Object Detection by Global Contour Shape

You might also like

Download as pptx, pdf, or txt
Download as pptx, pdf, or txt
You are on page 1of 27

Presented by S Raghavender B080520CS

Introduction
The aim of this work is to investigate the potential of

global shape as a cue for object detection and recognition. This was the research done in two fields, digital image processing and pattern recognition. This is method for object class detection in images based on global shape. If an image is given we have to find the objects that are in that image comparing it with given templates.

Example
Template Object found in the image

Basic Steps used in a Normal method


RGB to gray conversion Filtering the image

Thresholding
Edge detection Recognition of connected components Correlating the component with a template.

RGB to gray conversion


In this step a RGB (color) image is converted to a gray

scale image using following formula. I = 0.3R + 0.59G + 0.11B A gray scale (mxn pixels)image is represented using a two dimensional matrix (mxn size) A RGB image (mxn pixels)is represented using a three dimensional matrix (mxnx3 size)

Example RGB to Gray scale

RGB Image

Gray scale image

Filtering the image


An image is filtered to remove noises present in the

image. There are many types of filters Median filter Low pass filter High pass filter Band pass filter

Example of filtering

Noisy Image

Filtered Image

Thresholding
Thresholding is a technique which divides the image

in to different segments depending on the intensity of the color at that point. A threshold may have different intensities. If there are only 2 intensities then they will be balck and white. Thresholding is done based on the calculated threshold value of the image.

Edge Detection
In the stage of edge detection all the edges that are

present in the image are found. There are different types of operators to find the edges in the given image. Example : Sobel Operator Laplacian operator

Example edge detection

Recognition of connected components


Finding the connected components deals with finding

different objects that are present in the given image. There different types of algorithms available for finding the connected components.

Correlation
Correlation of the connected components with

template image deals with resizing the image to template size and matching the image with different templates and finding the most matching template from the set of the templates. Here we have to mainly deal with finding the contour (shape) of the given object and match it with the template image and find the corresponding object.

Disadvantages of above method


Cannot detect the object if it is in different angle.
Template should match properly with the object. Cannot find the objects which are oriented in different

way.

Detection by contour matching


The major steps involved in object detection are Segmentation Finding the shape model Matching the shape Computing space distance Detecting objects

Segmentation vs. Edge detection


The first step towards shape-based object detection is

to extract potential object contour points from the input image, which then are compared to a shape template. The short-comings of this basic edge detection has been one of the major difficulties for shape-based object detection. An example of segmentation is Statistical Region Merging

Example : Segmentation

Shape Model
A shape X is approximated by a closed polygon with a

fixed number of equally spaced vertices N. Since the points are equally spaced, the sequence of points can be parametrised by an integer arc length: X={x(u),u=0,,N-1}. The last vertex coincides with the first one for computational simplicity: x(N)=x(0). To dive the shape of the curvatures a method is used in which tangents are drawn to every point and matched with the corresponding shapes.

Matching shapes
Given are two shapes X and Y, one for the shape

template, and one for a candidate contour extracted from a test image. A matching between the two shapes is a function, which associates the point sets {x(u)} and {y(v)} (both parametrised by their arc length), such that each point on either curve has at least one corresponding point on the other curve.

Matching Shape (contd..)


To achieve a rotation of contour X relative to contour Y,

the tangent angles x(u) have to be shifted in a circular manner, and then re-normalised such that x(0)=0 for the new starting point.

Computing shape distance


This is used for comparing the objects or connected

components with templates.

Detecting Objects
The nonlinear elastic matching distance is a measure

for the similarity between two contours. Any group of neighboring super-pixels forms a closed contour, and the combinatorial set of all such contours is the search space for object detection. If the shape of segments and distance between them matches with that of template then object is detected. Since correlation is based on probability the results may not be 100% accurate.

Experimental results
A collection of four diverse object classes are used,

which have in common that they are mainly defined by their global shape, while they have either little texture at all, or strongly varying texture. In a detailed experimental evaluation, it has been shown to outperform previous methods, and has achieved a detection rate of 91%.

Example

References
1. Konard Schindler, David Suter, Obect detection by global contour matching. 2. R. Nock and F. Nielsen, Statistical region merging. IEEE Trans. Pattern Anal. Mach. Intell., 26 11 (2004), pp. 14521458. 3. L.J. Latecki and R. Lakmper, Shape similarity measure based on correspondence of visual parts. IEEE Trans. Pattern Anal. Mach. Intell., 22 10 4. Wikipedia

You might also like