ICCV 2005: Papers I Liked

1
ICCV 2005 Papers I Liked
Beijing, China October 15-21

• Vaibhav Vaish

2
Conference Statistics

• 244 papers
• 45 oral presentations
• 199 posters
• 1670 abstracts registered
• 11 workshops
• 7 short courses
• Computer Vision Contest

3
Papers I Liked A Categorisation

• Down with parameters!
• When light behaves badly
• Vision discovers Stephen Boyd
• SIGGRAPH Techniques
• Multi-dimensional Miscellany

4
Contest Where am I ?

• Given some images of a city taken from known
  locations, estimate locations of other images

• Contest web page with data sets, results, other
  contest ideas proposed
• Rick Szeliskis PPT (overview)

5
Non-parametric Methods
Space-time Video Completion, Wexler et al, CVPR
2004

• Space-time Scene Manifolds
• Y. Wexler, D. Simakov
• Detecting Irregularities in Images and Video
• O. Boiman, M. Irani

6
Space-time Scene Manifolds
Input Video
7
Space-time Scene Manifolds
8
Space-time Scene Manifolds
Input Video
9
Space-time Scene Manifolds
10
Detecting Irregularities in Images and Video

• How well can we compose an image region with
  patches from observed image(s) ?

11
Detecting unseen poses
Input database
12
Detecting unseen behavior
13
Detecting Saliency in Video
14
Detecting Saliency in a Single Image
15
Discussion

• Unsupervised learning method
• Training set easy to acquire
• Learn large class of poses / behaviors from small
  set of images
• Technical Contributions
• Patch descriptors
• Inference algorithm
• Potential applications
• Detecting click fraud in click stream ?
• Detecting suspicious behavior in network traffic
  ?
• Finding unusual cells in microscope images ?

Honorable Mention, Marr Prize
16
When Light Behaves Badly

• Structured Light in Scattering Media
• S. Narasimhan, S. Nayar, B. Sun, S. Koppal
• Dynamic Refraction Stereo
• N. Morris, K. Kutulakos
• A Theory of Refractive and Specular 3D Shape by
  Light-Path Triangulation
• K. Kutulakos, E. Steger
• A Theory of Inverse Light Transport
• S. Seitz, Y. Matsushita, K. Kutulakos

17
Structured Light in Scattering Media

• Stripe-based Range Scanning

• Detect stripe in presence of scattering
• Remove effect of scatter on appearance

18
Structured Light in Scattering Media

• Photometric Stereo

• Requires minimum 5 views
• Recovers normals and depth

Video
19
Refractive and Specular 3D Shape

• Given
• N calibrated cameras
• K light bounces (reflection or refraction)
• M 3D reference points for every pixels light
  path
• When can we reconstruct depth per pixel ?

20
Refractive and Specular 3D Shape

• lt1 camera, 1 bounce, 2 reference pointsgt is
  solvable
• (simply by ray intersection)

21
Refractive and Specular 3D Shape

• lt3 cameras, 2 bounces, 2 ref. pointsgt is solvable
  given refractive index / mirror knowledge
• ltN cameras, 3 bounces, 2 ref. pointsgt is not
  solvable

22
The Complete Analysis
23
Experiment lt5,2,2gt Triangulation
24
Experiment lt5,2,2gt Triangulation
Honorable Mention, Marr Prize
25
EE 364 _at_ ICCV Global Minima

• Multiple View Geometry and the L8 norm
• F. Kahl
• Quasiconvex Optimization for Robust Geometric
  Reconstruction
• Q. Ke, T. Kanade
• Globally Optimal Estimates for Geometric
  Reconstruction Problems
• F. Kahl, D. Henrion
• Globally Optimal Solutions for Energy
  Minimization in Stereo Vision using Reweighted
  Belief Propagation
• T. Meltzer, C. Yanover, Y. Weiss

26
Seeking the Global Minimum

• Reconstruction problems in vision
• Epipolar geometry from a stereo pair
• Homography between 2 planes
• Multi-view triangulation (3D reconstruction)
• Camera calibration
• Current Approaches
• Find correspondences
• Initial estimate (SVD, linear optimization)
• Bundle adjustment (gradient descent for local
  minima)

27
Seeking the Global Minimum

• Ke Kanade, Kahl
• Minimize the largest error (L8 norm)
• Quasiconvex problem (SOCP)
• Variant minimize mth largest error Ke Kanade
  to tolerate m-1 outliers
• Kahl Henrion
• Finds global minimum for L2 norm
• Progressive approximation by SDP
• Several parameters / thresholds to be set

28
Kahl Henrion Global L2 Minimum
Results indicate in practice, bundle adjustment
is a good estimator of global minimum
Marr Prize
29
Techniques _at_ SIGGRAPH 2005

• Geometric Context from a Single Image
• D. Hoiem, A. Efros, M. Herbert
• An Iterative Image Optimization Approach for
  Unified Image Segmentation and Matting
• J. Wang, M. Cohen
• An Algebraic Approach to Surface Reconstruction
  from Gradient Fields
• A. Agrawal, R. Chellappa, R. Raskar

30
Miscellaneous

• Sparse Image Encoding using a 3D Non-negative
  Tensor Factorization
• T. Hazan, S. Polak, A. Shashua
• A Unifying Approach to Hard and Probabilistic
  Clustering
• R. Zass, A. Shashua
• A Graph Cut Algorithm for Generalized Image
  Deconvolution
• A. Raj, R. Zabih
• Efficiently Registering Video into Panoramic
  Mosaics
• D. Steedley, C. Pal, R. Szeliski