## Machine learning for high-speed corner detection (2006)

### Cached

### Download Links

- [mi.eng.cam.ac.uk]
- [mi.eng.cam.ac.uk]
- [edwardrosten.com]
- [www.edwardrosten.com]
- [mi.eng.cam.ac.uk]
- CiteULike
- DBLP

### Other Repositories/Bibliography

Venue: | In European Conference on Computer Vision |

Citations: | 166 - 4 self |

### BibTeX

@INPROCEEDINGS{Rosten06machinelearning,

author = {Edward Rosten and Tom Drummond},

title = {Machine learning for high-speed corner detection},

booktitle = {In European Conference on Computer Vision},

year = {2006},

pages = {430--443}

}

### OpenURL

### Abstract

Where feature points are used in real-time frame-rate applications, a high-speed feature detector is necessary. Feature detectors such as SIFT (DoG), Harris and SUSAN are good methods which yield high quality features, however they are too computationally intensive for use in real-time applications of any complexity. Here we show that machine learning can be used to derive a feature detector which can fully process live PAL video using less than 7% of the available processing time. By comparison neither the Harris detector (120%) nor the detection stage of SIFT (300%) can operate at full frame rate.

### Citations

5227 | Distinctive image features from scale-invariant keypoints
- Lowe
- 2004
(Show Context)
Citation Context ...alysis of the computation of H, and find some suitable approximations which allow them to obtain a speed increase by computing only two smoothed images, instead of the three previously required. Lowe =-=[12]-=- obtains scale invariance by convolving the image with a Difference of Gaussians (DoG) kernel at multiple scales, retaining locations which are optima in scale as well as space. DoG is used because it... |

3382 | Induction of Decision Trees
- Quinlan
- 1986
(Show Context)
Citation Context ...artitions P into three subsets, Pd,Ps,Pb, where each p is assigned to PSp→x . Let Kp be a boolean variable which is true if p is a corner and false otherwise. Stage 2 employs the algorithm used in ID3=-=[31]-=- and begins by selecting the x which yields the most information about whether the candidate pixel is a corner, measured by the entropy of Kp. The entropy of K for the set P is: H(P) = (c + ¯c)log 2(c... |

1494 | Good features to track
- Shi, Tomasi
- 1994
(Show Context)
Citation Context ...igenvalues. It has been shown[6] that the eigenvalues are an approximate measure of the image curvature. Based on the assumption of affine image deformation, a mathematical analysis led Shi and Tomasi=-=[7]-=- conclude that it is better to use the smallest eigen value of H as the corner strength function: C = min (λ1,λ2). (3) A number of suggestion have [5,7,8,9] been made for how to compute the corner str... |

995 | Scale and affine invariant interest point detectors
- Mikolajczyk, Schmid
(Show Context)
Citation Context ...2 apart,sMachine learning for high-speed corner detection 3 speeds up computation by a factor of about two, compared to the striaghtforward implementation of Gaussian convolution [13]. It is noted in =-=[14]-=- that the LoG is a particularly stable scale-space kernel. Scale-space techniques have also been combined with the Harris approach in [15] which computes Harris corners at multiple scales and retains ... |

322 | Indexing based on scale invariant interest points
- Mikolajczyk, Schmid
- 2001
(Show Context)
Citation Context ...d implementation of Gaussian convolution [13]. It is noted in [14] that the LoG is a particularly stable scale-space kernel. Scale-space techniques have also been combined with the Harris approach in =-=[15]-=- which computes Harris corners at multiple scales and retains only those which are also optima of the LoG response across scales. Recently, scale invariance has been extended to consider features whic... |

299 | Evaluation of interest point detectors
- Schmid, Mohr, et al.
- 2000
(Show Context)
Citation Context ...features produced are unsuitable for downstream processing. In particular, the same scene viewed from two different positions should yield features which correspond to the same real-world 3D locations=-=[1]-=-. Hence the second contribution of this paper is a comparison corner detectors based on this criterion applied to 3D scenes. This comparison supports a number of claims made elsewhere concerning exist... |

211 | SUSAN – a new approach to low-level image processing
- SMITH, BRADY
- 1997
(Show Context)
Citation Context ...] assumes that a corner resembles a blurred wedge, and finds the characteristics of the wedge (the amplitude, angle and blur) by fitting it to the local image. The idea of the wedge is generalised in =-=[27]-=-, where a method for calculating the corner strength is proposed which computes self similarity by looking at the proportion of pixels inside a disc whose intensity is within some threshold of the cen... |

