Alessio Del Bue Homepage

2010 Tutorial

Category: Uncategorized / Tags: no tag / Add Comment

Lourdes Agapito

Adrien Bartoli

Alessio Del Bue

2nd Tutorial on

Computer Vision in a Non-rigid World

6-7 December 2010, Italian Institute of Technology (IIT), Genova, Italy

The tutorial presents modern computer vision techniques to deal with the registration and 3D reconstruction of deformable shapes using images. The aim of this 2-day course is to present a principled procedure for dealing with images where each object can arbitrarily change its shape — a common occurrence in human motion analysis, medical imaging and video-surveillance scenarios. The tutorial will discuss a set of techniques general in their formulation but that can be customized given the specific imaging problem of the user. In particular, an emphasis is put over the use of physical or statistical priors which can aid the solution of such an ill-posed problem. Real examples on the human motion analysis and medical imaging domains will show the effectiveness of the approaches in dealing with different deforming shapes.

Schedule

Venue

Monday 6th of December:

2:30pm – 5:30pm: Non-rigid Image Registration

Tuesday 7th of December:
9:30am – 12:30pm: Template-based image registration and 3D reconstruction

14:30pm – 17:30pm: Non-rigid Structure from Motion

The tutorial will take place at:
Italian Institute of Technology
Via Morego 30, Genova , Italy

How to get to IIT

Check this page for Hotels nearby IIT

Syllabus
Non-rigid Image Registration: • Overview and general points – The deformable image registration problem – The appearance may vary strongly – Priors: statistical (eg AAMs), physical (eg rigidity, smoothness) – Problem statement using optimization – Image modeling • Parametric warps – Definition – Smoothness and piecewise smoothness, self-occlusions – The Flow-Field – Mesh-Based warps – Radial Basis Function warps – Tensor-Product warps – Generalized warps • Estimation – Feature-based vs pixel-based – Bootstrapping • Feature-based estimation – Estimation from clean point matches – Keypoint detection and matching – Robustness – Self-occlusions • Pixel-based estimation – Basic iterations – Photometric variations – Occlusions and self-occlusions	Non-rigid 3D reconstruction: • Template based 3D reconstruction of deformable surfaces. • The factorization framework. - Global 3D reconstruction - Low rank bases model - 3D models for NRSfM - The non-rigid motion - The metric constraints • Factorization approaches to NR-SfM. - A closed form solution - Iterative approaches to NR-SfM - Non-linear optimization vs closed-form solutions. - Bundle Adjustment for NR-SfM - Coarse to fine approach - EM-PPCA for NR-SfM • Statistical priors: Gaussian priors, coarse to fine, dynamics, low rank pre-trained vs un-trained models, locally linear models. • Metric projections. - The motion manifold - Augmented Lagrangian multipliers for NRSfM • Piecewise approaches: planar/quadratic. - Global vs Local - The quadratic model - Piecewise reconstruction • Trajectory space – DCT trajectory basis. • Future directions.

Syllabus

Non-rigid Image Registration:
• Overview and general points
– The deformable image registration problem
– The appearance may vary strongly
– Priors: statistical (eg AAMs), physical (eg rigidity, smoothness)
– Problem statement using optimization
– Image modeling
• Parametric warps
– Definition
– Smoothness and piecewise smoothness, self-occlusions
– The Flow-Field
– Mesh-Based warps
– Radial Basis Function warps
– Tensor-Product warps
– Generalized warps
• Estimation
– Feature-based vs pixel-based
– Bootstrapping
• Feature-based estimation
– Estimation from clean point matches
– Keypoint detection and matching
– Robustness
– Self-occlusions
• Pixel-based estimation
– Basic iterations
– Photometric variations
– Occlusions and self-occlusions

Non-rigid 3D reconstruction:
• Template based 3D reconstruction of deformable surfaces.
• The factorization framework.
- Global 3D reconstruction
- Low rank bases model
- 3D models for NRSfM
- The non-rigid motion
- The metric constraints
• Factorization approaches to NR-SfM.
- A closed form solution
- Iterative approaches to NR-SfM
- Non-linear optimization vs closed-form solutions.
- Bundle Adjustment for NR-SfM
- Coarse to fine approach
- EM-PPCA for NR-SfM
• Statistical priors: Gaussian priors, coarse to fine, dynamics, low rank pre-trained vs un-trained models, locally linear models.
• Metric projections.
- The motion manifold
- Augmented Lagrangian multipliers for NRSfM
• Piecewise approaches: planar/quadratic.
- Global vs Local
- The quadratic model
- Piecewise reconstruction
• Trajectory space – DCT trajectory basis.
• Future directions.