Scribd
Upload
23 views

Closed-Form Optimal Two-View Triangulation Based On Angular Errors

This paper proposes closed-form solutions to optimal two-view triangulation based on minimizing angular reprojection errors. It derives exact solutions for L1, L2, and L∞ cost functions that correct feature point rays to intersect with minimum error. Experimental results show the proposed L1 median method yields lowest error in over 5.5 million tests for various criteria like reprojection error, triangulation speed, and different cost functions. The methods guarantee global optimality while being extremely simple and fast for two-view triangulation.

Uploaded by

Seong Hun Lee
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
23 views

Closed-Form Optimal Two-View Triangulation Based On Angular Errors

This paper proposes closed-form solutions to optimal two-view triangulation based on minimizing angular reprojection errors. It derives exact solutions for L1, L2, and L∞ cost functions that correct feature point rays to intersect with minimum error. Experimental results show the proposed L1 median method yields lowest error in over 5.5 million tests for various criteria like reprojection error, triangulation speed, and different cost functions. The methods guarantee global optimality while being extremely simple and fast for two-view triangulation.

Uploaded by

Seong Hun Lee
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

Closed-Form Optimal Two-View Triangulation Based on Angular Errors

Seong Hun Lee and Javier Civera (I3A, University of Zaragoza, Spain)
1. Two-View Triangulation 3. 𝐋𝟏 angle minimization 6. Qualitative results of the proposed 𝐋𝟏 method (median)
Locating the 3D point given its projections
in two views of known calibration and pose. If Rf0 × t ≤ f1 × t , then
′
Rf0 = Rf0 − Rf0 ∙ n1 n1 with n1 = f1 × t
′
f1 = f1
Else
′
Rf0 = Rf0 ,
′
f1 = f1 − f1 ∙ n0 n0 with n0 = Rf0 × t

4. 𝐋𝟐 sine of angle minimization


2. Optimal Method ′
Rf0 = Rf0 − Rf0 ∙ n n
Correct the rays (f0 and f1 ) to make them ′
f1 = f1 − f1 ∙n n 7. Percentage of the total experiments (>5,5M) for which each method yields the
intersect with a minimal image/angular lowest error in given criterion
reprojection cost, e.g., where n is the 2nd column of matrix V from
• 𝐋𝟏 𝐧𝐨𝐫𝐦: d0 + d1 [1] or θ0 + θ1 [ours] T T T
[1] [1] [1] [4] [2]
USV = SVD Rf0 f1 I−t t
2 2
• 𝐋𝟐 𝐧𝐨𝐫𝐦: d0 + d1 [1,4] or
2
sin (θ 0) +
2
sin (θ1 ) [3, ours] 5. 𝐋∞ angle minimization
• 𝐋∞ 𝐧𝐨𝐫𝐦: max(d0 , d1 ) [2] or ′
Rf0 = Rf0 − Rf0 ∙ n n
max(θ0 , θ1 ) [ours] ′
f1 = f1 − f1 ∙n n
where
8. Triangulation speed
na if na ≥ nb
n= n
b otherwise
[1] [1] [1] [2] [4] [4]
with
na = Rf0 + f1 × t
nb = Rf0 − f1 × t
9. Conclusions
• In this work, we derived the exact L1 , L2 and L∞ optimal solutions to two-view
[1] R. Hartley and P. Sturm. Triangulation. CVIU. 1997
[2] D. Níster. Automatic Dense Reconstruction from Uncalibrated Video Sequences. PhD thesis. 2001 triangulation based on angular reprojection errors.
[3] J. Oliensis. Exact Two-Image Structure from Motion. TPAMI. 2002 • Our methods are extremely simple and fast, and they guarantee global optimality under
[4] P. Lindstrom. Triangulation made easy. CVPR. 2010 respective cost functions.
[5] S. Lee and J. Civera. Triangulation: Why Optimize?. BMVC. 2019
ICCV 2019 Paper Nr: 1911 Poster Nr: 140 {seonghunlee, jcivera}@unizar.es

You might also like