPatchSize = patchSize / 2;
std::vector<int> umax(halfPatchSize + 2);
int v, v0, vmax = cvFloor(halfPatchSize * std::sqrt(2.f) / 2 + 1);
int vmin = cvCeil(halfPatchSize * std::sqrt(2.f) / 2);
for (v = 0; v vmax; ++v)
umax[v] = cvRound(std::sqrt((double)halfPatchSize * halfPatchSize - v * v));
for (v = halfPatchSize, v0 = 0; v >= vmin; --v){
while (umax[v0] == umax[v0 + 1])
++v0;
umax[v] = v0;
++v0;
}
allKeypoints.clear();
std::vector<KeyPoint> keypoints;
std::vector<int> counters(nlevels);
keypoints.reserve(nfeaturesPerLevel[0]*2);
for( level = 0; level < nlevels; level++ )
{
int featuresNum = nfeaturesPerLevel[level];
Mat img = imagePyramid(layerInfo[level]);
Mat mask = maskPyramid.empty() ? Mat() : maskPyramid(layerInfo[level]);
{使用fast进行角点检测
Ptr<FastFeatureDetector> fd = FastFeatureDetector::create(fastThreshold, true);
fd->detect(img, keypoints, mask);
}
KeyPointsFilter::runByImageBorder(keypoints, img.size(), edgeThreshold);
KeyPointsFilter::retainBest(keypoints, scoreType == ORB_Impl::HARRIS_SCORE ? 2 * featuresNum : featuresNum);
nkeypoints = (int)keypoints.size();
counters[level] = nkeypoints;
float sf = layerScale[level];
for( i = 0; i < nkeypoints; i++ ){
keypoints[i].octave = level;
keypoints[i].size = patchSize*sf;
}
std::copy(keypoints.begin(), keypoints.end(), std::back_inserter(allKeypoints));
}
std::vector<Vec3i> ukeypoints_buf;
nkeypoints = (int)allKeypoints.size();
if(nkeypoints == 0){
return;
}
Mat responses;
if( scoreType == ORB_Impl::HARRIS_SCORE ) {
HarrisResponses(imagePyramid, layerInfo, allKeypoints, 7, HARRIS_K);
std::vector<KeyPoint> newAllKeypoints;
newAllKeypoints.reserve(nfeaturesPerLevel[0]*nlevels);
int offset = 0;
for( level = 0; level < nlevels; level++ ) {
int featuresNum = nfeaturesPerLevel[level];
nkeypoints = counters[level];
keypoints.resize(nkeypoints);
std::copy(allKeypoints.begin() + offset,
allKeypoints.begin() + offset + nkeypoints,
keypoints.begin());
offset += nkeypoints;
KeyPointsFilter::retainBest(keypoints, featuresNum);
std::copy(keypoints.begin(), keypoints.end(), std::back_inserter(newAllKeypoints));
}
std::swap(allKeypoints, newAllKeypoints);
}
nkeypoints = (int)allKeypoints.size();
{
ICAngles(imagePyramid, layerInfo, allKeypoints, umax, halfPatchSize);
}
for( i = 0; i < nkeypoints; i++ ) {
float scale = layerScale[allKeypoints[i].octave];
allKeypoints[i].pt *= scale;
}
}
下面两张图分别是用Harris过滤前的结果,第二张图的角点比第一张图稍微少一些但是不明显:
下面是经过计算的点的角度,图中线相对于x轴正方向的夹角就是角点的角度:
计算特征点的描述子:
static void
computeOrbDescriptors( const Mat& imagePyramid, const std::vector<Rect>& layerInfo,
const std::vector<float>& layerScale, std::vector<KeyPoint>& keypoints,
Mat& descriptors, const std::vector<Point>& _pattern, int dsize, int wta_k )
{
int step = (int)imagePyramid.step;
int j, i, nkeypoints = (int)keypoints.size();
for( j = 0; j < nkeypoints; j++ ){
const KeyPoint& kpt = keypoints[j];
const Rect& layer = layerInfo[kpt.octave];
float scale = 1.f/layerScale[kpt.octave];
float angle = kpt.angle;
angle *= (float)(CV_PI/180.f);
