0、方法参考

RANSAC

1、ORSA (Optimized RANSAC)

1. 核心： balance：sigma & NFA
   sigma: a threshold on the residual error
   NFA: the number of false alarms
2. 公式及参数

• 

• 

• 

• For each possible k, we compute the NFA as above, and keep only the least of them, provided it is below some threshold, usually 1.

2、ORSA (Optimized RANSAC) + Homography

1. 算法流程图
   
2. 上述算法细节

• sampling:
  Initially, it is the identity, but whenever the best meaningful model so far is found(in 2.3), this array is set to inliers, so that next samples will be drawn among inliers.

• nIterReserve的设置
  The idea is that even though no meaningful model has been found, the best found model has a higher probability of containing true inliers. This can help in difficult cases of high outlier ratio.

• precision
  Remember that the actual precision is ultimately automatically chosen by ORSA, below the user's input precision.

• NFA threshold
  
  a standard value of 1 is common in a contrario methods.

3. 公式及参数

• 几何意义： ratio of areas of a radius 1 disk and of the right image.
• 在homography计算中，NFA的具体公式如上

3、其他

3.1、Duplicates Removal

remove duplicate correspondences (identical endpoints) before launching ORSA.

原因：
SIFT – keypoint matches :
the codes of these keypoints can match for these different orientations???疑惑

（原文参考） If a keypoint in one image has several orientations and another
keypoint in the other image also has several orientations, the codes
of these keypoints can match for these different orientations, and
several correspondences with identical endpoints are output. However,
we do not discard correspondences sharing only one endpoint

3.2、Failure ：very few matches.