Segmentation Course web page: vision.cis.udel.edu/~cv May 2, 2003 Lecture 29.

Segmentation

Course web page:vision.cis.udel.edu/~cv

May 2, 2003 Lecture 29

Announcements

• Read Forsyth & Ponce Chapter 14.4 and Chapter 25 on clustering and digital libraries, respectively

Outline

• Definition of segmentation• Grouping strategies• Segmentation applications

– Detecting shot boundaries– Background subtraction

What is Segmentation?

• Clustering image elements that “belong together” – Partitioning

• Divide into regions/sequences with coherent internal properties

– Grouping • Identify sets of coherent tokens in image

• Tokens: Whatever we need to group – Pixels – Features (corners, lines, etc.) – Larger regions (e.g., arms, legs, torso)– Discrete objects (e.g., people in a crowd)– Etc.

Example: Partitioning by Texture

courtesy of University of Bonn

Fitting

• Associate model(s) with tokens– Estimation: What are parameters of

model for a given set of tokens?• Least-squares, etc.

– Correspondence: Which token belongs to which model?• RANSAC, etc.

Approaches to Grouping

• Bottom up segmentation– Tokens belong together because they are

locally coherent• Top down segmentation

– Tokens belong together because they lie on the same object—must recognize object first

– RANSAC implements this in a very basic form• Not clear how to apply to higher-level concepts (i.e.,

objects for which we lack analytic models)

• Not mutually exclusive—successful algorithms generally require both

Gestalt Theory of Grouping

• Psychological basis for why/how things are grouped• Figure-ground discrimination

– Grouping can be seen in terms of allocating tokens to figure or ground

• Factors affecting token coherence– Proximity– Similarity: Based on color, texture, orientation (aka parallelism), etc.– Common fate: Parallel motion (i.e., segmentation of optical flow by

similarity)– Common region: Tokens that lie inside the same closed region tend

to be grouped together.– Closure: Tokens or curves that tend to lead to closed curves tend to

be grouped together.– Symmetry: Curves that lead to symmetric groups are grouped

together– Continuity: Tokens that lead to “continuous” — as in “joining up

nicely,” rather than in the formal sense — curves tend to be grouped– Familiar Configuration: Tokens that, when grouped, lead to a

familiar object—e.g., the top-down recognition that allows us to see the dalmation from Forsyth & Ponce

Gestalt Grouping Factors

from Forsyth & Ponce

Example: Bottom-Up Segmentation

Segmenting cheese curds by texture(note importance of scale!)

Example: Top-Down Segmentation

from Forsyth & Ponce

Application:Shot Boundary Detection

• The problem: Divide video footage into a set of shots– Each shot is a continuous sequence of frames

from one camera• Types

– Cut: Shot changes in one frame– Fade, wipe, dissolve, etc.: Multi-frame

transition• Applications

– Video editing easier since shots become tokens– Can summarize video with key frames from

each shotfrom M. Smith

& T. Kanade

Shot Boundary Detection

• Basic approach: Threshold inter-frame difference

• Possible metrics – Raw: SSD, correlation, etc.

• More sensitive to camera motion– Histogram– Edge comparison– Break into blocks

• Use hysteresis to handle gradual transitions

from M. Smith & T. Kanade

Graph of frame-to-frame histogram difference

Application: Background Subtraction

• The problem: Assuming static camera, discrimin-ate moving foreground objects from background

• Applications– Traffic monitoring– Surveillance/security– User interaction

Current imagefrom C. Stauffer and W. Grimson

Background image Foreground pixels

courtesy of C. Wren

Pfinder

Background Subtraction: Simple Approaches

• Adjacent Frame Difference: Each image is subtracted from previous image in sequence. Absolute pixel differences greater than a threshold are

marked as foreground (It > ¿)

• Mean & Threshold: Pixel-wise mean values are computed during training phase; pixels within fixed threshold of the mean are considered background

adapted from K. Toyama et al.

Results & Problems for Simple Approaches

from K. Toyama et al.

Background Subtraction: Issues

• Noise models– Unimodal: Pixel values vary over time even for

static scenes– Multimodal: Features in background can “oscillate”,

requiring models which can represent disjoint sets of pixel values (e.g., waving trees against sky)

• Gross illumination changes– Continuous: Gradual illumination changes alter the

appearance of the background (e.g., time of day)– Discontinuous: Sudden changes in illumination and

other scene parameters alter the appearance of the background (e.g., flipping a light switch)

• Bootstrapping– Is a training phase with “no foreground” necessary,

or can the system learn what’s static vs. dynamic online?

Pixel RGB Distributions over time

Perceived color values of solid objects (e.g., tree trunk) have roughly Gaussian distributions due to CCD noise, etc. Leaf & monitor pixels

have bimodal distributions because of waving & flickering, respectively

courtesy of J. Buhmann

Improved Approaches to Background Subtraction

• Mean & Covariance: Mean and covariance of pixel values are updated continuously: – Moving average is used to adapt to slowly

changing illumination (low-pass temporal filter)

Foreground pixels are determined using a threshold on the Mahalanobis distance

• Mixture of Gaussians: A pixel-wise mixture of multiple Gaussians models the background adapted from K. Toyama et al.

Ellipsoids of Constant Probability for Gaussian

Distributions

from Duda et al.

Fitting Gaussians to Color Distributions

Can parametrize scaling, rotation, translationof ellipsoid with SVD of covariance matrix

Mahalanobis Distance

• Distance of point from Gaussian distribution– Along axes of fitted

ellipsoid– In units of standard

deviations (i.e., scaled)

covariance matrix

X(2, 2)

adapted from Duda & Hart

Example: Background Subtraction for Surveillance

courtesy of Elgammal et al.