文档搜索 > Slide 1

Slide 1


Segmentation and Grouping 

Computer Vision

CS 543 / ECE 549

University of Illinois 

Derek Hoiem 

02/23/10


Outline 
 

  • Goals of segmentation
    • Over-segmentation (for efficiency or features)
    • Get support regions for computing features (multiple segmentations)
    • Get object regions
  • Soft boundaries + watershed  
  • Mean-shift 
  • Gestalt cues 
  • Normalized cuts 
  • Other methods 
    • Felzenszwalb and Huttenlocher

 


Last week 
 

  • Clustering
  • EM 
     

Today’s class 
 

  • More on EM
  • Segmentation and grouping 
    • Gestalt cues
    • By boundaries (watershed)
    • By clustering (mean-shift)

 


EM


Gestalt grouping


German: Gestalt - "form" or "whole” 
 
 

Berlin School, early 20th century 

Kurt Koffka, Max Wertheimer, and Wolfgang K?hler  

View of brain:

  • whole is more than the sum of its parts
  • holistic
  • parallel
  • analog
  • self-organizing tendencies
 
 

Slide from S. Saverese 

Gestalt psychology or gestaltism


The Muller-Lyer illusion 

Gestaltism


Explanation


Principles of perceptual organization 

From Steve Lehar: The Constructive Aspect of Visual Perception


Principles of perceptual organization


From Steve Lehar: The Constructive Aspect of Visual Perception 

Grouping by invisible completion


From Steve Lehar: The Constructive Aspect of Visual Perception 

Grouping involves global interpretation


Grouping involves global interpretation 

From Steve Lehar: The Constructive Aspect of Visual Perception


Gestaltists do not believe in coincidence


Emergence

 


Gestalt cues 
 

  • Good intuition and basic principles for grouping
  • Difficult to implement in practice 
  • Sometimes used for occlusion reasoning 

Moving on to image segmentation … 

    Goal: Break up the image into meaningful or perceptually similar regions


Segmentation for feature support 

50x50 Patch 

50x50 Patch


Segmentation for efficiency 

[Felzenszwalb and Huttenlocher 2004] 

[Hoiem et al. 2005, Mori 2005] 

[Shi and Malik 2001]


Segmentation as a result 

Rother et al. 2004


Types of segmentations 

Oversegmentation 

Undersegmentation 

Multiple Segmentations


Major processes for segmentation 
 

  • Bottom-up: group tokens with similar features
  • Top-down: group tokens that likely belong to the same object
 

[Levin and Weiss 2006]


Segmentation using clustering 
 

  • Kmeans
  • Mean-shift 

 


Source: K. Grauman 

Feature Space


Image 

Clusters on intensity 

Clusters on color 

K-means clustering using intensity alone and color alone


K-Means pros and cons 

  • Pros
    • Simple and fast
    • Easy to implement
  • Cons 
    • Need to choose K
    • Sensitive to outliers
  • Usage 
    • Rarely used for pixel segmentation
  • Versatile technique for clustering-based segmentation
 

D. Comaniciu and P. Meer, Mean Shift: A Robust Approach toward Feature Space Analysis, PAMI 2002.  

Mean shift segmentation


Kernel density estimation 

Kernel density estimation function 

Gaussian kernel


Mean shift algorithm 
 

  • Try to find modes of this non-parametric density

Region of

interest 

Center of

mass 

Mean Shift

vector 

Slide by Y. Ukrainitz & B. Sarel 

Mean shift


Region of

interest 

Center of

mass 

Mean Shift

vector 

Slide by Y. Ukrainitz & B. Sarel 

Mean shift


Region of

interest 

Center of

mass 

Mean Shift

vector 

Slide by Y. Ukrainitz & B. Sarel 

Mean shift


Region of

interest 

Center of

mass 

Mean Shift

vector 

Mean shift 

Slide by Y. Ukrainitz & B. Sarel


Region of

interest 

Center of

mass 

Mean Shift

vector 

Slide by Y. Ukrainitz & B. Sarel 

Mean shift


Region of

interest 

Center of

mass 

Mean Shift

vector 

Slide by Y. Ukrainitz & B. Sarel 

Mean shift


Region of

interest 

Center of

mass 

Slide by Y. Ukrainitz & B. Sarel 

Mean shift


Simple Mean Shift procedure:

  • Compute mean shift vector
  • Translate the Kernel window by m(x) 
 

Computing the Mean Shift


Real Modality Analysis


  • Attraction basin: the region for which all trajectories lead to the same mode
  • Cluster: all data points in the attraction basin of a mode
 
 

Slide by Y. Ukrainitz & B. Sarel 

Attraction basin


Attraction basin


Mean shift clustering 

  • The mean shift algorithm seeks modes of the given set of points
    1. Choose kernel and bandwidth
    2. For each point:
      1. Center a window on that point
      2. Compute the mean of the data in the search window
      3. Center the search window at the new mean location
      4. Repeat (b,c) until convergence
    3. Assign points that lead to nearby modes to the same cluster

 


  • Find features (color, gradients, texture, etc)
  • Set kernel size for features Kf and position Ks
  • Initialize windows at individual pixel locations
  • Perform mean shift for each window until convergence
  • Merge windows that are within width of Kf and Ks
 

Segmentation by Mean Shift


http://www.caip.rutgers.edu/~comanici/MSPAMI/msPamiResults.html 

Mean shift segmentation results


http://www.caip.rutgers.edu/~comanici/MSPAMI/msPamiResults.html


Mean-shift: other issues 

  • Speedups
    • Uniform kernel (much faster but not as good)
    • Binning or hierarchical methods
    • Approximate nearest neighbor search
  • Methods to adapt kernel size depending on data density
  • Lots of theoretical support

        D. Comaniciu and P. Meer, Mean Shift: A Robust Approach toward Feature Space Analysis, PAMI 2002.  

 


Mean shift pros and cons 

  • Pros
    • Good general-practice segmentation
    • Finds variable number of regions
    • Robust to outliers
  • Cons
    • Have to choose kernel size in advance
    • Original algorithm doesn’t deal well with high dimensions
  • When to use it
    • Oversegmentatoin
    • Multiple segmentations
    • Other tracking and clustering applications

Watershed algorithm


Watershed segmentation 

Image 

Gradient 

Watershed boundaries


Meyer’s watershed segmentation 
 

  1. Choose local minima as region seeds
  2. Add neighbors to priority queue, sorted by value
  3. Take top priority pixel from queue
    1. If all labeled neighbors have same label, assign to pixel
    2. Add all non-marked neighbors
  4. Repeat step 3 until finished
 

Meyer 1991 

Matlab: seg = watershed(bnd_im)


Simple trick 

  • Use median filter to reduce number of regions

Watershed usage 
 

  • Use as a starting point for hierarchical segmentation
    • Ultrametric contour map (Arbelaez 2006)
  • Works with any soft boundaries 
     
     
    • Pb
    • Canny
    • Etc.

 


Watershed pros and cons 
 

  • Pros
    • Fast (< 1 sec for 512x512 image)
    • Among best methods for hierarchical segmentation
  • Cons
    • Only as good as the soft boundaries
    • Not easy to get variety of regions for multiple segmentations
    • No top-down information
  • Usage 
     
    • Preferred algorithm for hierarchical segmentation

Things to remember 

  • Gestalt cues and principles of organization
  • Uses of segmentation 
    • Efficiency
    • Better features
    • Want the segmented object
  • Mean-shift segmentation 
    • Good general-purpose segmentation method
    • Generally useful clustering, tracking technique
  • Watershed segmentation 
    • Good for hierarchical segmentation
    • Use in combination with boundary prediction

Further reading 
 

  • Mean-shift paper by Comaniciu and Meer

http://www.caip.rutgers.edu/~comanici/Papers/MsRobustApproach.pdf 
 

  • Adaptive mean shift in higher dimensions

http://mis.hevra.haifa.ac.il/~ishimshoni/papers/chap9.pdf 
 

  • Contours to regions (watershed): Arbelaez et al. 2009

http://www.eecs.berkeley.edu/~arbelaez/publications/Arbelaez_Maire_Fowlkes_Malik_CVPR2009.pdf 
 

 


Next class 
 

  • Graph-based segmentation
    • Normalized cuts
    • Graph cuts

设为首页 | 加入收藏 | 昂纲搜索

All Rights Reserved Powered by 文档下载网

Copyright © 2011
文档下载网内容来自网络,如有侵犯请和我们联系。tousu#anggang.com
返回顶部