This action might not be possible to undo. Are you sure you want to continue?

Konstantinos G. Derpanis

August 15, 2005.

Mean shift represents a general non-parametric mode ﬁnding/clustering proce-

dure. In contrast to the classic K-means clustering approach (Duda, Hart & Stork,

2001), there are no embedded assumptions on the shape of the distribution nor the

number of modes/clusters. Mean shift was ﬁrst proposed by Fukunaga and Hostetler

(Fukunaga & Hostetler, 1975), later adapted by Cheng (Cheng, 1995) for the purpose

of image analysis and more recently extended by Comaniciu, Meer and Ramesh to

low-level vision problems, including, segmentation (Comaniciu & Meer, 2002), adap-

tive smoothing (Comaniciu & Meer, 2002) and tracking (Comaniciu, Ramesh & Meer,

2003).

The main idea behind mean shift is to treat the points in the d-dimensional feature

space as an empirical probability density function where dense regions in the feature

space correspond to the local maxima or modes of the underlying distribution. For

each data point in the feature space, one performs a gradient ascent procedure on the

local estimated density until convergence. The stationary points of this procedure

represent the modes of the distribution. Furthermore, the data points associated (at

least approximately) with the same stationary point are considered members of the

same cluster.

Next we review several of the technical details behind mean shift (for further

details, see (Cheng, 1995; Comaniciu & Meer, 2002)).

Given n data points x

i

∈ R

d

, the multivariate kernel density estimate using a

radially symmetric kernel

1

(e.g., Epanechnikov and Gaussian kernels), K(x), is given

by,

ˆ

f

K

=

1

nh

d

n

i=1

K

_

x −x

i

h

_

, (1)

where h (termed the bandwidth parameter) deﬁnes the radius of kernel. The radially

symmetric kernel is deﬁned as,

K(x) = c

k

k(x

2

), (2)

1

Note that the mean shift procedure has been extended to anisotropic kernels (Wang, Thiesson,

Xu & Cohen, 2004).

1

where c

k

represents a normalization constant. Taking the gradient of the density

estimator (1) and some further algebraic manipulation yields,

∇

ˆ

f(x) =

2c

k,d

nh

d+2

_

n

i=1

g

_

_

_

_

_

x −x

i

h

_

_

_

_

2

__

. ¸¸ .

term 1

_

¸

¸

_

n

i=1

x

i

g

_

_

_

x−x

i

h

_

_

2

_

n

i=1

g

_

_

_

x−x

i

h

_

_

2

_

−x

_

¸

¸

_

. ¸¸ .

term 2

, (3)

where g(x) = −k

**(x) denotes the derivative of the selected kernel proﬁle. The ﬁrst
**

term is proportional to the density estimate at x (computed with the kernel G =

c

g

g(x

2

)). The second term, called the mean shift vector, m, points toward the

direction of maximum increase in density and is proportional to the density gradient

estimate at point x obtained with kernel K. The mean shift procedure for a given

point x

i

is as follows: (see Fig. 1):

1. Compute the mean shift vector m(x

t

i

).

2. Translate density estimation window: x

t+1

i

= x

t

i

+ m(x

t

i

).

3. Iterate steps 1. and 2. until convergence, i.e., ∇f(x

i

) = 0.

For a proof of convergence, see (Comaniciu & Meer, 2002).

The most computationally expensive component of the mean shift procedure cor-

responds to identifying the neighbours of a point in space (as deﬁned by the kernel

and its bandwidth); this problem is known as multidimensional range searching in the

computational geometry literature. This computation becomes unwieldly for high di-

mensional feature spaces. Proposed solutions to this problem include, embedding the

mean shift procedure into a ﬁne-to-coarse hierarchical bandwidth approach (DeMen-

thon & Megret, 2002) and employing approximate nearest-neighbour hashing-based

search (Georgescu, Shimshoni & Meer, 2003).

Finally, a limitation of the standard mean shift procedure is that the value of the

bandwidth parameter is unspeciﬁed. For representative solutions to the bandwidth

selection problem, see (Comaniciu, Ramesh & Meer, 2001; Singh & Ahuja, 2003;

Wang, Thiesson, Xu & Cohen, 2004).

References

Cheng, Y. (1995). Mean shift, mode seeking, and clustering. IEEE Transactions on

Pattern Analysis and Machine Intelligence, 17(8), 790–799.

Comaniciu, D. & Meer, P. (2002). Mean shift: A robust approach toward feature

space analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence,

24(5), 603–619.

2

0

i

x

1

i

x

n

i

x

Figure 1: Mean shift procedure. Starting at data point x

i

, run the mean shift pro-

cedure to ﬁnd the stationary points of the density function. Superscripts denote the

mean shift iteration, the shaded and black dots denote the input data points and

successive window centres, respectively, and the dotted circles denote the density

estimation windows.

Comaniciu, D., Ramesh, V. & Meer, P. (2001). The variable bandwidth mean shift

and data-driven scale selection. In International Conference on Computer Vision

(pp. I: 438–445).

