27
Scene Reconstruction Seminar presented by Anton Jigali Advanced Topics in Computer Vision (048921)

Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

Embed Size (px)

Citation preview

Page 1: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

Scene Reconstruction

Seminar presented by Anton Jigalin

Advanced Topics in Computer Vision (048921)

Page 2: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

2

Scene Reconstruction

Scene Visualization:

Photo Tourism: 3D photo browser

Path Finder: 3D navigation interface

References: Scene Reconstruction and Visualization from Internet Photo Collections. N. Snavely. 2008.

Page 3: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

3

Scene Reconstruction

Page 4: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

4

The goal

Automatically reconstruct 3D geometry from large collections of photos from the Internet

Page 5: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

Motivation

5

New way to organize photo collections Experience of being at a place

Page 6: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

6

Related works

Image based modeling

Structure from motion (SfM)

Moviemaps

Page 7: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

7

Challenges

Unknown cameras position

Unknown internal camera parameters

Correspondence problem

Visualization

Algorithmic complexity

Running the complete pipeline on a 3.80GHz Intel Xeon machine with 4GB of core memory. The keypoint detection and matching phases were run in parallel on 10 such machines.

Page 8: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

8

Results

Page 9: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

9

How was this result achieved?

Page 10: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

10

Finding correspondence

Feature detection - SIFT

Feature matching - RANSAC

Epipolar constraint

Page 11: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

11

Structure from Motion Algorithm

Page 12: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

12

Initial two-frame reconstruction

Intermediate stage Final reconstruction

Structure from Motion Algorithm

Page 13: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

13

Failure modes

Repeating textures

Insufficient overlap or texture

Cascading errors

Bad initialization

Page 14: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

14

Connectivity graphs

Page 15: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

15

Scene Visualization:

Photo Tourism: 3D photo browser

Page 16: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

16

Photo Tourism: Main components

Rendering engine

Navigation interface

Tools for scene annotation

Page 17: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

17

Transitions between photographs

Page 18: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

18

Navigation interface

Object-based navigation

Stabilized slideshow

Page 19: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

19

Scene annotation

Annotate regions

Register new photographs

Page 20: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

20

Scene Visualization:

Path Finder: 3D navigation interface

Page 21: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

21

Best possible controls

The goals of 3D navigation:

• Exploration• Search• Object inspection

Principles for 3D scene navigation:

• Controls are scene and task dependent

• Constrained and automatic navigation

• User’s position and motion should be clear

Page 22: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

22

Pathfinder: Main components

Viewpoint scoring

Navigation controls for a scene

Rendering engine

User interface

Page 23: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

25

Scene-specific controls: Panoramas and Orbits

Panoramas

Orbits

Page 24: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

26

Path planning

High-quality rendering

Shortest path

Piecewise linear path

Path smoothing

Connectivity graph

Page 25: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

29

Limitations

Non circular orbits

User needs

Appearance compensation

Larger scenes

(a) Pisa Doumo

(b) Stonehenge

Page 26: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

30

Conclusion

Structure from motion

Photo Tourism - 3D photo browser

Path Finder - 3D navigation system

Reconstructing Rome

Capturing appearance

New photo-sharing community

Future work

Page 27: Scene Reconstruction Seminar presented by Anton Jigalin Advanced Topics in Computer Vision ( 048921 )

31

Thank you