Download presentation
Presentation is loading. Please wait.
1
Extension of M-VOTE: Improving Feature Detection
Presented by Wyman Group Meeting 24 Oct 2005 Good afternoon everybody, today I would like to present one possible extension of my past FYP project. How to improve the feature selection stage’s algorithm. I will present 2 feature selection algorithms from 2 papers. I am still looking for a MPhil topic, and this may be one of my choice. It may be a over easy topic.
2
Outline Feature Detection Rotational motion detection
KLT detector SUSAN detector Rotational motion detection How zooming of camera enhances interaction with user Though three possible extensions. But today I will just focus on the feature selection part. Focus on two detector KLT and Susan.
3
Feature Detection To detect points of interest (feature) out of an image Requirement on points of interest High contrast and different from nearby blocks Corners are used in motion analysis Motion is ambiguous at an edge but unambiguous at a corner Corner detectors Plessey detector KLT detector Susan detector Purpose of feature selection on an image is to … Requirement… Corner is always used… 3 common corner detectors,
4
What Is a Corner? Corners are local image features characterised by locations where variations of intensity function f(x, y) in both X and Y directions are high Intersection points between two or more edge segments Before i start to introduce what is the two corner detectors, I would like to give a definition on Corner. Image corners are the projections of physical 3-D corners, termed objective structures. Corners are …
5
Corner Detectors Corner detection should satisfy a number of important criteria: All the true corners should be detected No false corners should be detected Corner points should be well localized Corner detector should be robust with respect to noise Corner detector should be efficient Mark corner point at the corner Perform well even when the image has noise fast
6
Families of Corner Detectors
Work directly with the values of brightness of images Usually based on the study of derivatives (orientation, magnitude) of grey-level or color image Extract object boundaries first then analyze its shape Use edge detectors first then analyze the curvature of boundaries Less reliable Two types Work on intensity of image Work on the shape of boundaries Today only talk about the corner detector of the first type
7
KLT Corner Detector Search for points where variations in two orthogonal directions are large Based on local structure matrix Derivatives of the intensity function f(x, y) are calculated in each point Then, the entries of the matrix are obtained Each of the entries is smoothed by Gaussian filter What It do is to … It is based on the … Image is represented by an intensity function f(x, y) The derivatives (fx, fxfy, fy) are first calculated at each point We obtain the entries of the matrix (fx, fxfy) at each point Normally the entries are smoothed
8
KLT Corner Detector The local structure matrix Cstr can be diagonalised by rotation of the coordinate axes. The diagonal entries will be the two eigenvalues The eignevalues are nonnegative Assume λ1>=λ2>=0 For a perfectly uniform image: Cstr = 0 and λ1=λ2=0 For a perfectly black-and-white step edge: λ1>0, λ2=0, where the eigenvector associated with λ1 is orthogonal to the edge For a corner of black square against a white background: λ1>=λ2>0. The higher contrast in that direction, the larger the eigenvalue A corner is identified by two strong edges. That is, a corner is a location where the smaller eigenvalue,λ2 , is large enough We need to get the eigenvalues of this local structure matrix The eigenvalues have these properties…
9
KLT Corner Detector The KLT corner detector has two parameters: the threshold on λ2 denoted by λthr, and the linear size of a square window D The algorithm Compute fx and fy over the entire image f(x, y) For each image point p: form the matrix Cstr over a DxD neighbourhood of p; compute λ2, the smaller eigenvalue of Cstr; if λ 2 >λthr , save the p into a list, L. Sort L in decreasing order of λ2. Scan the sorted list from top to bottom. For each current point, p, delete all points appearing further in the list, which belong to the neighbourhood of p
10
KLT Corner Detector The threshold λthr can be estimated from the histogram of λ2 There is no simple criterion for the window size D. Values between 2 and 10 are adequate in most practical cases But some corners which are close to each other may be lost for larger D Window size D controls how close the corner point can be
11
Example of corner detection
12
The SUSAN Detector Proposed by Smith and Brady in 1995
It doesn’t use any derivatives SUSAN stands for Smallest “Univalue Segment Assimilating Nucleus” It is based on the fact that each point within an image has associated with it a local area of comparable brightness The Principle It generates a circular mask around a given point in an image It compares the intensity of neightbouring pixels with that of the centre pixel (nucleus of the mask), the area with similar intensity to the nucleus is called USAN area Repeat the procedure for each pixel within the image
13
The SUSAN Detector USAN area varies within the image depending on its location with respect to special features of the image The USAN area is maximum within the rectangular area but falls to a minimum at an edge and to an even smaller value corresponding to a local minimum at a corner This is the property upon which the corner finder algorithm is based, hence the name SUSAN Different location, different USAN With reference to following figure, it can be seen that the USAN area varies within the image depending on its location with respect to special features of the image. The USAN\ area is maximum within the rectangular area but falls to a minimum at an edge and to an even smaller value corresponding to a local minimum at a corner. This is the property upon which the corner finder algorithm is based, hence the name SUSAN , in detecting uni-dimensional (edge) or two-dimensional (corner) features within an image. This project will, however, only consider the second-order or corner features in the context of motion detection.
