Interest point characterisation through textural analysis for rejection of bad correspondences

Pattern Recognition Letters 26 (2005) 1587–1596

www.elsevier.com/locate/patrec

Interest point characterisation through textural analysisfor rejection of bad correspondences

Viorela Ila *, Rafael Garcia, Xavier Cufi, Joan Batlle

Computer Vision and Robotics Group, University of Girona, Campus Montilivi, 17071 Girona, Spain

Received 18 November 2003; received in revised form 16 December 2004

Available online 14 April 2005

Communicated by E. Backer

Abstract

This work proposes a tool based on texture analysis to characterise incorrect point correspondences in underwater

image pairs. Interest point correspondences are first detected through region correlation, obtaining pairs of matched

points in both images. For every pair of points, their textural characteristics are computed. These textural properties

are stored in two corresponding characterisation vectors, which are then compared by means of similarity measures.

This measure can be considered as a reliable threshold for outlier rejection. Experiments with real underwater images

were carried out.

� 2005 Elsevier B.V. All rights reserved.

Keywords: Correspondence problem; Outlier rejection; Texture characterisation; Texture operators

1. Introduction

Detecting correct correspondences in a pair of

images taken at two consecutive time instants is

an important issue in computer vision. Often this

means detecting features in one image, and match-

0167-8655/$ - see front matter � 2005 Elsevier B.V. All rights reserv

doi:10.1016/j.patrec.2005.01.008

* Corresponding author.

E-mail addresses: [email protected] (V. Ila), [email protected].

es (R. Garcia), [email protected] (X. Cufi), [email protected]

(J. Batlle).

ing them in the second one. Typical features to be

matched are interest points (Sethi and Jain, 1987),straight line segments (Deriche and Faugeras,

1990) or, less frequently, image contours (Sch-

wartz and Sharir, 1986). The selection of features

may depend on the application, although points

are commonly used because they can be easily ex-

tracted and are quite robust to noise (Harris and

Stephens, 1988). However, the matching of these

features in the second image is normally a complextask. In this paper we deal with point correspon-

dences. Quite often, local gray level correlation is

applied to detect matchings in the pair of images.

ed.

mailto:[email protected]

mailto:[email protected].



1588 V. Ila et al. / Pattern Recognition Letters 26 (2005) 1587–1596

Although this technique provides good results in

standard images (Giachetti, 2000), it may lead to

detection of incorrect correspondences in under-

water sequences. Underwater images are difficult

to process due to the medium transmission proper-ties (Funk et al., 1972). These difficulties provoke a

blurring of the elements in the image with high

clutter in the regions of interest and lack of distinct

features. Automatic detection of unreliable match-

ings can be achieved by means of outlier rejection

techniques, such as LMedS (Rousseeuw and

Leroy, 1987) or RANSAC (Bolles and Fischler,

1981). However, these probabilistic algorithmsare based on random sampling and robust regres-

sion. Due to their probabilistic nature, they

may produce incorrect results, although with a

bounded error probability, but their main problem

is the high computational cost associated with the

search in the space of possible estimates generated

from the data. In all cases, the aim is to find a

set of point pairs which minimise the square sumover the residuals. Given a set of point correspon-

dences, a M-estimators robust technique could

also be applied (Zhang, 1998). The M-estimators

are based on replacing the residual squares with

a weighted function of residuals to make the esti-

mation less sensitive to outliers. Some authors

have reported a robust behaviour of M-estimators

in the presence of bad correspondence localisa-tions but not to false matchings (Xu and Zhang,

1996). Moreover, M-estimators is an iterative

algorithm with a considerable computational cost.

This article extends our previous work of

improving image correspondences (Garcia et al.,

2001) by proposing a method to characterise incor-

rect correspondences through textural analysis.

While in (Garcia et al., 2001) the texture operatorswere used for choosing the best correspondence

from a set of candidates, our new proposal uses

intensity-based techniques to detect point pairs,

and texture information to eliminate possible out-

cðp1; p2Þ ¼Pa

i¼�a

Paj¼�aðI1ðx1 þ i; y1 þ jÞ � I1ðx1; y1ÞÞ � ðI

n2ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffir2ðI1Þ � r2ðI2Þ

p

liers. In this new approach the outlier rejection

process is independent of the matching problem.

Previous work in this topic is an extensive test of

different texture operators to improve the robust-

ness of the correspondence problem. On the otherhand, in this new paper we take advantage of the

previous results to try out a different approach

for outlier rejection. These techniques provide

the basis of a new methodology to improve point

correspondences with a reduced computational

burden.

The remainder of the paper is structured as fol-

lows: Section 2 details the proposed methodologyfor feature characterisation. Then, comparative

results to demonstrate the validity of the described

approach are presented in Section 3. Finally, Sec-

tion 4 outlines conclusions and future work.

2. Feature characterisation through texture analysis

This proposed approach performs characterisa-

tion of interest points based on texture analysis.

Consider a pair of underwater images (I1, I2). The

first step to solve the correspondence problems is

the detection of a set of well-contrasted points in

the first image. Based on previous experiments

on underwater images, Harris–Stephens (Harris

and Stephens, 1988) corner detector was chosenfor this implementation. For each interest point

p1 = (x1,y1) of the first image, a normalised corre-

lation (Giachetti, 2000) provides corresponding

matching p2 = (x2,y2) in the second image. Given

a point p1, the correlation score c(p1,p2) is com-

puted for all the points in the second image p2which are located at the neighbourhood of the

coordinates defined by p1 (Eq. (1)). However,experimental results show that the correlation

technique described in this equation may produce

incorrect correspondences in underwater imaging

(Garcia et al., 2001).

2ðx2 þ i; y2 þ jÞ � I2ðx2; y2ÞÞ ð1Þ

V. Ila et al. / Pattern Recognition Letters 26 (2005) 1587–1596 1589

where n · n is the size of the correlation window,

a = (n � 1)/2, I1ðx1; y1Þ and I2ðx2; y2Þ are the ave-

rage intensity of both correlation windows and

r2(Æ) is the variance as defined in Eq. (2).

r2ðIÞ ¼Pa

i¼�a

Paj¼�aðIðxþ i; y þ jÞ2Þ

n2� Iðx; yÞ2:

ð2Þ

2.1. Textural characterisation

Different techniques for solving the matching

ambiguities can be found in the literature (Zhang

et al., 1995; Bolles and Fischler, 1981). Our previ-

ous work proposed a method to improve the

robustness of correspondence points in two con-

secutive images. In that work (Garcia et al.,

2001), several configurations of a number of tex-ture operators were tested: energy filters (Laws,

1980), co-occurrence matrix (Haralick et al.,

1973), Local Binary Patterns (Ojala and Pietikai-

nen, 1999) along with a combination of these.

Based on this previous study, the present work

uses texture information to characterise bad corre-

spondences among a set of point pairs in two

images. Laws texture energy filters were tested inour approach, being the ones that worked better

in the previous work. These operators come from

a computation of different statistical measures:

Absolute mean, Standard Deviation, Positive

Mean and Negative Mean over a pre-filtered

image. Filters are based on 1 · 3 or 1 · 5 vectors,

namely: Level, Edge and Spot.

L3 ¼ ½ 1 2 1 �; L5 ¼ ½ 1 4 6 4 1 �E3 ¼ ½�1 0 1 �; E5 ¼ ½�1 �2 0 2 1 �S3 ¼ ½�1 2 �1 �; S5 ¼ ½�1 0 2 0 �1 �:

ð3ÞConsider P interest points kp1, and k = 1, . . . ,P,

defined in I1. Texture-based characterisation is

performed by considering an m · m window in

their neighbourhood. For every kth point, a tex-

ture operator is then computed in its neighbour-hood. The same operation is performed for its

corresponding matching kp2 in I2. In this way,

two characterisation vectors kv1 and kv2 of the

kth interest point and its matching are obtained:

kv1 ¼ kv11; kv12; . . . ; kv1N� �

kv2 ¼ kv21; kv22; . . . ; kv2N� � ð4Þ

where N ¼ ðm�1s þ 1Þ � ðm�1

s þ 1Þ is the size of the

characterisation vector, and s is the value of the

subsampling of the characterisation window.