176 |
Multi-view matching for unordered image sets, or “How do I organize my holiday snaps
- Schaffalitzky, Zisserman
- 2002
(Show Context)
Citation Context ... scales and retains only those which are also optima of the LoG response across scales. Recently, scale invariance has been extended to consider features which are invariant to affine transformations =-=[14,16,17]-=-. An edge (usually a step change in intensity) in an image corresponds to the boundary between two regions. At corners of regions, this boundary changes direction rapidly. Several techniques were deve... |

125 |
Obstacle Avoidance and Navigation in the Real World by a Seeing Robot Rover
- Moravec
- 1980
(Show Context)
Citation Context ...f feature detection algorithms work by computing a corner response function (C) across the image. Pixels which exceed a threshold cornerness value (and are locally maximal) are then retained. Moravec =-=[4]-=- computes the sum-of-squared-differences (SSD) between a patch around a candidate corner and patches shifted a small distance in a number of directions. C is then the smallest SSD so obtained, thus en... |

124 |
Gray-level corner detection
- Kitchen, Rosenfeld
- 1982
(Show Context)
Citation Context ...code[18], finding maxima of curvature [19,20,21], change in direction [22] or change in appearance[23]. Others avoid chaining edges and instead look for maxima of curvature[24] or change in direction =-=[25]-=- at places where the gradient is large. Another class of corner detectors work by examining a small patch of an image to see if it “looks” like a corner. Since second derivatives are not computed, a n... |

107 | Invariant features from interest point groups
- Brown, Lowe
- 2002
(Show Context)
Citation Context ... scales and retains only those which are also optima of the LoG response across scales. Recently, scale invariance has been extended to consider features which are invariant to affine transformations =-=[14,16,17]-=-. An edge (usually a step change in intensity) in an image corresponds to the boundary between two regions. At corners of regions, this boundary changes direction rapidly. Several techniques were deve... |

106 |
M.: A combined corner and edge detector. Alvey Vision Conference
- HARRIS, STEPHENS
- 1988
(Show Context)
Citation Context ...s shifted a small distance in a number of directions. C is then the smallest SSD so obtained, thus ensuring that extracted corners are those locations which change maximally under translations. Harris=-=[5]-=- builds on this by computing an approximation to the second derivative of the SSD with respect to the shift The approximation is: � H = � �I 2 x � IxIy �IxIy � I 2 y , (1) where � denotes averaging pe... |

88 | Comparing and evaluating interest points
- Schmid, Mohr, et al.
- 1998
(Show Context)
Citation Context ...nterestingly, the Harris detector outperforms Shi and Tomasi detector in this case. Mikolajczyk and Schmid [15] evaluate the repeatability of the Harris-Laplace detector evaluated using the method in =-=[34]-=-, where planar scenes are examined. The results show that Harris-Laplace points outperform both DoG points and Harris points in repeatability. For the box dataset, our results verify that this is corr... |

84 | Fusing points and lines for high performance tracking
- Rosten, Drummond
- 2005
(Show Context)
Citation Context ...ing algorithm to yield a large speed increase. In addition, the approach allows the detector to be generalised, producing a suite of high-speed detectors which we currently use for real-time tracking =-=[2]-=- and AR label placement [3]. To show that speed can been obtained without necessarily sacrificing the quality of the feature detector we compare our detector, to a variety of wellknown detectors. In S... |

66 |
Robust image corner detection through curvature scale space
- Mokhtarian, Suomel
- 1998
(Show Context)
Citation Context ... rapidly. Several techniques were developed which involved detecting and chaining edges with a view to finding corners in the chained edge by analysing the chain code[18], finding maxima of curvature =-=[19,20,21]-=-, change in direction [22] or change in appearance[23]. Others avoid chaining edges and instead look for maxima of curvature[24] or change in direction [25] at places where the gradient is large. Anot... |

61 | Finding corners
- Noble, J
- 1988
(Show Context)
Citation Context ...e). Harris then defines the corner response to be C = |H| − k(traceH) 2 . (2) This is large if both eigenvalues of H are large, and it avoids explicit computation of the eigenvalues. It has been shown=-=[6]-=- that the eigenvalues are an approximate measure of the image curvature. Based on the assumption of affine image deformation, a mathematical analysis led Shi and Tomasi[7] conclude that it is better t... |