float a = (float)cos(angle), b = (float)sin(angle);
const uchar* center = &imagePyramid.at<uchar>(cvRound(kpt.pt.y*scale) + layer.y, cvRound(kpt.pt.x*scale) + layer.x);
float x, y;
int ix, iy;
const Point* pattern = &_pattern[0];
uchar* desc = descriptors.ptr<uchar>(j);
#if 1
#define GET_VALUE(idx) \
(x = pattern[idx].x*a - pattern[idx].y*b, \
y = pattern[idx].x*b + pattern[idx].y*a, \
ix = cvRound(x), \
iy = cvRound(y), \
*(center + iy*step + ix) )
#else
#endif
if( wta_k == 2 ){
for (i = 0; i < dsize; ++i, pattern += 16){
int t0, t1, val;
t0 = GET_VALUE(0); t1 = GET_VALUE(1);
val = t0 < t1;
t0 = GET_VALUE(2); t1 = GET_VALUE(3);
val |= (t0 < t1) << 1;
t0 = GET_VALUE(4); t1 = GET_VALUE(5);
val |= (t0 < t1) << 2;
t0 = GET_VALUE(6); t1 = GET_VALUE(7);
val |= (t0 < t1) << 3;
t0 = GET_VALUE(8); t1 = GET_VALUE(9);
val |= (t0 < t1) << 4;
t0 = GET_VALUE(10); t1 = GET_VALUE(11);
val |= (t0 < t1) << 5;
t0 = GET_VALUE(12); t1 = GET_VALUE(13);
val |= (t0 < t1) << 6;
t0 = GET_VALUE(14); t1 = GET_VALUE(15);
val |= (t0 < t1) << 7;
desc[i] = (uchar)val;
}
}
else if( wta_k == 3 ){
for (i = 0; i < dsize; ++i, pattern += 12){
int t0, t1, t2, val;
t0 = GET_VALUE(0); t1 = GET_VALUE(1); t2 = GET_VALUE(2);
val = t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0);
t0 = GET_VALUE(3); t1 = GET_VALUE(4); t2 = GET_VALUE(5);
val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 2;
t0 = GET_VALUE(6); t1 = GET_VALUE(7); t2 = GET_VALUE(8);
val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 4;
t0 = GET_VALUE(9); t1 = GET_VALUE(10); t2 = GET_VALUE(11);
val |= (t2 > t1 ? (t2 > t0 ? 2 : 0) : (t1 > t0)) << 6;
desc[i] = (uchar)val;
}
}
else if( wta_k == 4 ){
for (i = 0; i < dsize; ++i, pattern += 16){
int t0, t1, t2, t3, u, v, k, val;
t0 = GET_VALUE(0); t1 = GET_VALUE(1);
t2 = GET_VALUE(2); t3 = GET_VALUE(3);
u = 0, v = 2;
if( t1 > t0 ) t0 = t1, u = 1;
if( t3 > t2 ) t2 = t3, v = 3;
k = t0 > t2 ? u : v;
val = k;
t0 = GET_VALUE(4); t1 = GET_VALUE(5);
t2 = GET_VALUE(6); t3 = GET_VALUE(7);
u = 0, v = 2;
if( t1 > t0 ) t0 = t1, u = 1;
if( t3 > t2 ) t2 = t3, v = 3;
k = t0 > t2 ? u : v;
val |= k << 2;
t0 = GET_VALUE(8); t1 = GET_VALUE(9);
t2 = GET_VALUE(10); t3 = GET_VALUE(11);
u = 0, v = 2;
if( t1 > t0 ) t0 = t1, u = 1;
if( t3 > t2 ) t2 = t3, v = 3;
k = t0 > t2 ? u : v;
val |= k << 4;
t0 = GET_VALUE(12); t1 = GET_VALUE(13);
t2 = GET_VALUE(14); t3 = GET_VALUE(15);
u = 0, v = 2;
if( t1 > t0 ) t0 = t1, u = 1;
if( t3 > t2 ) t2 = t3, v = 3;
k = t0 > t2 ? u : v;
val |= k << 6;
desc[i] = (uchar)val;
}
}
else
CV_Error( Error::StsBadSize, "Wrong wta_k. It can be only 2, 3 or 4." );
#undef GET_VALUE
}
}
4 参考文献
- ORB:an efficient alternative to sift or surf
- Features from Accelerated Segment Test (FAST)
- Machine learning for high-speed corner detection
- ID3 Descision Tree
- BRIEF: Binary Robust Independent Elementary Features⋆
Original: https://blog.csdn.net/GrayOnDream/article/details/127378847
Author: 落樱弥城
Title: ORB算法与opencv实现
原创文章受到原创版权保护。转载请注明出处:https://www.johngo689.com/637093/
转载文章受原作者版权保护。转载请注明原作者出处!