Error message here!

Hide Error message here!

忘记密码?

Error message here!

请输入正确邮箱

Hide Error message here!

密码丢失?请输入您的电子邮件地址。您将收到一个重设密码链接。

Error message here!

返回登录

Close

Orbslam note 2 - feature matching method

Cc1924 2022-01-23 13:10:15 阅读数:7 评论数:0 点赞数:0 收藏数:0

2. Feature matching method

1.SearchForInitialization—— The matching used during initialization

  1. Initialization must be performed between two adjacent ordinary frames , Therefore, it can be considered that the pose difference between the two will not be too large
  2. stay F2( The first 2 frame ) Directly with F1( The first 1 frame ) The feature point coordinates in the are centered , In the radius of 100 Search inside the circle . This way can avoid violent matching , At the same time, the reason for matching in a circular area around is 1 Described in , Because the pose difference between the two frames is not big , So the matching feature points must be very close .
  3. Because it is initialization , Therefore, the requirements for matching are relatively high , So we used Match uniqueness and Rotational consistency Screening . problem : The proportion of matching uniqueness is 0.9 Relatively high , Logically, this is the initialization part, which should be more unique , That is, the proportion should be lower ?

3. I think bug

1.KeyFrame.cc Medium UpdateConnections() function

KeyFrame::UpdateConnections() in , The established common view relationship does not seem to be symmetrical .
See the following code for details : For The common view key of the current frame Establishment and The current frame When the connection between , A connection can only be established if it is greater than the threshold . And in the end The current frame Establish and The common view key of the current frame When the connection between , But directly assigned , Not removed Those common view keyframes that are less than the threshold , So this should be a problem .

Copyright statement
In this paper,the author:[Cc1924],Reprint please bring the original link, thank you