2.2. Comparing textural characterisation of

interest points and their correspondent matching

The characterisation vector of every point is

compared with the characterisation vector of

its correspondence point in the second image.

Thus, we are searching for similarities in terms

of texture. Normalised correlation was used to

measure similarity between these vectors: kv1and kv2:

cðkv1; kv2Þ ¼PN

i¼1kv1i � kv1� �

� kv2i � kv2� �

N �ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffir2 kv1ð Þ � r2ðkv2Þ

p ð5Þ

where kv1 and kv2 are the average value of vectorskv1,

kv2, respectively, and r2(Æ) defines the variance.This similarity measure can be used to detect

bad correspondences. A set of experiments were

carried out in order to select the best texture

operators to be applied for outlier rejection inunderwater images.

3. Experimental results

3.1. Experimental methodology

Our experimental methodology has two parts.

The first part consists of selecting the best tex-

ture operators from a list of 20 energy filters.The second step is applying every operator and

a combination of them to different underwater

sequences and comparing the results with robust

methods.

Consider a pair of underwater images, a set of

point-matchings was obtained through normalised

correlation algorithm, as defined in Eq. (1). As

part of our experimental methodology, we markedall the visually incorrect matches, thus providing a

list of bad correspondences. Human experts select

the incorrect matchings from the set of pair point-

matching. Since presence of human factor may

1 http://eia.udg.es/~viorela/download_images.htm2 The copyright is owned by ISR-IST.


introduce errors, five different persons performed

this operation. When the pairs of points are repre-

sented like in Fig. 3, by drawing the motion vector,

incorrect matchings can be easily detected being

the vectors having different orientation comparingto the dominant orientation. In some cases, excep-

tions appears due to underwater conditions such

as presence of moving objects (fishes, algae) or

nonuniform see floor (3D aspects), etc., the motion

vector of some points can have different orienta-

tion regarding the dominant orientation but the

correspondence between points may be correct.

Every pair of points and their correspondingmatchings are textural-characterised. The similar-

ity value between texture characterisation vectors

for every point and its matching is analysed. We

chose the highest value corresponding to manually

detected bad correlated point as the threshold.

That means that, when sorting all the points

according to the similarity value, all the outliers

lay on the right side of the threshold and on theleft side we can guarantee all absence of bad corre-

spondences, as can be seen in Fig. 1. The threshold

should ensure the rejection of bad correspondences

while maintaining a great number of good corre-

spondences. This threshold is used to automati-

cally reject all the outliers at the price of loosing

a number of correct matchings. This proposed

technique looks for the texture characterisationmethods which provide a higher separation be-

tween correct and false correspondences. Several

experiments were carried out to find the best tex-

ture operator.

In the second part of our experiment the pairs

point-matchings from several sequences, the

majority differing from those used in the first part

of the experiment, were characterised using thebest texture operator. In this experiment the false

correspondences were unknown. We assumed that

the whole sequence can have a fraction � of possi-ble outliers. The new threshold was computed like

in Fig. 2 and applied to the sequences under test.

This threshold is a function of � and the mean

value of the similarity measure between texture

characterisation vectors. As we can see in the fig-ure, even if the number of outliers is bigger than

the estimated � the new threshold eliminate them

all. The new threshold is computed by finding a

mean value between the average error measure-

ments and the one corresponding to �.

