ORB: an effiient alternative to SIFT or SURF
ORB
1.Oriented FAST
(Features from Accelerated Segment Test)
keypoint detector
2.Rotated BRIEF
(Binary Robust Independent Elementary Features)
binary descriptor
• 1. Oriented FAST
• 1.1 FAST Detector
FAST-9+Harris cornr measure+top N points
阙值t>=<灰度值之差
像素点周围的16个像素
有连续12个像素和中心
点不同,那么它就是一
个角点
• 1.2 Orientation by Intensity Centroid
Centroid is
• 2.Rotated BRIEF
• 2.1 BRIEF
1.以关键点P为圆心,以d为半径做圆O;
2.在圆O内选取Gauss分布的N个点对,
如图N=4;
3. 定义T
4.将得到的结果进行组合,如最终的描
述子为:1011
• 2.2 Steered BRIEF
• not invariant to rotation
• according to the orientation of keypoints
• 例如特征点A、B的描述子如下
A:10101011
B:10101010
AB异或操作,计算出A和B的相似度。
BRIEF has a large variance and a mean near 0.5;
Steered BRIEF present a more uniform appearance and low variance
• 2.3 rBRIEF
• high variance makes a feature more discriminative
• to recover from the loss of variance in Steered BRIEF
and reduce correlation
• learning good binary features->greedy search
•
patch 31*31
test 5*5
N = (31-5)^2 possible sub-windows
select pairs of two from these
binary tests
eliminate tests that overlap
M=205590 possible tests
• 设一张图有4个特征点
• 每个特征点有五个待测描述子,希望得到四个描述
子(类比,有M个待测描述子,希望得到256个)
筛
选
掉
1
1
0
0
0.5
0
1
0
1
0.5
1
0
1
0
0.5
平均值
1
1
0
0
1
0
1
1
0.75 0.5
10111
11010
00101
01001
1
0
1
1
• 把110放进R,T中去除110后排第一的010与R中所有
列向量求correlation,超过阈值,舍弃,否则将010
放进R,依次循环,直到R中有四个列向量