global optimization methods for computer...

Daniel CremersDepartment of Computer Science

University of Bonn

Global Optimization Methodsfor Computer Vision

Bonn Vision Workshop, October 8, 2009

2Global Optimization Methods for Computer VisionDaniel Cremers

Motion analysis3D reconstruction

Shape analysis Segmentation & tracking

Overview

3Global Optimization Methods for Computer VisionDaniel Cremers

Motion analysis3D reconstruction

Shape analysis Segmentation & tracking

Overview

4Global Optimization Methods for Computer VisionDaniel Cremers

Multiple View Reconstruction

5Global Optimization Methods for Computer VisionDaniel Cremers

Photoconsistency function:

Determine a surface of optimal photoconsistency by minimizing

Kolev, Klodt, Brox, Cremers, Int. J. of Computer Vision ’09:Theorem: Globally optimal surfaces can be computed via convex relaxation.

Multiple View Reconstruction

6Global Optimization Methods for Computer VisionDaniel Cremers

Non-convex energy

Non-convex versus Convex Energies

Convex energy

7Global Optimization Methods for Computer VisionDaniel Cremers

Middlebury Benchmark

8Global Optimization Methods for Computer VisionDaniel Cremers

Multiple View Reconstruction

Kolev, Cremers, ECCV ’08

9Global Optimization Methods for Computer VisionDaniel Cremers

Multiple View Reconstruction

10Global Optimization Methods for Computer VisionDaniel Cremers

Super-Resolution Texture Map

Goldlücke, Cremers, ICCV ’09, DAGM ’09* * Best Paper

Award

11Global Optimization Methods for Computer VisionDaniel Cremers

Super-Resolution Texture Map

Goldlücke, Cremers, ICCV ’09, DAGM ‘09

Weighted average Super-resolution texture

12Global Optimization Methods for Computer VisionDaniel Cremers

One of two images Depth reconstruction

Convex Multilabel Optimization

Pock, Schoenemann, Graber, Bischof, Cremers, ECCV ’08

13Global Optimization Methods for Computer VisionDaniel Cremers

Convex Multilabel Optimization

Pock, Chambolle, Bischof, Cremers, CVPR ’09

Input color image 10 label segmentation

14Global Optimization Methods for Computer VisionDaniel Cremers

Motion analysis3D reconstruction

Shape analysis Segmentation & tracking

Overview

15Global Optimization Methods for Computer VisionDaniel Cremers

Input video

High Accuracy Motion Estimation

Realtime optical flow

16Global Optimization Methods for Computer VisionDaniel Cremers

Input video

Realtime optical flow *

* 30 fps at 640 x 480 resolution

High Accuracy Motion Estimation

17Global Optimization Methods for Computer VisionDaniel Cremers

Quantitative Performance

Wedel et al. ICCV ’09

18Global Optimization Methods for Computer VisionDaniel Cremers

Wedel et al., ECCV ’08

Scene Flow: Motion & 3D StructureIn collaboration with Daimler Research

19Global Optimization Methods for Computer VisionDaniel Cremers

Motion analysis3D reconstruction

Shape analysis Segmentation & tracking

Overview

20Global Optimization Methods for Computer VisionDaniel Cremers

Klassen et al., PAMI ’03:Shape similarity as length of the shortest path (geodesic)

Shape Similarity via Geodesics

21Global Optimization Methods for Computer VisionDaniel Cremers

Schmidt, Clausen, Cremers DAGM ’06

Variational length minimizationShooting Methode

symmetric, stable, 1000 x faster

Shape Similarity via Geodesics

22Global Optimization Methods for Computer VisionDaniel Cremers

Schmidt, Farin, Cremers, ICCV ’07

Optimal correspondence betweenPoints of similar curvature

Optimal Matching & Correspondence

Matching as shortest cylic path

23Global Optimization Methods for Computer VisionDaniel Cremers

Fast Matching in Subcubic Runtime

24Global Optimization Methods for Computer VisionDaniel Cremers

Motion analysis3D reconstruction

Shape analysis Segmentation & tracking

Overview

25Global Optimization Methods for Computer VisionDaniel Cremers

Schoenemann, Cremers ICCV 2007, PAMI 2009 Theorem: Any matching of the template to the image corresponds to

a cyclic path in the product space of template and image.

template

?

Globally Optimal Tracking

26Global Optimization Methods for Computer VisionDaniel Cremers

25 fps3 fps

Schoenemann & Cremers CVPR ’08:Real-time by parallel implementation of polynomial algorithm.

Globally Optimal Tracking

27Global Optimization Methods for Computer VisionDaniel Cremers

Globally Optimal Tracking

28Global Optimization Methods for Computer VisionDaniel Cremers

Tracking deformable shapes

Multiple view reconstruction

Scene flow estimation

Summary

Super-resolution texturing