3.2. Experimental results

A set of experiments were performed during this

work. The proposed approach was tested with 15

underwater image sequences. These images were

selected because they are a good representation

of the possible conditions found in underwater

environments, such as: blurring, lack of well-

defined contours, bad visibility, low contrast, scat-

tering effects, nonuniform illumination, lightingartifacts generated by the waves, etc. Moreover,

different scenarios were selected: rocky seafloor,

sand and algae seafloor, moving fishes, as well as

some man-made objects like an old submerged

chain, etc.

Some of the underwater image sequences were

acquired by the URIS Underwater Robot built

at the University of Girona. Two of the sequencesused in our experiments are available on the web1.

One of the sequences (Fig. 2(c) and (d)) includes

997 images which have been taken by the URIS

underwater robot at Costa Brava (Spain), next to

Palamos. The robot was navigating at a depth of

2 m, and its altitude was approximately 3 m over

the sea floor. The images from Fig. 2(a) and (b)

were acquired on a different dive and using a col-our camera. The last pair of images presented in

this paper is part of a sequence of images acquired

at Platja d�Aro (Costa Brava). These are colour

images from a sequence where the robot is follow-

ing a submerged chain. Some other images used in

these experiments were acquired at Nice (France)2

in September 2000 at 6 m depth using a PAL

camera.For every image sequence, a set of pairs point-

matching was obtained as described in Section 2,

considering a correlation window size of 25 · 25

and a search window of 61 · 61, according to the

possible displacement between two frames. Fig. 3

(left) shows three of the tested underwater se-

quences and the point-matching obtained through

http://eia.udg.es/~viorela/download_images.htm

0 50 100 150–1

–0.8

–0.6

–0.4

–0.2

0

0.2

0.4

0.6

0.8

1

points

corr

elat

ion

mea

sure

outliers threshold0 50 100 150

–1

–0.8

–0.6

–0.4

–0.2

0

0.2

0.4

0.6

0.8

1

corr

elat

ion

mea

sure

points outliers(a) (b)

0 50 100 150–1

–0.8

–0.6

–0.4

–0.2

0

0.2

0.4

0.6

0.8

1

points

corr

elat

ion

mea

sure

thresholdoutliers0 50 100 150

–1

–0.8

–0.6

–0.4

–0.2

0

0.2

0.4

0.6

0.8

1

points

corr

elat

ion

mea

sure

outliers(c) (d)

0 50 100 150–1

–0.8

–0.6

–0.4

–0.2

0

0.2

0.4

0.6

0.8

1

points

corr

elat

ion

mea

sure

outliersthreshold

0 50 100 150–1

–0.8

–0.6

–0.4

–0.2

0

0.2

0.4

0.6

0.8

1

points

corr

elat

ion

mea

sure

outliers(e) (f)

Fig. 1. Outliers distribution when computing the similarity measure between characterisation vectors for all 150 interest points

corresponding to the images from Fig. 3, left side. (a), (c), (e): sorted and (b), (d), (f): unsorted according to their correlation score.


region-based correlation. Since texture is a prop-

erty of regions, the textural characterisation of

the interest points depends on the size of the neigh-

bourhood taken into consideration. For every tex-

ture operator, four different neighbourhoods of

size m · m were considered: 11 · 11, 15 · 15,

19 · 19 and 25 · 25. The neighbourhoods size does

not depend on the image resolution. Moreover,

within a neighbourhood, there is a high degree

of redundancy when texture characterisation is

Table 1

Size of the characterisation vector according to different

neighbourhood sizes and their subsampling

Samples Window size

11 · 11 15 · 15 19 · 19 25 · 25

1 121 225 361 625

2 36 64 81 169

3 9 9 9 9

0 50 100 150–1

–0.8

–0.6

–0.4

–0.2

0

0.2

0.4

0.6

0.8

1

POINTS

SIM

ILA

RIT

Y M

EA

SU

RE

Sequence 8 Method L5S5_1_2_w25

threshold

median

E

Fig. 2. Threshold for rejecting the outliers.


applied to every pixel. By choosing different sub-

sampling strategies, our approach will verify if

redundancy introduces robustness in the system.

Three significative subsampling modalities were se-

lected: considering every point: s = 1, every two

points: s = 2 or only nine points of the neighbour-hood: N = 9 (see Section 2.2). Table 1 shows the

corresponding size of the characterisation vector

depending on both the selected neighbourhood

and its subsampling strategy. For every pair of

images, we tested 20 texture operators, every oper-

ator in four different modalities dependent on the

size of the characterisation window, and every

window having three different subsamplings. Thefollowing Laws filters, resulting from combina-

tions of (3) basic vectors, were applied to the

image: L3L3, E3E3, L5S5, E5S5, E5L5. These fil-

ters are detailed below (6).

L3L3 ¼ 1 4 6 4 1½ �;E3E3 ¼ 1 0 �2 0 1½ �;

L5S5 ¼

�1 0 2 0 �1

�4 0 8 0 �4

�6 0 12 0 �6

�4 0 8 0 �4

�1 0 2 0 �1

26666664

37777775;

E5S5 ¼

�1 0 2 0 �1

�2 0 4 0 �2

0 0 0 0 0

2 0 �4 0 2

1 0 �2 0 1

26666664

37777775;

E5L5 ¼

�1 �4 �6 �4 �1

�2 �8 �12 �8 �2

0 0 0 0 0

2 8 12 8 2

1 4 6 4 1

26666664

37777775:

ð6Þ

Four statistical measures were considered: Abso-

lute Mean, Standard Deviation, Positive Mean and

Negative Mean. The goal was to find the most ade-

quate textural operators to solve the correspon-dence problem. Operators were compared in terms

of number of rejected outliers and number of surviv-

ing good correspondences. By observing the distribu-

tion of outliers when representing the similarity

measure between characterisation vectors, we can

select the highest value, giving rise to a to bad corre-

spondence as a threshold. By rejecting all the points

having the similarity value lower than this thresh-old, we can analyse the number of remaining points

in order to see the efficiency of this method, see Fig.

1(a), (c) and (e). In order to evaluate the results, all

pairs point-matching obtained using normalised

correlation were visually verified and classified in

good and bad correspondences. Even if an error is

present in the analysis provided by the human

experts, there is a small probability that this errorcoincides with the ‘‘worst case’’ which determines

the threshold.

Fifteen sequences were tested. For every se-

quence, 20 texture operators were applied using

four window sizes and three subsamplings. Fig. 4

shows, for every window size, the average percent-

Fig. 3. Selected images from the 15 tested sequences: (a) strong shading effects produced by the waves; (c) rotation; (e) scattering

produces blurring in the image. Detection of correspondences in two consecutive images using intensity-based techniques. (b), (d), (f)

corresponding pairs of points after outlier rejection using texture characterisation.


age of good correspondences after rejecting the

outliers using the texture information. As we ex-

pected, the best percentage of correct correspon-

dences after eliminating the outliers corresponds


to the 19 · 19 and 25 · 25 neighbourhood size. It

can also be observed that a moderate subsampling

(considering every two points) provides similar

results to processing every point. On the other

hand, higher subsampling drastically decreases per-formance. Six texture operators providing the best

performance can be selected according to Fig. 4:

1. mask=L5S5, Absolute Mean;

2. mask=E3E3, Absolute Mean;

3. mask=E3E3, Positive Mean;

Am Sd Pm Nm Am Sd Pm Nm Am Sd Pm Nm Am Sd Pm Nm Am Sd Pm Nm0

10

20

30

40

50

60

70

80

90

100L3L3 E3E3 L5S5 E5S5 E5L5

METHOD

AV

ER

AG

E P

ER

CE

NT

AG

E

subsam1subsam2subsam3

(a) (


10

20

30

40

50

60

70

80

90

100L3L3 E3E3 L5S5 E5S5 E5L5

METHOD

AV

ER

AG

E P

ER

CE

NT

AG

E


(c) (

Fig. 4. Average percentage of remaining correct correspondences aft

method. (a) Neighbourhood size 11 · 11. (b) Neighbourhood size 15

25 · 25.

4. mask=L5S5, Negative Mean;

5. mask=L5S5, Positive Mean;

6. mask=E5S5, Negative Mean.

Looking for the possibility of improving the re-sults, combinations of these six operators have

also been tested and compared with single opera-

tor performance. Different ways of combining tex-

ture characterisation were tested: summing the

similarity measure obtained using two texture

operators or applying their minimum value. Fif-


10

20

30

40

50

60

70

80

90

100L3L3 E3E3 L5S5 E5S5 E5L5

METHOD

AV

ER

AG

E P

ER

CE

NT

AG

E


b)


10

20

30

40

50

60

70

80

90

100L3L3 E3E3 L5S5 E5S5 E5L5

METHOD

AV

ER

AG

E P

ER

CE

NT

AG

E


d)

er eliminating all the outliers using the proposed texture based

· 15. (c) Neighbourhood size 19 · 19. (d) Neighbourhood size

Table 2

Average percentage of correct correspondences after eliminating all the outliers using the six proposed texture operators and their

combination

SumnOneOp.nMin L5S5_Am E3E3_Am E3E3_Pm L5S5_Nm L5S5_Pm E5S5_Nm

L5S5_Am 96.0722 95.3538 96.1968 96.2814 95.2629 93.4897

E3E3_Am 96.0989 95.5880 96.3751 96.0911 95.2749 93.2692

E3E3_Pm 96.6912 96.3472 95.3515 95.3277 96.0041 94.8135

L5S5_Nm 96.4366 96.5661 95.7628 94.2817 96.0194 94.4341

L5S5_Pm 95.8195 95.6658 96.4316 96.1840 93.4180 92.9625

E5S5_Nm 95.9136 95.5415 95.9599 96.3436 95.1490 93.0360

25 · 25 Characterisation window and every two pixel subsampling were applied. Under diagonal—sum of correspondence measures

between characterisation vectors; above diagonal—minimum of correspondence measures between characterisation vectors; diagonal—

apply single textural operator.


teen combinations were tested using a 25 · 25 win-

dow sampled every two pixels. The results are

shown in Table 2: Data under the diagonal are

the average percentage when the texture operatorswere combined by summing the similarity values;

the diagonal contains the results in case of apply-

ing a single operator and above the diagonal, the

minimum similarity value to reject the outliers. It

can be observed that better results were obtained

following the summing approach. But the differ-

ences between these three approaches are quite

small, therefore, only one textural operator willbe considered for reducing computation time.

The second step in our experimental methodol-

ogy is applying the best textural operators for out-

lier rejection and comparing the results with robust

methods. One important issue is to find a thres-

hold to ensure the rejection of all outliers. In this

experiment we observed that by sorting the pairs

point-matching according to the value of theircorrelation measurement between texture charac-

terisation vectors and applying the threshold

proposed in Section 3.1 we obtained satisfactory

results. This strategy was applied to other 15

sequences, different from those used in the first

experiment. Fig. 3(b), (d) and (f) shows the results

in the case of applying texture based strategy. Due

to space limitation only three sequences are illus-trated in this paper. Results from all the sequences

are available on the web.3

3 http://eia.udg.es/~viorela/image_texture_chr.htm

4. Conclusions and future work

In this paper we have presented an analysis of

using different texture operators for feature char-acterisation. Tests on real underwater images show

an adequate characterisation of the incorrect cor-

respondences. We selected different sequences of

underwater images in order to test how the algo-

rithm can tackle underwater imaging problems.

Texture provides a rich source of information for

feature characterisation, this being an essential

characteristic when having difficulties in success-fully using standard techniques. The robustness

of the proposed technique is based on the exploita-

tion of gray level information complemented by

texture cues. Twenty texture operators were tested,

and six best combinations of energy filters and sta-

tistical measures were chosen taking into consider-

ation the number of surviving inliers after outliers

rejection.A threshold which take into consideration an

estimated numbers of outliers corresponding to a

sequence of underwater images and the mean

value of the similarity measure between texture

characterisation vectors proved to work good in

bad correspondences rejection for a large number

of tested sequences. When the algorithm runs on

a real system performing a real mission, we canconsider quite similar conditions for the whole

set of images. In this case, by analysing a set of

some images from the sequence we can choose a

value for the � factor which can be applied to the

whole sequence. This new approach outperforms

random sampling techniques like LMedS or RAN-

SAC, in terms of computational cost. On the other

http://eia.udg.es/~viorela/image_texture_chr.htm


hand, due to their probabilistic nature, robust

methods may produce incorrect results. An impor-

tant observation is that robust methods eliminate

all the points with different displacement compar-

ing to the majority of points, while texture basedtechniques eliminate only bad correspondences.

While robust methods need a forehand estimation

of either fundamental matrix, in case of 3D scene,

or homography in case of planar scene, the pre-

sented new texture-based approach does not re-

quire any a priori information. Presently, we are

carrying out hardware implementation of a regis-

tration algorithm. Thus, this new approach canbe helpful due to its facility of parallelization.

References

Bolles, R., Fischler, M., 1981. A ransac-based approach to

model fitting and its application to finding cylinders in range

data. In: 6th Internat. Joint Conf. on Artificial Intelligence,

Vancouver, Canada, pp. 637–643.

Deriche, R., Faugeras, O., 1990. Tracking line segments. In:

First European Conf. on Computer Vision. pp. 259–268.

Funk, C.J., Bryant, S., Beckman Jr., P., 1972. Handbook of

underwater imaging system design. Technical report, Ocean

Technology Department, Naval Undersea Center.

Garcia, R., Cufi, X., Batlle, J., 2001. Detection of matchings in

a sequence of underwater images through texture analysis.

In: Internat. Conf. on Image Process. pp. 361–364.

Giachetti, A., 2000. Matching techniques to compute image

motion. Image and Vision Computing (18), 247–260.

Haralick, R., Shanguman, K., Dinstein, I., 1973. Textural

features for image classification. IEEE Trans. Systems Man

Cybernet. (November), 610–621.

Harris, C., Stephens, M., 1988. A combined corner and edge

detector. In: Fourth Alvey Vision Conf., Manchester, pp.

147–151.

Laws, K., 1980. Textured image segmentation. Ph.D. thesis,

Processing Institute, University of Southern California, LA.

Ojala, T., Pietikainen, M., 1999. Unsupervised texture segmen-

tation using feature distributions. Pattern Recogn. 32, 477–

486.

Rousseeuw, P.J., Leroy, A.M., 1987. Robust Regression and

Outlier Detection. John Wiley, New York.

Schwartz, J., Sharir, M., 1986. Identification of partially

occluded objects in two and three dimensions by

matching noisy characteristic curves. Int. J. Robot. Res. 6,

29–43.

Sethi, I.K., Jain, R., 1987. Finding trajectories of feature points

in a monocular image sequence. IEEE Trans. Pattern Anal.

Mach. Intell. 9, 56–73.

Xu, G., Zhang, Z., 1996. Computational imaging and vision.

In: Epipolar Geometry in Stereo, Motion and Object

Recognition: A Unified Approach, vol. 6. Springer.

Zhang, Z., 1998. Determining the epipolar geometry and its

uncertainty: A review. Int. J. Comput. Vision 27 (2), 161–

198.

Zhang, Z., Deriche, R., Faugeras, O.D., Luong, Q.-T., 1995. A

robust technique for matching two uncalibrated images

through the recovery of the unknown epipolar geometry.

Artificial Intell. 78 (1–2), 87–119, Available from:

<citeseer.nj.nec.com/article/zhang94robust.html> .

http://citeseer.nj.nec.com/article/zhang94robust.html

Documents

Interest point characterisation through textural analysis for rejection of bad correspondences