46 |
Corner detection and curve representations using cubic b-splines
- Medioni, Yasumoto
- 1987
(Show Context)
Citation Context ... rapidly. Several techniques were developed which involved detecting and chaining edges with a view to finding corners in the chained edge by analysing the chain code[18], finding maxima of curvature =-=[19,20,21]-=-, change in direction [22] or change in appearance[23]. Others avoid chaining edges and instead look for maxima of curvature[24] or change in direction [25] at places where the gradient is large. Anot... |

40 |
Real-time corner detection algorithm for motion estimation
- Wang, Brady
- 1995
(Show Context)
Citation Context ...edge by analysing the chain code[18], finding maxima of curvature [19,20,21], change in direction [22] or change in appearance[23]. Others avoid chaining edges and instead look for maxima of curvature=-=[24]-=- or change in direction [25] at places where the gradient is large. Another class of corner detectors work by examining a small patch of an image to see if it “looks” like a corner. Since second deriv... |

37 |
Digital Communications
- Sklar
- 1988
(Show Context)
Citation Context ...his kernel is symmetric, this is equivalent to matched filtering for objects with that shape. The robustness is achieved because matched filtering is optimal in the presence of additive Gaussian noise=-=[36]-=-. FAST, however, is not very robust to the presence of noise. This is to be expected: Since high speed is achieved by analysing the fewest pixels possible, the detector’s ability to average out noise ... |

35 |
Description of image surfaces
- Noble, J
- 1989
(Show Context)
Citation Context ...mation, a mathematical analysis led Shi and Tomasi[7] conclude that it is better to use the smallest eigen value of H as the corner strength function: C = min (λ1,λ2). (3) A number of suggestion have =-=[5,7,8,9]-=- been made for how to compute the corner strength from H and these have been all shown [10] to be equivalent to various matrix norms of H Zheng et al.[11] perform an analysis of the computation of H, ... |

34 |
Fast corner detection
- Trajkovic, Hedley
- 1998
(Show Context)
Citation Context ...ts surroundings. A set of rules is used to suppress qualitatively “bad” features, and then local minima of the, SUSANs, (Smallest USAN) are selected from the remaining candidates. Trajkovic and Hedley=-=[28]-=- use a similar idea: a patch is not self-similar if pixels generally look different from the centre of the patch. This is measured by considering a circle. fC is the pixel value at the centre of the c... |

27 |
Early jump-out corner detectors
- Cooper, Svetha, et al.
- 1993
(Show Context)
Citation Context ...etecting and chaining edges with a view to finding corners in the chained edge by analysing the chain code[18], finding maxima of curvature [19,20,21], change in direction [22] or change in appearance=-=[23]-=-. Others avoid chaining edges and instead look for maxima of curvature[24] or change in direction [25] at places where the gradient is large. Another class of corner detectors work by examining a smal... |

25 |
Analysis of gray level corner detection
- Zheng, wang, et al.
- 1999
(Show Context)
Citation Context ...(λ1,λ2). (3) A number of suggestion have [5,7,8,9] been made for how to compute the corner strength from H and these have been all shown [10] to be equivalent to various matrix norms of H Zheng et al.=-=[11]-=- perform an analysis of the computation of H, and find some suitable approximations which allow them to obtain a speed increase by computing only two smoothed images, instead of the three previously r... |

19 | A Fast Radial Symmetry Transform for Detecting
- Loy, Zelinsky
- 2002
(Show Context)
Citation Context ...on the response is too low, then the potential corner is rejected. To speed up the method further, this fast check is first applied at a coarse scale. A fast radial symmetry transform is developed in =-=[29]-=- to detect points. Points have a high score when the gradient is both radially symmetric, strong, and of a uniform sign along the radius. The scale can be varied by changing the size of the area which... |

16 | A condition number for point matching with application to registration and 200 postregistration error estimation
- Kenney, Manjunath, et al.
(Show Context)
Citation Context ...mation, a mathematical analysis led Shi and Tomasi[7] conclude that it is better to use the smallest eigen value of H as the corner strength function: C = min (λ1,λ2). (3) A number of suggestion have =-=[5,7,8,9]-=- been made for how to compute the corner strength from H and these have been all shown [10] to be equivalent to various matrix norms of H Zheng et al.[11] perform an analysis of the computation of H, ... |

16 | Fast Computation of Characteristic Scale using a Half-Octave Pyramid
- Crowley, Riff, et al.
- 2002
(Show Context)
Citation Context ...ed that scales are √ 2 apart,sMachine learning for high-speed corner detection 3 speeds up computation by a factor of about two, compared to the striaghtforward implementation of Gaussian convolution =-=[13]-=-. It is noted in [14] that the LoG is a particularly stable scale-space kernel. Scale-space techniques have also been combined with the Harris approach in [15] which computes Harris corners at multipl... |

12 | Real-time video annotations for augmented reality
- Rosten, Reitmayr, et al.
- 2005
(Show Context)
Citation Context ...rge speed increase. In addition, the approach allows the detector to be generalised, producing a suite of high-speed detectors which we currently use for real-time tracking [2] and AR label placement =-=[3]-=-. To show that speed can been obtained without necessarily sacrificing the quality of the feature detector we compare our detector, to a variety of wellknown detectors. In Section 3 this is done using... |

12 | A mathematical comparison of point detectors
- Zuliani, Kenney, et al.
(Show Context)
Citation Context ...lest eigen value of H as the corner strength function: C = min (λ1,λ2). (3) A number of suggestion have [5,7,8,9] been made for how to compute the corner strength from H and these have been all shown =-=[10]-=- to be equivalent to various matrix norms of H Zheng et al.[11] perform an analysis of the computation of H, and find some suitable approximations which allow them to obtain a speed increase by comput... |

9 |
Corner characterization by differential geometry techniques
- Guiducci
- 1988
(Show Context)
Citation Context ... this is that they tend to perform poorly on images with only large-scale features such as blurred images. The corner detector presented in this work belongs to this category. The method presented in =-=[26]-=- assumes that a corner resembles a blurred wedge, and finds the characteristics of the wedge (the amplitude, angle and blur) by fitting it to the local image. The idea of the wedge is generalised in [... |

9 |
Demo software: SIFT keypoint detector
- Lowe
- 2005
(Show Context)
Citation Context ...s per frame (typical numbers, probably commonly used); the results are somewhat less convincing in the other datasets, where points undergo non-projective changes. In the sample implementation of SIFT=-=[35]-=-, approximately 1000 points are generated on the images from the test sets. We concur that this a good choice for the number of features since this appears to be roughly where the repeatability curve ... |

6 |
A.: A comparison of corner detection techniques for chain coded curves
- Rutkowski, Rosenfeld
- 1978
(Show Context)
Citation Context ...s, this boundary changes direction rapidly. Several techniques were developed which involved detecting and chaining edges with a view to finding corners in the chained edge by analysing the chain code=-=[18]-=-, finding maxima of curvature [19,20,21], change in direction [22] or change in appearance[23]. Others avoid chaining edges and instead look for maxima of curvature[24] or change in direction [25] at ... |

4 |
Curve Encoding and Detection of Discontinuities
- Langridge
(Show Context)
Citation Context ... rapidly. Several techniques were developed which involved detecting and chaining edges with a view to finding corners in the chained edge by analysing the chain code[18], finding maxima of curvature =-=[19,20,21]-=-, change in direction [22] or change in appearance[23]. Others avoid chaining edges and instead look for maxima of curvature[24] or change in direction [25] at places where the gradient is large. Anot... |

2 |
A neural network based corner detection method
- Dias, Kassim, et al.
- 1995
(Show Context)
Citation Context ... see if it looks like a corner is to use machine learning to classify patches of the image as corners or non-corners. The examples used in the training set determine the type of features detected. In =-=[30]-=-, a three layer neural network is trained to recognise corners where edges meet at a multiple of 45 ◦ , near to the centre of an 8 × 8 window. This is applied to images after edge detection and thinni... |

1 |
L.G.: Computer and robot vision. Volume 1. AdisonWesley
- Haralick, Shapiro
- 1993
(Show Context)
Citation Context ...e developed which involved detecting and chaining edges with a view to finding corners in the chained edge by analysing the chain code[18], finding maxima of curvature [19,20,21], change in direction =-=[22]-=- or change in appearance[23]. Others avoid chaining edges and instead look for maxima of curvature[24] or change in direction [25] at places where the gradient is large. Another class of corner detect... |

1 |
eds.: Performenace Evaluation of Corner Detection Algorithms under Affine and Similarity Transforms
- Cootes, Taylor
(Show Context)
Citation Context ...o the Pentium III. 3 A comparison of detector repeatability Although there is a vast body of work on corner detection, there is much less on the subject of comparing detectors. Mohannah and Mokhtarian=-=[33]-=- evaluate performance by warping test images in an affine manner by a known amount. They define the ‘consistency of corner numbers’ as CCN = 100 × 1.1 −|nw−no| , where nw is the number of features in ... |