Download presentation

Presentation is loading. Please wait.

Published byGeorge Wimbrow Modified over 2 years ago

1
1

2
2 An extreme occurrence of the missing data W I D E B A S E L I N E – no point in more than 2 images!

3
3 Difficult cases Coinciding camera centers panorama zoom Dominant planes no problem

4
4 Uneven image capture 26 images 325 image pairs Some important, but very few matches

5
5 Uneven image capture

6
6 Our method can solve all previous examples.

7
7 Algorithm Technical contribution of this paper matches – uncalibrated EG [Matas et al. BMVC’02] focal length calibration [Stewenius et al. CVPR’05], [Nister PAMI’04] [Chum] EG importance consistent rotations linearly bundle adjustment with constrained rotations consistent translations using SOCP [Kahl ICCV’05] dense stereo [Kostkova & Sara BMVC’03]

8
8 Calibrated RANSAC and planes The six-point algorithm found only points on the wall. [Stewenius et al. CVPR’05] Two-View Geometry Unaffected by a Dominant Plane. [Chum et al. CVPR'05] use inliers as a pool for drawing samples in RANSAC on epipolar geometry

9
9 Full calibration The “five-point algorithm” on all pairs. [Nister PAMI’04] Partial calibration – unknown focal length The “six-point algorithm” on all pairs. [Stewenius et al. CVPR’05] mean focal length

10
10 Consistent rotations – previous work [Uyttendaele et al., CG\&A '04] – dense video self-intersecting paths vanishing points [Martinec, Pajdla CVPR'05] – gluing projective reconstructions metric upgrade needed! loosely coupled components – ambiguity!

11
11 Rotation registration into a reference frame rotation matrices rotations w.r.t. a reference frame relative rotation consistent rotations

12
12 Consistent rotations – solution fast: ~ 1 sec for 1000 image pairs close to orthonormal orthonormal and solve large & sparse matrix rewrite as eigenvalue problem global minimum well conditioned rotations – projection to orthonormal matrices

13
13 Refining rotations in each partial reconstruction:

14
14 Refining rotations in each partial reconstruction:

15
15 Refining rotations in each partial reconstruction: replace rotations by the consistent ones,

16
16 Refining rotations in each partial reconstruction: reprojection errors grow bundle adjustment needed change in relative rotation replace rotations by the consistent ones,

17
17 Refining rotations in each partial reconstruction:

18
18 Refining rotations in each partial reconstruction: refine all reconstructions together, each in independant coordinate frame, but with corresponding rotations constrained to be same re-estimate camera translations and points using [Kahl ICCV'05]

19
19 consistent rotations same rotations, translations unknown

20
20 0.8 / 18 pxl consistent rotations low errors stability

21
21 consistent rotations 0.8 / 18 pxl 0.20 / 1.6 pxl refine Refining rotations

22
22 Translations consistent rotations 0.8 / 18 pxl 0.20 / 1.6 pxl refine 0.24 / 1.3 pxl consistent translations [Kahl ICCV'05] 0.19 / 1.1 pxl refine

23
23 Final reconstruction

24
24 Experiments ICCV’05 Contest finals mean / maximum error 3.01 / 4.87 meters

25
25 Experiments ICCV’05 Contest finals St. Martin rotunda – 104 images

26
26 support Experiments ICCV’05 Contest finals Head2 St. Martin rotunda correct surfacesurface use triplets importance uneven image capture few data

27
27 Summary New algorithm for 3D reconstruction: EG importance consistent rotations linearly bundle adjustment with constrained rotations Acknowledgements: Ondrej Chum … code for EG unaffected by a dominant plane Fred Schaffalitzki … code for the six-point algorithm (publicly available) Lourakis et al. … base code for bundle adjustment (publicly available) Jana Kostkova … routines for dense stereo Richard Szeliski … the ICCV'05 Contest data (publicly available) Difficult scenarios: coinciding camera centers only two-view matches uneven image capture, wide base-line recent results on 260 views practical algorithm

Similar presentations

OK

Triangulation and Multi-View Geometry Class 9 Read notes Section 3.3, 4.3-4.4, 5.1 (if interested, read Triggs’s paper on MVG using tensor notation, see.

Triangulation and Multi-View Geometry Class 9 Read notes Section 3.3, 4.3-4.4, 5.1 (if interested, read Triggs’s paper on MVG using tensor notation, see.

© 2017 SlidePlayer.com Inc.

All rights reserved.

Ads by Google

Ppt on general etiquettes meaning Ppt on indian politics today Ppt on business etiquettes training wheels Ppt on tunnel diode detector Ppt on question tags exercises Ppt on credit card processing Ppt on principles of peace building institute of africa Ppt on shell structures analysis Numeric display ppt online Ppt on hot and cold places