14
The SUSAN Detector The algorithm
Determine a circular mask, typically of 37 pixels around a nucleus for each point within the image Calculate the difference in brightness between each pixel of the mask and that of its nucleus and Sum the number of pixels within the circular mask which have similar intensity levels to that of the nucleus basic similarity function where r is a given pixel location within the mask and ro is the position of the mask nucleus in the image frame. I(r) refers to the intensity of the pixel at location r , t is the brightness threshold, and c is the output of the comparison in intensities at locations r and ro . The sum of the comparison outputs (c) is then taken and that represents the total number of pixels in the USAN region, in other words, the USAN area:
15
The SUSAN Detector Compare n with g, the geometric threshold which is set to half of the maximum value that n can be (nmax/2) At a perfect corner (where two straight edges intersect) the USAN area will always be less than half the size of the mask area, and will be a local minimum is the corner response. The above rule is a simple formulation of the SUSAN principle since the response is inversely proportional to the USAN area.
16
The SUSAN Detector (Improved)
However, the above process may generate “false” corners There are a couple of experiments to remove false positives arising from strong edges and noise The first test consists in finding the centre of gravity of the USAN area Only those pixels falling on a straight line from the nucleus of the mask to its centre of gravity and forming part of the USAN area will constitute a proper corner Unfortunately, other scene phenomena can also generate second-order intensity variations in the image and hence be marked incorrectly as ``false'' corners, termed subjective structures. Therefore, once a corner strength image has been derived from the geometric threshold, the resulting image is subject to a couple of experiments to remove false positives arising from strong edges and noise. The first task consists in finding the centre of gravity [120, page 10,] of the USAN area. Those USAN areas corresponding to proper corners will have centres of gravity sufficiently far from the nucleus of the mask (cf. Figure 5.3) unlike false positives (for example a straight edge crossing a region of uniform brightness twice that of the edge) where the centre of gravity will be very close to the nucleus. This observation can be used to screen out the false positives derived from edges. The second test which only applies to real images subject to excessive noise looks up for contiguity within the USAN area. Only those pixels falling on a straight line from the nucleus of the mask to its centre of gravity and forming part of the USAN area will constitute a proper corner. The remaining areas of the mask will only contain false corners and so can be discarded.
17
The SUSAN Detector The initial corner response image is subjected to non-maximal suppression Local maxima are searched for in smaller five-by-five pixel regions Find a pixel which has a response greater than a given threshold is also greater than all its immediate neighbors In other words, if at least one of those neighboring pixels has a magnitude higher than that of the examined pixel then the latter is assigned a zero value
18
The SUSAN Detector SUSAN has better accuracy, reliability and speed than Plessey corner detector SUSAN took 0.3 seconds to process this picture on a single Sun SPARC-2 processor; the Plessey corner finder took 3.5 seconds Output of the SUSAN corner detector (t=10) given the test image. (0.3 sec) Output of the Plessey corner detector (o =2.0) given the test image. (3.5 sec)
19
SUSAN Corner Detection Result
Depend on threshold, higher threshold will give better result, yet less corner detected
20
Advantages of SUSAN Detector
It performs well even in the presence of noise No image derivatives are used Integrating effect and non-linear response give strong noise rejection The use of controlling parameters is much simpler and less arbitrary
21
Problems Are these corner detection algorithms fast enough for mobile devices? Floating point operation consideration How to extend the algorithms to detect corners in color image? Robust feature matching across widely separated color images. Proceedings of the 17th International Conference on Pattern Recognition (ICPR’04)
22
Reference S.M. Smith. SUSAN - a new approach to low level image processing. Internal Technical Report TR95SMS1, Defence Research Agency, Chobham Lane, Chertsey, Surrey, UK, 1995. J. Shi and C. Tomasi. Good features to track. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR94), Seattle, June 1994.
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.