Comaniciu, D., Ramesh, V. & Meer, P. (2003). Kernel-based object tracking. IEEE

Transactions on Pattern Analysis and Machine Intelligence, 25(5), 564–577.

DeMenthon, D. & Megret, R. (2002). Spatio-temporal segmentation of video by

hierarchical mean shift analysis. Technical report, University of Maryland.

Duda, R., Hart, P. & Stork, D. (2001). Pattern Classiﬁcation. Wiley.

Fukunaga, K. & Hostetler, L. (1975). The estimation of the gradient of a density

function, with applications in pattern recognition. IEEE Transactions on Infor-

mation Theory, 21(1), 32–40.

Georgescu, B., Shimshoni, I. & Meer, P. (2003). Mean shift based clustering in high

dimensions: A texture classiﬁcation example. In International Conference on

Computer Vision (pp. 456–463).

3

Singh, M. & Ahuja, N. (2003). Regression based bandwidth selection for segmentation

using parzen windows. In International Conference on Computer Vision (pp. 2–

9).

Wang, J., Thiesson, B., Xu, Y. & Cohen, M. (2004). Image and video segmentation

by anisotropic kernel mean shift. In European Conference on Computer Vision

(pp. Vol II: 238–249).

4

- Chirp Z Transform
- ADSP
- Usb 1.1 vs Usb 2.0
- A Collection of Computer Tips
- Computer Vision
- Cam Shift
- Object detection
- CAMSHIFT
- CAMSHIFT
- An Improved CamShift Algorithm for Target Tracking in Video
- Improved CAMSHIFT
- Improved object tracking
- non rigid object tracking
- People Tracking via CAMSHIFT
- Region of interest extraction
- 200 Ways to Revive Hard Drive
- Inspiron 15 Intel n5040 Owner's Manual
- discrete kalman filter
- laviola_acc2003
- color object detection
- Pls Iccv
- Tutorial Uk f
- Unified Mtd
- ukf
- Advanced Mat Lab Tutorial

Mean shift clustering

Mean shift clustering

- EDGE DETECTION OF MICROSCOPIC IMAGEby IAEME Publication
- Ijret - Mammogram Image Segmentation Using Rough Clusteringby International Journal of Research in Engineering and Technology
- Comparison Between C-mea and Kby Mohd Nawawi Mohamad Sahini
- Segmentation of Brain Tumour from MRI image – Analysis of K-means and DBSCAN Clusteringby International Journal of Research in Engineering and Science

- IIJIT-2014-06-12-024by editorijettcs1
- An Evaluation Approach for Detection of Contours With 4-d Images a Reviewby esatjournals
- Performance Analysis of Color Image Segmentation using KMeans Clustering Algorithm in Different Color Spacesby International Organization of Scientific Research (IOSR)
- Survey on Clustering Based Color Image Segmentationby International Journal of Research in Engineering and Technology

- InTech-Image Segmentation Through Clustering Based on Natural Computing Techniquesby karteekak
- An approach for segmentation of medical images using pillar K-means algorithmby seventhsensegroup
- Data Clustering for Image Segmentationby Hai Hung
- Segmentation of Unhealthy Region of Plant Leaf Using Image Processing Techniques a Surveyby International Journal of Research in Engineering and Technology

- EDGE DETECTION OF MICROSCOPIC IMAGE
- Ijret - Mammogram Image Segmentation Using Rough Clustering
- Comparison Between C-mea and K
- Segmentation of Brain Tumour from MRI image – Analysis of K-means and DBSCAN Clustering
- IIJIT-2014-06-12-024
- An Evaluation Approach for Detection of Contours With 4-d Images a Review
- Performance Analysis of Color Image Segmentation using KMeans Clustering Algorithm in Different Color Spaces
- Survey on Clustering Based Color Image Segmentation
- Survey on Clustering Based Color Image Segmentation and Novel Approaches to Fcm Algorithm
- Image and Video Compression
- Image Segmentation Using Biogeography Based Optimization
- Mammogram Image Segmentation Using Rough Clustering
- InTech-Image Segmentation Through Clustering Based on Natural Computing Techniques
- An approach for segmentation of medical images using pillar K-means algorithm
- Data Clustering for Image Segmentation
- Segmentation of Unhealthy Region of Plant Leaf Using Image Processing Techniques a Survey
- Segmentation FrankWyman Slides
- Color Segmentation k-means algorithm
- MIP AND UNSUPERVISED CLUSTERING FOR THE DETECTION OF BRAIN TUMOUR CELLS
- Region Based Segmentation
- Manasi Final Project
- Image segmentation
- CS 223-B L11 Segmentation
- Using k Means Cluster and Fuzzy c Means for Defect Segmentation in Fruits
- Segmentation
- Extraction of Buildings in Urban Areas From Very High Resolution Satellite Images
- K-Means
- 40120130406009
- Ppt on Color Image Segmentation
- Segmentation of Lung Glandular Cells using Multiple Color Spaces
- Mean Shift Clustering [Derpanis]

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd