POSTER SESSION BOOKLET - iplab.dmi.unict.it · latent topic model based on nonnegative matrix...

POSTER SESSION BOOKLET

http://www.dmi.unict.it/icvss

University of Catania - University of Cambridge

International Computer Vision Summer School 2011

Registration, Recognition and Reconstruction in Images and Video

Sicily, 11-16 July 2011


http://iplab.dmi.unict.it/

http://mi.eng.cam.ac.uk/research/vision/

International Computer Vision Summer School

Computer vision is the science and technology of making machines that see. It

is concerned with the theory, design and implementation of algorithms that can

automatically process visual data to recognize objects, track and recover their

shape and spatial layout.

The International Computer Vision Summer School - ICVSS was established

in 2007 to provide both an objective and clear overview and an in-depth analysis

of the state-of-the-art research in Computer Vision. The courses are delivered by

world renowned experts in the field, from both academia and industry, and cover

both theoretical and practical aspects of real Computer Vision problems.

The school is organized every year by University of Cambridge (Computer

Vision and Robotics Group) and University of Catania (Image Processing Lab).

The general entry point for past and future ICVSS editions is:


ICVSS Poster Session

The International Computer Vision Summer School is especially aimed to provide

a stimulating space for young researchers and Ph.D. Students. Participants have

the possibility to present the results of their research, and to interact with their

scientific peers, in a friendly and constructive environment.

This booklet contains the abstract of the posters accepted to ICVSS 2011.

Best Presentation Prize

A subset of the submitted posters is selected by the school committee for short

oral presentation. A best presentation prize is given to the best presentation

selected by the school committee.

Scholarship

A scholarship is awarded to the best PhD student attending the school. The

decision is made by the School Committee at the time of the School, taking into

account candidates’cv, poster and oral presentation.

Sicily, May 2011 Roberto Cipolla

Sebastiano Battiato

Giovanni Maria Farinella

i


List of Posters1

1. HIDING FINGERPRINT INFORMATION IN FACE IMAGES

Ahmadian P., Rahmati M.

2. FINE-GRAINED IMAGE CLASSIFICATION USING FISHER VECTORS

Akata Z., Sanchez J., Perronnin F.

3. GRAPHTRACK: FAST AND OPTIMAL TRACKING

Amberg B.

4. MULTI-TARGET TRACKING BY CONTINUOUS ENERGY MINIMIZATION

Andriyenko A., Schindler K.,

5. TRAINABLE V4-LIKE FILTERS FOR DETECTING RETINAL VASCULAR

BIFURCATIONS

Azzopardi G., Petkov N.

6. ONLINE APPEARANCE MODELS FOR VISUAL TRACKING

Bachoo AK., Nicolls F.

7. MAKING INTELLIGENT TUTORS EMOTIONALLY AWARE

Banda N., Robinson P.

8. MULTI-POSE FACE RECOGNITION FOR PERSON RETRIEVAL IN CAMERA

NETWORKS

Bouml M., Bernardin K., Ekenel H., Stiefelhagen R.

9. AUTOMATED CATEGORIZATION OF ABNORMALITIES BASED ON COMPUTER

VISION ANALYSIS OF LARGE MEDICAL IMAGING COLLECTIONS

Burner A., Donner R., Mayerhoefer M., Kainberger F., Langs G.

10. FEATURE EXTRACTION FOR NON-TEXTURED OBJECT RECOGNITION

USING A STEREO CAMERA

Byeon W.

11. FLEXIBLE INTELLIGENT VISUAL SURVEILLANCE SYSTEM

Chang H. J., Yi K. M., Yin S., Kim S. W., Choi J. Y.

12. TOWARDS THE EXAGGERATED IMAGE STEREOTYPES

Chen C., Lauze F., Igel C., Feragen A., Loog M., Nielsen M.

13. UNCOOPERATIVE 3D FACE RECOGNITION FOR INTELLIGENT SURVEILLANCE

APPLICATIONS

Cheng X., Lakemond R., Fookes C., Sridharan S.

1Posters are ordered by surname of first author. Each poster is identified by a number. Thepage of a poster in this booklet corresponds with the ID of the poster.

ii

14. AUTOMATIC ANNOTATION OF HISTOPATHOLOGICAL IMAGES USING A

LATENT TOPIC MODEL BASED ON NONNEGATIVE MATRIX FACTORIZATION

Cruz-Roa A., Dıaz G., Gonzılez F.

15. A PROBABILISTIC GENERATIVE APPROACH TO INVARIANT VISUAL

INFERENCE AND LEARNING

Dai Z., Lucke J.

16. EVENT TYPE CLASSIFICATION FOR PERSONAL PHOTO ALBUMS BASED ON

SIGNATURE IMAGE

Dang-Nguyen D.-T., Dao M.-S., Boato G., DeNatale F.

17. CYLINDRICAL PANORAMA MATCHING

De Carufel J.-L., Laganiere R.

18. AUGMENTED PERCEPTION AND INTERACTION WITH HANDHELD DEVICES

De Tommaso D., Calinon S.

19. NON-PARAMETRIC SUB-PIXEL LOCAL POINT SPREAD FUNCTION

ESTIMATION

Delbracio M., Muse P., Almansa A., Morel JM.

20. A METHOD FOR NOISY IRIS SEGMENTATION

Donida Labati R., Scotti F.

21. A SEEDED FRAMEWORK FOR TRACKING OF GENE EXPRESSION DYNAMICS

IN INDIVIDUAL CELL COMPARTMENTS

Du C.-J., Marcello M., Spiller D.G., White M.R.H., Bretschneider T.

22. SPATIAL INVARIANCE FOR SCENE CLASSIFICATION

Dunlop H.

23. A SIMPLE INPAINTING METHOD AND ITS GPU IMPLEMENTATION

Fassold H.

24. CLASSIFICATION OF THE ACROSOME INTEGRITY OF BOAR SPERMATOZOA

HEADS USING SURF AGAINST TRADITIONAL DESCRIPTORS

Fernandez-Robles L., Alegre E.

25. UNSER FEATURES ESTIMATION FOR REAL-TIME TISSUE

CHARACTERIZATION

Galluzzo F., Testoni N.

26. WORD RECOGNITION METHODS THROUGH WORD SHAPE CODING

TECHNICS COMPARISON

Garcıa-Ordas M.T., Alegre E.

27. EVALUATION AND IMPROVEMENT OF ADAPTIVE FILTERS FOR SHARPNESS

ENHANCEMENT AND NOISE REMOVAL

Garcıa-Olalla O., Alegre E.

iii

28. NOISE REDUCTION USING BIO-INSPIRED AND SOFT COMPUTING

Gonzalez Jaime L.

29. ENHANCED SUPER-RESOLUTION FOR MEDICAL DIAGNOSIS

Gonzalez Villanueva L., M. Callico G., Tobajas, F.

30. HIERARCHICAL ARM-HAND GESTURES MODELLING AND RECOGNITION

Gori I., Fanello S.R., Demiris Y.

31. DISTRIBUTED FRAMEWORK FOR MULTI-CAMERA PEOPLE TRACKING

Gruenwedel S., Nino J., Jelaca V., Philips W.

32. EMOTION ANALYSIS ON HUMAN FACE THROUGH THERMAL IMAGING

Guler P., Dumlu Seda N.

33. A PIPELINE FOR MODELING URBAN STREET FACADES FROM TERRESTRIAL

LASER AND IMAGE DATA

Hammoudi K., Dornaika F., Soheilian B., Paparoditis N.

34. ESTIMATION ERROR ANALYSIS IN STEREO COLOR MAPPING

Hasan S.F., Stauder J., Tremeau A.

35. RIEMANNIAN PERONA-MALIK DIFFUSION FOR ORIENTATION DISTRIBUTION

FUNCTION IMAGES

Krajsek K., Heinemann C., Scharr H.

36. LOGATOM RECOGNIZABILITY OF FINGER ALPHABET IN VIDEO

Heribanova P., Polec J.

37. USING HIGH LEVEL INFORMATION FOR LOW LEVEL TRACKING

Horbert E., Mitzel D., Leibe B.

38. EXPLORING PHOTOBIOS

Kemelmacher-Shlizerman I., Shechtman E., Garg R., Seitz S.M.

39. LEARNING FACE RECOGNITION IN VIDEOS FROM ASSOCIATED

INFORMATION SOURCES

Kostinger M.

40. ONLINE DISCRIMINATIVE DICTIONARY LEARNING FOR IMAGE

CLASSIFICATION

Kong S., Wang D.

41. ANALYSING COMPLEX ACTIVITIES IN VIDEO SEQUENCES

Kuehne H., Gehrig D.

42. EYE-TRACKING GUIDANCE FOR EXOSKELETON REHABILITATION

Loconsole C.

iv

43. AUTOMATIC ABDOMINAL ORGANS SEGMENTATION WITH MATHEMATICAL

MORPHOLOGY IN MR IMAGES

Lopez-Mir F., Naranjo V., Angulo J., Alcaniz M.

44. HIERARCHICAL SHAPE MODEL FOR OBJECT DETECTION

Macak J., Drbohlav O.

45. COMBINING 2D AND 3D OBJECT CATEGORIZATION FOR TASK CONSTRAINED

GRASPING

Madry M., Song D., Kragic D.

46. INTERPRETING HAND-OVER-FACE GESTURES

Mahmoud M. , Baltrusaitis T., Robinson P.

47. STEREO PERFORMANCE FOR CLUTTERED SCENES

Mannan F., Langer M. S.

48. VISUAL SERVOING FOR MICRO AERIAL VEHICLES

Maurer M., Katusic M., Bischof H.

49. INTER-MODALITY REGISTRATION TO GUIDE CARDIAC PROCEDURES

McManigle J.E., Arai A., Noble J.A.

50. CAMERA CALIBRATION AND IMAGE DISTORTION CORRECTION

FOR SUPERIOR VISUALIZATION IN MEDICAL ENDOSCOPY

Melo R., Barreto J, Falcao G.

51. REDUCING THE PROBLEM OF OCCLUSIONS IN LASER-TRIANGULATION

RECONSTRUCTION

Munaro M., Michieletto S., Menegatti E.

52. SYMBOL RETRIEVAL BY IDENTIFYING REPEATING PATTERNS

Nayef N., Breuel T.

53. HUMAN SHAPE AND POSE RECOVERY FROM MONOCULAR IMAGES USING

STATISTICAL MODELS

Neophytou A., Guillemaut J.-Y.,Hilton A.

54. REAL-TIME OBJECT DETECTION FOR MULTI-CAMERA SURVEILLANCE

Nino J., Pizurica A., Philips W.

55. EYE REFLECTION ANALYSIS AND APPLICATIONS

Nitschke C., Nakazawa A., Takemura H.

56. FACE RECOGNITION FROM SINGLE SAMPLE

Omelina L.

57. RECONSTRUCTING A 3D TRAJECTORY UNDER PERSPECTIVE PROJECTION

Park H.S., Shiratori T., Matthews I., Sheikh Y.

v

58. AUTOMATIC LATENT FINGERPRINT MATCHING

Paulino A. A., Jain A. K., Feng J.

59. UNMANNED AERIAL SYSTEMS FOR WILDLIFE MONITORING

Perez F., Albo C., Viguria A., Ollero A.

60. ROBUST PEOPLE DETECTION BASED ON APPEARANCE AND SHAPE

Pishchulin L., Jain A., Wojek C., Andriluka M., Thormaehlen T., Schiele B.

61. SUPER-RESOLUTION OF BRAIN MR IMAGES: A SPARSE REPRESENTATION

APPROACH

Rueda A., Malpica N., Romero E.

62. DEPTH ACCURACY GAINS FOR 3D RECONSTRUCTION FROM MULTIPLE VIEWS

Rumpler M., Irschara A., Bischof H.

63. PARALLELISING BUNDLE ADJUSTMENT

Salas-Moreno R.

64. SPATIO-TEMPORAL REGISTRATION FOR DIGITAL ROBUST WATERMARKING

IN VIDEOS

Schaber P., Kopf S., Effelsberg W.

65. BODY HEIGHT ESTIMATION USING A SINGLE CAMERA

Scharfenberger C., Chakraborty S., Faerber G.

66. DETECTION OF PEOPLE AND MOTION ANALYSIS IN AERIAL IMAGE SEQUENCES

Schmidt F.

67. A COMPUTATIONAL METHOD FOR QUANTITATIVE ASSESSMENT OF THE

MOVEMENT OF THE THORACIC AORTA

Schwartz E., Holfeld J., Czerny M., Langs G.

68. FROM LDA TO VISION VIA POPULATION STRUCTURE

Sharmanska V., Lampert C.H.

69. THE IMPORTANCE OF SHAPE TO ECHOCARDIOGRAM SEGMENTATION

Stebbing R.

70. DYNAMIC TEXTURE PREDICTION FOR H.264/AVC INTER CODING

Stojanovic A., Ohm J.-R.

71. A FREE-VIEWPOINT VIRTUAL MIRROR WITH MARKER-LESS USER

INTERACTION

Straka M., Hauswiesner S., Ruther M., Bischof H.

72. FUNCTIONAL ANALYSIS OF NEURAL NETWORKS: NEURONS SEGMENTATION

AND ACTIVATION MAPS

Ullo S., Del Bue A., Murino V., Maccione A., Berdondini L.

vi

73. A PORTABLE LOW VISION AID

Urena R., Morillas C., Pelayo F.

74. RANDOM DEPTH: THREE FOR THE PRICE OF ONE

Velasco-Forero S.

75. MACHINE LEARNING FOR TARGET DETECTION

Vink J.P.

76. CAMERA-BASED ANALYSIS OF ROTARY KILNS

Waibel P., Matthes J., Keller H.B.

77. INVESTIGATIONS ON ACTIVE SENSORS FOR COMPUTER VISION

Weinmann M.

78. VISUAL LANDMARK-BASED OUTDOOR LOCALIZATION FOR MAVS

Wendel A., Irschara A., Bischof H.

79. STRUCTURE BASED MOSAICKING OF AERIAL IMAGES FROM LOW

ALTITUDE OF NON-PLANAR SCENES

Wischounig-Strucl D., Quartisch M., Rinner B.

80. FACADE EXTRACTION FROM OBLIQUE AIRBORNE IMAGES

Xiao J., Gerke M., Vosselman G.

81. TRACKING VIA LOCAL PATCHES AND HIERARCHICAL SAMPLING

Yi K. M., Kim S. W., Jeong H., Choi J. Y

82. LATENT FINGERPRINT ENHANCEMENT

Yoon S., Jain A. K.

vii

HIDING FINGERPRINT INFORMATION IN

FACE IMAGES

Ahmadian P., Rahmati M.

Abstract: With the wide spread of biometric identification systems, establishing the au-

thenticity of biometric data has emerged as an important issue. This work presents a water-

marking technique which can hide fingerprint information in face image without any noticeable

damage to image. Thus, the reliability of identification results increases and if for any reason

one biometric feature fails the other one can be used. The method is also robust to blurring,

rotation, JEPG compression and cropping attacks.

Contact: [email protected]

Presentation Type: Poster

Date: Monday 11 July 2011

Time: 17:20 - 19:00

Room: PS1

1

FINE-GRAINED IMAGE CLASSIFICATION US-

ING FISHER VECTORS

Akata Z., Sanchez J., Perronnin F.

Abstract: -Fine-grained visual classification (FGVC) aims at the fine distinction of spe-

cific image categories (e.g fungus) -We motivate Fisher Kernel framework for FGVC and show

experimentally that it yields excellent results

Contact: zeynep [email protected]



Time: 17:20 - 19:00

Room: PS1

2

GRAPHTRACK: FAST ANDOPTIMAL TRACK-

ING

Brian A.

Abstract: Special effects in movies require tracks of features through scenes. Tracks are

found in an interactive process. The artist marks a position, and the computer proposes a track

which is then further refined by the artist. This is a difficult problem due to three aspects. -

Sudden appearance changes due to lighting and pose - Occlusions - Speed: Interactive editing

requires higher than framerate speed We formulated tracking as path search in a large graph,

and solve it efficiently with a modificiation of Dijkstra’s algorithm. The method is based on

[Buchanan and Fitzgibbon, 06]. Our main contributions are - Efficient incorporation of a

background appearance model - Formulation as a shortest path problem - Correct handling of

occlusions - High-Efficiency implementation with up to 200 fps for a high resolution video




Time: 17:20 - 19:00

Room: PS1

3

MULTI-TARGET TRACKING BY CONTINU-

OUS ENERGY MINIMIZATION

Andriyenko A., Schindler K.

Abstract: We propose to formulate multi-target tracking as minimization of a continuous

energy function. Other than a number of recent approaches we focus on designing an energy

function that represents the problem as faithfully as possible, rather than one that is amenable

to elegant optimization. To find strong local minima of the proposed energy we extend the

conjugate gradient method with periodic trans dimensional jumps. Experiments on public

datasets validate our approach.




Time: 17:20 - 19:00

Room: PS1

4

TRAINABLE V4-LIKE FILTERS FORDETECT-

ING RETINAL VASCULAR BIFURCATIONS

Azzopardi G., Petkov N.

Abstract: We propose a novel method to detect vascular bifurcations in retinal fundus

images. Our method is implemented in trainable filters that mimic the properties of some

neurons in area V4 of visual cortex. Such a filter is configured by combining given channels

of a bank of Gabor filters using an AND-type operation. Their selection is determined by an

automatic analysis of a user-specified feature. With only 25 filters we report a recall rate of

98.52% at a precision rate of 95.19% on 40 images.




Time: 17:20 - 19:00

Room: PS1

5

ONLINE APPEARANCEMODELS FOR VISUAL

TRACKING

Bachoo AK., Nicolls F.

Abstract: An ongoing research area in object tracking is the representation and online

update of the target appearance. This work examines online appearance models for improved

visual tracking. The primary idea is to use both background and foreground models in a

statistical framework for improved target tracking. Motivating works are discussed for the

general research and then results are presented for a basic tracking system. The tracker uses

template matching within a particle filter. The template is updated online and a robust error

function detects occlusions and outliers. The results are very promising and the tracker is

robust to occlusions, white caps, low contrast and camera motion when tested in a maritime

environment.




Time: 17:20 - 19:00

Room: PS1

6

MAKING INTELLIGENT TUTORS EMOTION-

ALLY AWARE

Banda N., Robinson P.

Abstract: Our research is directed toward equipping intelligent tutoring systems with the

ability to infer complex mental states from visual and audio cues. This will allow the tutoring

system to maximize the learning potential of a student by adjusting the study material in

response to the detected mental state. Inference is achieved by analysing facial expressions,

head gestures and audio features.




Time: 17:20 - 19:00

Room: PS1

7

MULTI-POSE FACE RECOGNITION FOR PER-

SON RETRIEVAL IN CAMERA NETWORKS

Bauml M., Bernardin K., Ekenel H., Stiefelhagen R.

Abstract: We study the use of facial appearance features for the re-identification of persons

using distributed camera networks in a realistic surveillance scenario. In contrast to features

commonly used for person re-identification, such as whole body appearance, facial features offer

the advantage of remaining stable over much larger intervals of time. The challenge in using

faces for such applications, apart from low captured face resolutions, is that their appearance

across camera sightings is largely influenced by lighting and viewing pose. Here, a number of

techniques to address these problems are presented and evaluated on a database of surveillance-

type recordings. A system for online capture and interactive retrieval is presented that allows to

search for sightings of particular persons in the video database. Evaluation results are presented

on surveillance data recorded with four cameras over several days.




Time: 17:20 - 19:00

Room: PS1

8

AUTOMATED CATEGORIZATION OF ABNOR-

MALITIES BASED ON COMPUTER VISION

ANALYSIS OF LARGE MEDICAL IMAGING

COLLECTIONS

Burner A., Donner R., Mayerhoefer M., Kainberger F., Langs G.

Abstract: The purpose of this PhD thesis is to make use of the vast amount of medical

imaging data available in hospitals to learn, categorize, and to retrieve similarities of abnormal-

ities. Currently, computer vision research is taking place on a pathology-by-pathology basis,

applying simple tasks of classifying images into modality and anatomic region and individually

developed supervised learning methods that identify pathologies for a certain anatomical part.

Such an approach is limited, especially when it comes to massive amounts of data. In con-

trast, the goal of this research is to automatically learn the appearance of frequently occurring

pathologies, and, if a certain case is presented, let a CAD system retrieve the most similar

images. To achieve this, a 3D local binary pattern (3D LBP) algorithm was implemented to

model the appearance and spatial configuration of anatomies. A special focus was given to

the high volume of medical imaging system and the requirement that the system performs and

scales well with the increasing demand for medical images.




Time: 17:20 - 19:00

Room: PS1

9

FEATURE EXTRACTION FORNON-TEXTURED

OBJECT RECOGNITION USING A STEREO

CAMERA

Byeon W.

Abstract: Recognizing object is one of the central issues in computer vision. There are a

bunch of features available for object recognition such as shape, color, and texture. Among

these features, texture particularly turns out useful to distinguish object in natural scenes. It is

robust to changes in illumination and scale, occlusions as well as view point variations. However

recognizing object which lacks of textures has been rarely touched. This work presents a novel

solution to extract the features for recognizing object which has limited texture information,

such as cup, door, plate and bookcase, using a stereo camera. Considering that very few textures

are available to represent this type of object, we instead make use of the two different source

information: color and three dimensional (3D) scale which is based of high-curvature points. In

our experiments, the proposed method has been tested in several challenging datasets in which

various condition changes, such as illumination, occlusion, scale, and rotation, are incorporated.

As indicated in the experiments, by using the result of non-textured object feature extraction,

high recognition precision has been achieved. The proposed approach also demonstrated more

robust recognition ability in non-textured objects in natural scenes.




Time: 17:20 - 19:00

Room: PS1

10

FLEXIBLE INTELLIGENT VISUAL SURVEIL-

LANCE SYSTEM

Chang H. J., Yi K. M., Yin S., Kim S. W., Choi J. Y.

Abstract: We develop a real-time intelligent visual surveillance system named as percep-

tion and intelligence lab - enhance your eye (PIL-EYE) by applying flexible modular system

architecture. Any functional module and algorithm can be added or removed independently.

Also, functional flow can be designed by simply placing the order of modules. Algorithm opti-

mization becomes easy by checking computational load in real time and commercialization can

be easily achieved by packaging of modules.


Presentation Type: To be announced

Date: To be announced

Time: To be announced

Room: To be announced

11

TOWARDS THE EXAGGERATED IMAGE STEREO-

TYPES

Chen C., Lauze F., Igel C., Feragen A., Loog M., Nielsen M.

Abstract: Given a training set of images and a binary classifier, we introduce a concept of

Exaggerated Image Stereotype for one of the two classes, based on the combination of generative

and discriminative models, respectively built from the training set and the classifier. The

exaggerated image stereotype should emphasize / exaggerate patterns in an image, resulting

in an optimal trade-off between classification result and likelihood of being generated from the

class of interest.




Time: 17:20 - 19:00

Room: PS1

12

UNCOOPERATIVE 3D FACE RECOGNITION

FOR INTELLIGENT SURVEILLANCE APPLI-

CATIONS

Cheng X., Lakemond R., Fookes C., Sridharan S.

Abstract: Uncooperative face recognition at a distance is a challenging problem since sub-

ject pose and lighting is not controllable. Compared to 2D approaches, 3D face recognition is

relatively invariant to pose or illumination variations. Existing 3D techniques either rely on

special devices or are too computationally expensive. This research aims to develop a 3D face

recognition method to identify uncooperative subjects from video sequences captured by single

or multiple video cameras.




Time: 17:20 - 19:00

Room: PS1

13

AUTOMATIC ANNOTATION OF HISTOPATHO-

LOGICAL IMAGES USING A LATENT TOPIC

MODEL BASED ONNONNEGATIVEMATRIX

FACTORIZATION

Cruz-Roa A., Dıaz G., Gonzılez F.

Abstract: Histopathological images are an important resource for clinical diagnosis and

biomedical research.The automatic analysis and annotation of these images is particularly chal-

lenging from an image understanding point of view; in this type of images, visual patterns are

generally a complex combination of fundamental visual features involving texture, color and

shape. This paper presents a novel method for automatic histopathological image annotation

using nonnegative matrix factorization. The proposed method uses a part-based image repre-

sentation, called bag of features, to represent the visual information of a histopathology image

collection in conjunction with a latent-topic-model based on nonnegative matrix factorization.

The method was evaluated over a histopathology dataset used to diagnosis of a skin cancer

known as basal cell carcinoma. The preliminary results are promising, showing an improved

generalization of the proposed method when compared to a baseline image annotation strategy

based on support vector machines.




Time: 17:20 - 19:00

Room: PS1

14

A PROBABILISTIC GENERATIVE APPROACH

TO INVARIANT VISUAL INFERENCE AND

LEARNING

Dai Z., Lucke J.

Abstract: In this work we study a probabilistic generative approach that explicitly ad-

dresses the translation invariance of objects in visual data. Object location is modeled using

an explicit hidden variable while the object itself is encoded by a specific spatial combination

of features. The investigated generative model autonomously learns from unlabeled data with

object identity and position. By using a probabilistic generative approach, we can show that an

object’s feature combination can reliably be learned based on a maximum likelihood approach.

We demonstrate the algorithm using artificial and more realistic visual data.




Time: 17:20 - 19:00

Room: PS1

15

EVENT TYPE CLASSIFICATION FOR PER-

SONAL PHOTO ALBUMS BASED ON SIGNA-

TURE IMAGE

Dang-Nguyen D.-T., Dao M.-S., Boato G., DeNatale F.

Abstract: Analyzing personal photo albums for understanding the conveyed events is an

emerging trend. It can help to annotate events and their components in order to support for

organizing and sharing the event-related information among the users. Therefore, having a

fast event-type classifier for personal photo albums could be considered as a basic requirement.

In this paper, a novel method for fast event-type classification of personal photo albums is

presented. Distinct from previous approaches, the proposed method does not process photos

of an event as individuals but as a whole by which not only gist and saliency but also time

information are captured and represented as low-level features towards mimicking biological

vision. To capture both gist and saliency of an event, a 2D histogram, called a GS-SIB, is

created by extracting dominant colors and salience map from all photos, i.e., each photo is

projected to a point of the GS-SIB image according to its dominant color and salience map

pattern value. Photos is then sorted by time and built as a sequence of symbols, called a

T-SIB. Each symbol and its order in T-SIB represent for photo coordinate in GS-SIB and

chronological order, respectively. The weighted sum of the differences between GS-SIBs, and

between T-SIBs is the discriminant value to classify event-type. A highly challenging database

of 19.101 photos from five varied event-types was used for evaluating the performance. The

experimental results show that the proposed method is capable of classifying with high accuracy

and low computational cost when comparing to other methods.




Time: 17:20 - 19:00

Room: PS1

16

CYLINDRICAL PANORAMA MATCHING

De Carufel J.-L., Laganiere R.

Abstract: Classical image matching methods such as SURF, SIFT and ASIFT are meant

to be used on planar images. SURF and SIFT are partially affine invariant, while ASIFT is

fully affine invariant. When we use these methods on cylindrical panoramas, the results are

not as good as expected. We present a method inspired by ASIFT that simulates different

transformations of the panoramas to be matched and then tries to match them.




Time: 17:20 - 19:00

Room: PS1

17

AUGMENTED PERCEPTION AND INTERAC-

TION WITH HANDHELD DEVICES

De Tommaso D., Calinon S.

Abstract: We propose a method for enabling handheld devices to share a unique 3D spatial

representation of the same environment using mixed reality. This method provides an architec-

ture for collecting environmental data from different kinds of portable devices by their relative

perceptual space, but also enables humans to actively interact inside the scene. In our real

scenario we consider a RGBD sensor mounted on a robotic arm, used to collect 3D informa-

tion from the environment, a pico-projector to visualize augmented digital information on real

objects and a tablet pc to visualize virtual trajectories of the arm in the real spatial position.




Time: 17:20 - 19:00

Room: PS1

18

NON-PARAMETRIC SUB-PIXEL LOCAL POINT

SPREAD FUNCTION ESTIMATION

Delbracio M., Muse P., Almansa A., Morel JM.

Abstract: - We prove formally that the non-parametric sub-pixel PSF estimation problem

is well-posed with a single well chosen observation. - Near-optimal accuracy achieved with a

Bernoulli(0.5) noise calibration pattern. - Local PSF estimated by solving a well conditioned

linear system that does not require regularizers. - Relative estimation error of 2% to 5%. -

Such a regularization and model free subpixel PSF estimation scheme is the first of its kind, to

our knowledge.




Time: 17:20 - 19:00

Room: PS1

19

A METHOD FOR NOISY IRIS SEGMENTA-

TION

Donida Labati R., Scotti F.

Abstract: We present an innovative algorithm for the segmentation of the iris in noisy

images, with boundaries regularization and the removal of the possible existing reflections.

The method achieves the iris segmentation by three main steps: estimation of the pupil and iris

centers; iris boundaries extraction, linearization, and regularization; detection of reflections and

occlusions. The proposed algorithm ranked seventh in the international Noisy Iris Challenge

Evaluation (NICE.I) [1].




Time: 17:20 - 19:00

Room: PS1

20

A SEEDED FRAMEWORK FOR TRACKING

OF GENE EXPRESSION DYNAMICS IN IN-

DIVIDUAL CELL COMPARTMENTS

Du C.-J., Marcello M., Spiller D.G., White M.R.H., Bretschneider T.

Abstract: For tracking of gene expression dynamics in individual cell compartments, we

propose to initialize a few keyframes with a novel constrained interactive segmentation method,

which uses the constrained density weighted Nystrom method for eigenvector decomposition

and the geodesic commute distance for pixel classification. The tracking is achieved by both

forward and backward propagating from two keyframes using a group of overlapping subwindow

segmenters around nuclei and cytoplasm boundaries. Experiments demonstrate that the system

is able to track quite complex cell sequences reasonable well.




Time: 17:20 - 19:00

Room: PS1

21

SPATIAL INVARIANCE FOR SCENE CLASSI-

FICATION

Dunlop H.

Abstract: We present a method for categorizing scenes using properties from local, interme-

diate, and global scales with a learned spatially invariant model. By applying a spatial pyramid

to regions of varying sizes and in different spatial locations, we are able to better characterize

the appearance of images. Adding spatial invariance that is modeled for each scene category,

these mid-sized region descriptors complement well the local and global methods already in

common use for scene classification.




Time: 17:20 - 19:00

Room: PS1

22

A SIMPLE INPAINTING METHOD AND ITS

GPU IMPLEMENTATION

Fassold H.

Abstract: A simple image inpainting method is proposed, and its efficient GPU implemen-

tation for NVIDIA GPUs is described. A speedup factor of 7 - 11 is observed for the GPU

implementation, compared with an optimized CPU implementation.




Time: 17:20 - 19:00

Room: PS1

23

CLASSIFICATION OF THE ACROSOME IN-

TEGRITY OF BOAR SPERMATOZOAHEADS

USING SURF AGAINST TRADITIONAL DE-

SCRIPTORS

Fernandez-Robles L., Alegre E.

Abstract: Automatic assessment of sperm quality is an important challenge in the veteri-

nary field. Our proposal is to characterize the acrosomes of boar spermatozoa heads as intact

or damaged using SURF descriptors, and compare them with Local Binary Pattern (LBP),

Flusser, Hu, Zernike and Legendre descriptors. We classify the images with k-Nearest Neigh-

bours. Experimental results point out that SURF descriptors are better, reaching an accuracy

of 94.88




Time: 17:20 - 19:00

Room: PS2

24

UNSER FEATURES ESTIMATION FOR REAL-

TIME TISSUE CHARACTERIZATION

Galluzzo F., Testoni N.

Abstract: In Ultrasound (US) guided prostate cancer detection, Computer Aided Detec-

tion (CAD) systems can be used to reduce the number of false positives. Diagnostic information

must be provided with a minimum frame rate of 2-3 fps. The work proposes a parallel imple-

mentation of Unser textural features computation algorithm suitable for analyzing US images.

Our CUDA implementation is effective to speed up features estimation leading to execution

times suitable for real-time tissue characterization. A CAD system exploiting real-time Unser

features overcomes a classical biopsy protocol.




Time: 17:20 - 19:00

Room: PS2

25

WORDRECOGNITIONMETHODS THROUGH

WORD SHAPE CODING TECHNICS COMPAR-

ISON

Garcıa-Ordas M.T., Alegre E.

Abstract: The aim of this work is to recognize words through their shape. Two methods

have been studied and modified to improve the results: holistic word recognition[1] and retrieval

of machine-printed Latin documents through word shape coding[2].




Time: 17:20 - 19:00

Room: PS2

26

EVALUATION AND IMPROVEMENT OF ADAP-

TIVE FILTERS FOR SHARPNESS ENHANCE-

MENT AND NOISE REMOVAL

Garcıa-Olalla O., Alegre E.

Abstract: Two adaptive filter algorithms for sharpness enhancement and noise removal

have been evaluated. A modification to the Adaptive Bilateral Filter(ABF) method [1] have

been carried out improving the former results. A metric comparision with adaptive Nonlinear

Complex Diffusion Filter (NCDF) algorithm [2] has been done using three methods: MSE

(mean square error), ENL (equivalent number of looks) and CNR (Contrast to noise ratio).




Time: 17:20 - 19:00

Room: PS2

27

NOISE REDUCTION USING BIO-INSPIRED

AND SOFT COMPUTING

Gonzalez Jaime L.

Abstract: Image denoising is still a challenge for the research community. There exist

many approaches to deal with it. However, the majority of these techniques are only capable

of efficiently dealing with white Gaussian noise. Medical images are characterized for more

complex noise. Our aim is to find new solutions to this issue applying different bio-inspired and

soft computing methods, as fuzzy logic or genetic algorithm among others. To that end, they

will be applied in different transform domains.




Time: 17:20 - 19:00

Room: PS2

28

ENHANCED SUPER-RESOLUTION FORMED-

ICAL DIAGNOSIS

Gonzalez Villanueva L., M. Callico G., Tobajas, F.

Abstract: Nowadays, images are employed in several areas of medicine for early diagnosis.

However, images related to pathological anatomy present in many situations poor quality, which

complicates the diagnostic process. This work is focused on the quality enhancement of this

type of images through a system based on super-resolution techniques. The results show that

the proposed methodology can help medical specialists in the diagnostic of several pathologies.




Time: 17:20 - 19:00

Room: PS2

29

HIERARCHICAL ARM-HANDGESTURESMOD-

ELLING AND RECOGNITION

Gori I., Fanello S.R., Demiris Y.

Abstract: We present an original, reliable and real-time gesture recognition system. Our

system exploits motion information without prior knowledge of the presence of humans in the

scene. We adopted a hierarchical approach; in particular we define a set of action primitives

that can be combined in order to obtain complex gestures. Accordingly, our system consists of

three levels: the first one is based on features extraction, the second one regards the modelling

and the recognition of simple actions, and the third one has been conceived for the modelling

and the recognition of complex gestures.




Time: 17:20 - 19:00

Room: PS2

30

DISTRIBUTED FRAMEWORK FOR MULTI-

CAMERA PEOPLE TRACKING

Gruenwedel S., Nino J., Jelaca V., Philips W.

Abstract: This work presents a distributed framework to track people in a multi-camera

network designed for real-time and scalability. For the task of foreground detection, we propose

an edge-based approach to robustly overcome the problem of lighting changes. We use edge

dependencies as statistical features of foreground and background regions and define foreground

as regions containing moving edges. Experiments prove the robustness of our method in the

presence of lighting changes as well as show promising results for 3D tracking.




Time: 17:20 - 19:00

Room: PS2

31

EMOTION ANALYSIS ONHUMAN FACE THROUGH

THERMAL IMAGING

Guler P., Dumlu Seda N.

Abstract: In this paper, emotion which can be categorized as valence and arousal, will

be detected from the human face during a game playing session. Thermal camera measures

the temperature changes according to different types of emotions such as valence or arousal.

These temperature measurements are used in understand emotion changes of human. Analyz-

ing human emotions using thermal infrared technology is useful in understanding people who

has disabilities like autism or paralyzer. It is not only crucial for detecting disorders but it

is also helpful for developing more complex robots that understands human emotions. Using

thermal camera in analyzing human emotions through temperature changes in face is getting

more popular since it is cheaper and more reachable than other methods like functional mag-

netic resonance imaging (fMRI). The hypothesis in this research lies on this mechanism; the

researchers of this study wonder whether the blood perfusion on the face regions is correlated

with the anxiety and stress levels of humans for the game technology or not. In this research,

for fetaure extraction, co-occurence matrices of the region of interests (ROIs) in each frame are

calculated. Energy, entropy, contrast, homogeneity, correlation features are calculated based on

co-occurence matrix. Thus, the dataset is obtained. ROIs are the regions where the tempara-

ture change can be seen in the face. Then, principle component analysis (PCA) is applied for

reducing dimensionality of fetaure set. Lastly, for discrimation of the frames where the player is

excited (exciting frames) from the frames where the player is not excited (non-exciting frames),

k-means clustering is applied to the dataset. According to the findings, the autors are able to

discriminate the exciting frames with 75% ratio. However, 50% of the non-exciting frames are

labeled as exciting frames.

Contact: puren [email protected]



Time: 17:20 - 19:00

Room: PS2

32

A PIPELINE FORMODELINGURBAN STREET

FACADES FROMTERRESTRIAL LASER AND

IMAGE DATA

Hammoudi K., Dornaika F., Soheilian B., Paparoditis N.

Abstract: This poster presents researches dealing with the 3D modeling of urban street

facades. In the last decade, the mapping field has strongly evolved due to the needs of civil

and military applications. At the French National Mapping Agency, approaches have been

developed in order to model street facades from laser and image data collected by terrestrial

mobile mapping systems (i.e., vehicle of acquisition). One of the objectives is the generation of

a realistic 3D viewer of street facades.




Time: 17:20 - 19:00

Room: PS2

33

ESTIMATION ERRORANALYSIS IN STEREO

COLOR MAPPING

Hasan S.F., Stauder J., Tremeau A.

Abstract: Color differences between views of a stereo pair is a challenging problem.Various

applications such as compression of stereo, 3D texture transfer, view interpolation etc. demand

the compensation of color differences which is typically done by color mapping. A large number

of existing color mapping is based on geometric feature correspondences which generates the

color correspondences and that’s the basis of color mapping model. Our focus here is to analyze

the impact of noise in feature correspondences and thus on the estimation of color mapping

model.




Time: 17:20 - 19:00

Room: PS2

34

RIEMANNIAN PERONA-MALIK DIFFUSION

FOR ORIENTATION DISTRIBUTION FUNC-

TION IMAGES

Krajsek K., Heinemann C., Scharr H.

Abstract: We generalize the Perona-Malik diffusion equation to ODF Images within a

Riemannian framework. To this end, we derive the PM diffusion equation from an energy

functional. Discretization as well as a numerical update scheme for solving resulting initial value

problems are developed. We provide a stability analysis of our update scheme and propose an

effective implementation by means of spherical harmonics. We demonstrate the performance of

our approach on synthetic as well as real data.




Time: 17:20 - 19:00

Room: PS2

35

LOGATOM RECOGNIZABILITY OF FINGER

ALPHABET IN VIDEO

Heribanova P., Polec J.

Abstract: This paper deals with the problem of the cued speech ( fingers alphabet ) recog-

nition methods in video. Cued speech is a specific gesture language used for communication

between deaf people. The aim of this paper is to show new objective method of testing con-

sonant sign recognizability in single-handed finger alphabet (dactyl) analogically to acoustics.

We used the sign logatoms to testing intelligibility because they have no meaning and can-

not be part of common words. From the results we construct the minimum coded bit-rate

recommendations for every spatial resolution.




Time: 17:20 - 19:00

Room: PS2

36

USINGHIGH LEVEL INFORMATION FOR LOW

LEVEL TRACKING

Horbert E., Mitzel D., Leibe B.

Abstract: Level Set Tracking is a low level tracking approach, where the object of interest is

segmented in the first frame, then tracked through the subsequent frames, in which the contour

is adapted. We integrate LS tracking with a tracking-by-detection framework and thereby show

how LS tracking can benefit from high level information, e.g. a ground plane estimate. The

LS tracker can also pass information to higher levels, such as object positions, a car’s rotation

angle or detailed appearance models.




Time: 17:20 - 19:00

Room: PS2

37

EXPLORING PHOTOBIOS

Kemelmacher-Shlizerman I., Shechtman E., Garg R., Seitz S.M.

Abstract: We generate face animations from large image collections of a person’s face, by

computing an optimized, aligned subsequence. This approach is the basis for the Face Movies

feature of Picasa. A key contribution is proving why the cross dissolve produces a strong motion

effect [1].




Time: 17:20 - 19:00

Room: PS2

38

LEARNING FACE RECOGNITION IN VIDEOS

FROMASSOCIATED INFORMATION SOURCES

Kostinger M.

Abstract: Videos are often associated with additional information that could be valuable

for interpretation of its content. This especially applies for the recognition of faces within video

streams, where often cues such as transcripts and subtitles are available. However, this data

is not completely reliable and might be ambiguously labeled. To overcome these limitations,

we propose a new semi supervised multiple instance learning algorithm, where the contribution

is twofold. First, we can transfer information on labeled bags of instances, thus, enabling us

to weaken the prerequisite of knowing the label for each instance. Second, we can integrate

unlabeled data, given only probabilistic information in form of priors.




Time: 17:20 - 19:00

Room: PS2

39

ONLINE DISCRIMINATIVE DICTIONARY LEARN-

ING FOR IMAGE CLASSIFICATION

Kong Shu, Wang Donghui

Abstract: Dictionary learning (DL) is an important technique for many vision perception

tasks, e.g., classification. But there are two problems of the classical DL (reconstructive DL)

arising in classification task, i.e., large-scale problem and better discriminability. For the two

issues, Online and discriminative framework are proposed, which means the ability to deal with

large-scale dataset, and the capability of distinguishing different kinds of objects, respectively.

We believe it makes sense to merge the two frameworks into one, and our work is focusedon

this issue.




Time: 17:20 - 19:00

Room: PS2

40

ANALYSING COMPLEX ACTIVITIES IN VIDEO

SEQUENCES

Kuehne H., Gehrig D.

Abstract: Video-based applications rely more and more on the fast and reliable recognition

of human actions. Although current recognition algorithms compete for better recognition

performance, their usability for real-world applications is still limited in terms of complexity,

runtime and robustness. In this context, a system for the continuous online recognition of

human actions from video-based motion information is proposed. It uses histograms of sparse

feature point flow with a Hidden-Markov-Model (HMM) based decoder system inspired from

speech recognition. To allow a recognition of ongoing tasks, complex sequences are split into

action units and during the recognition assembled by a context free grammar. The evaluation

of the presented system shows a good performance even compared to marker-based recognition

approaches.




Time: 17:20 - 19:00

Room: PS2

41

EYE-TRACKINGGUIDANCE FOR EXOSKELE-

TON REHABILITATION

Loconsole C.

Abstract: Eye-tracking systems are playing an increasingly important role in assistive

robotics as hand-free interaction interfaces for motor impaired people, but no noticeable ap-

plications have been developed so far for enhancing the robotic assisted training in functional

rehabilitation. In this poster it is proposed a new gaze based control to provide active guidance

to the upper limb movement, through a robotic exoskeleton, in the functional rehabilitation of

pick and place tasks. Experimental results on healthy subjects demonstrate the feasibility of the

proposed approach and the breakthrough that the system introduces in the field of eye-based

rehabilitation systems.



Date: Tuesday 12 July 2011

Time: 17:50 - 19:30

Room: PS1

42

AUTOMATIC ABDOMINAL ORGANS SEGMEN-

TATIONWITHMATHEMATICALMORPHOL-

OGY IN MR IMAGES

Lopez-Mir F., Naranjo V., Angulo J., Alcaniz M.

Abstract: This work proposes an automatic segmentation method based on the watershed

transformation. This method is tested in magnetic resonance datasets and focused on the

segmentation of the aorta artery [1] and the liver. After this segmentation process, a 3D

model is created with the purpose of projecting it over the patient to help the surgeon in the

trocar placement in laparoscopy surgery [2]. The watershed algorithm is a robust segmentation

method if the input image has well-defined boundaries and if the image minima represent

relevant objects. In abdominal MR images these hypotheses are two problems and for this

reason a pre-processing step is required. The problem of the necessity well-defined boundaries

are solved in the case of the aorta artery with the gradient image and different morphological

filters to eliminate structures with similar grayscale values. In the case of the liver this step is

more complex and the gradient is obtained applying an opening area operator. The problems of

the minima are solved with a variant of the watershed transformation called marked-controlled

watershed, which consists of using a set of markers to modify the gradient image. These markers

will be the new minima of the image and they prevent the over-segmentation, typical in the

basic watershed algorithm. In the case of the aorta artery two markers are defined, an internal

marker obtained as the geodesic centre of the adjacent slice (in the z axis) and an external

marker obtained as the perimeter of the dilatation of the adjacent slice (in the z axis). In the

liver, the external marker is obtained as in the aorta artery (the perimeter of a dilatation of

the adjacent slice) but in the case of internal markers, several internal markers are calculated

to prevent the break of the liver and the possibility of the hepatic tree isolates some parts of

the liver. Finally a threshold decided if a region is liver or not.




Time: 17:50 - 19:30

Room: PS1

43

HIERARCHICAL SHAPEMODEL FOR OBJECT

DETECTION

Macak J., Drbohlav O.

Abstract: Hierarchical models are very suitable for detection and classification of objects

from high number of classes because of their natural ability to model the exponential growth of

the complexity of image scenes as the resolution and the number of objects in scene increases.

The model is build in layers and each layer consists of a set of compositions that are frequent

in the learning data. Each layer describes relations between compositions from the preceding

layer. The lowest layer describes small areas of objects like line segments, arcs, etc. The highest

layer can describe a whole object or a big part of an object. In this framework, compositions of

all but especially lower layers can be easily shared within categories which leads to an effective

description of a high number of classes. This approach is based on the work of Leonardis et al.




Time: 17:50 - 19:30

Room: PS1

44

COMBINING 2D AND 3D OBJECT CATEGO-

RIZATION FOR TASK CONSTRAINEDGRASP-

ING

Madry M., Song D., Kragic D.

Abstract: We present a system able to transfer grasp knowledge between object categories

defined by geometric properties and functionality. In the center lies an Object Categorization

Module (OCM) based on 2D and 3D visual data that is integrated with a grasp planner. The

system runs on a robot (ARMAR-III) equipped with active stereo cameras. The experimental

evaluation compares individual 2D and 3D categorization with the fused 2D-3D OCM, and

shows the usefulness of the approach in task-based grasping.




Time: 17:50 - 19:30

Room: PS1

45

INTERPRETINGHAND-OVER-FACE GESTURES

Mahmoud M. , Baltrusaitis T., Robinson P.

Abstract: People often hold their hands near their faces as a gesture in natural conversa-

tion, which can interfere with affective inference from facial expressions. However, these gestures

are valuable as an additional channel for multi-modal inference. We have collected a 3D multi-

modal corpus of naturally evoked complex mental states, and labelled it using crowd-sourcing.

The database will be made generally available.




Time: 17:50 - 19:30

Room: PS1

46

STEREO PERFORMANCE FOR CLUTTERED

SCENES

Mannan F., Langer M. S.

Abstract: Previous works on stereo evaluation did not address performance for different

types of scenes. This work [1], evaluates MRF-based stereo formulations for cluttered scenes

[2]. Three types of methods are considered: basic (Basic)[3], uniqueness (KZ-Uni)[4], and

visibility (KZvis)[ 5]. These are evaluated based on mislabeled pixels of different types (binoc-

ular/monocular) in different regions (on or away from occlusion boundary).




Time: 17:50 - 19:30

Room: PS1

47

VISUAL SERVOING FORMICROAERIAL VE-

HICLES

Maurer M., Katusic M., Bischof H.

Abstract: We present a visual height controller for a micro aerial vehicle (MAV). The MAV

is part of an autonomous visual inspection setup for power pylons. We demonstrate our progress

in visual servoing and focus on a height controller using fuzzy logic. In preliminary results we

show that the MAV can reach a desired height at a speed of 0.06m/s and a height accuracy of

0.000475m (MSE). The generalization to full 6 DoF will allow an accurate positioning even in

outdoor environments.




Time: 17:50 - 19:30

Room: PS1

48

INTER-MODALITY REGISTRATION TOGUIDE

CARDIAC PROCEDURES

McManigle J.E., Arai A., Noble J.A.

Abstract: Percutaneous cardiac procedures are currently guided by real-time echocardio-

graphy and fluoroscopy. However, targets of these procedures are often identified by cardiac

magnetic resonance (CMR) and x-ray computed tomography (CT) imaging. This project in-

troduces wavelet fusion echo images as an intermediate registration step to permit rapid regis-

tration from intra-operative echo to pre-operative CMR and CT. This may allow beating-heart

guidance with these more effective modalities.




Time: 17:50 - 19:30

Room: PS1

49

CAMERA CALIBRATION AND IMAGE DIS-

TORTION CORRECTION FOR SUPERIOR VI-

SUALIZATION IN MEDICAL ENDOSCOPY

Melo R., Barreto J, Falcao G.

Abstract: Medical endoscopy is used in a wide variety of diagnostic and surgical procedures.

These procedures are renowned for the difficulty of orienting the camera and instruments inside

the human body cavities. The small size of the lens causes radial distortion of the image, which

hinders the navigation process and leads to errors in depth perception and object morphology.

This article presents a complete software-based system to improve the visualization in clinical

endoscopy by correcting radial distortion in real time. Our system can be used with any

type of medical endoscopic technology, including oblique-viewing endoscopes and HD image

acquisition. The initial camera calibration is performed in an unsupervised manner from a

single checkerboard pattern image. For oblique-viewing endoscopes the changes in calibration

during operation are handled by a new adaptive camera projection model and an algorithm that

infer the rotation of the probe lens using only image information. The workload is distributed

across the CPU and GPU through an optimized CPU+GPU hybrid solution. This enables real-

time performance, even for HD video inputs. The system is evaluated for different technical

aspects, including accuracy of modeling and calibration, overall robustness and runtime profile.




Time: 17:50 - 19:30

Room: PS1

50

REDUCING THE PROBLEMOF OCCLUSIONS

IN LASER-TRIANGULATION RECONSTRUC-

TION

Munaro M., Michieletto S., Menegatti E.

Abstract: This poster presents a method for reducing the problem of occlusions in a laser-

triangulation system for 2.5D models creation. It is focused on reducing occlusions in the

direction of movement of the camera by exploiting two laser projectors instead of one and

providing the corresponding calibration and registration algorithms.




Time: 17:50 - 19:30

Room: PS1

51

SYMBOL RETRIEVAL BY IDENTIFYING RE-

PEATING PATTERNS

Nayef N., Breuel T.

Abstract: Content analysis of images is essential for search engines and retrieval applica-

tions. This work presents a method for content analysis in technical line drawings. This is

achieved by: 1) Identifying patterns by statistical grouping 2) Clustering the patterns using

geometric matching. The clusters form a symbol library to be used in Symbol retrieval. The

results on a standard dataset of architectural drawings are significantly higher than all the

results reported so far.




Time: 17:50 - 19:30

Room: PS1

52

HUMAN SHAPE AND POSE RECOVERY FROM

MONOCULAR IMAGES USING STATISTICAL

MODELS

Neophytou A., Guillemaut J.-Y.,Hilton A.

Abstract: In this work we investigate the use of Statistical Models in recovering human

shape and pose from single images. From a set of aligned 3D scans we aim to construct a

deformable human model that captures both the variation in human shape and pose as well as

the correlation between them. By fitting this model to an image, a realistic human body model

of the person on the image can be extracted. It is also possible to estimate other meaningful

attributes such as height, weight, etc.




Time: 17:50 - 19:30

Room: PS1

53

REAL-TIME OBJECT DETECTION FORMULTI-

CAMERA SURVEILLANCE

Nino J., Pizurica A., Philips W.

Abstract: Real-Time video surveillance presents issues related to illumination. For the

task of object detection, we propose a method that makes use of an intensity-invariable but

fragmented LBP foreground mask. A C4.5 Decision Tree is trained with the purpose of inte-

grating foreground fragments as object components. The resulting classified segments are then

integrated in a Dynamic Bayesian Network framework. Promising results are obtained using

few examples for training the classifier, which can ease the set-up of the detector in real life

multi-camera environments.




Time: 17:50 - 19:30

Room: PS1

54

EYE REFLECTION ANALYSIS AND APPLI-

CATIONS

Nitschke C., Nakazawa A., Takemura H.

Abstract: Recently, the geometric relation between a human eye and its image has been

formalized to analyze corneal reflections [NN06]. Proceeding with these efforts, we aim in

strategies exploring the relation of camera, eyes, and scene in arbitrary environments to enable

insights for human-scene interaction. We study the light transport under multiple eyes, includ-

ing calibration, feature matching, back and forward projection. The findings enable a novel

method for display-camera calibration.




Time: 17:50 - 19:30

Room: PS1

55

FACE RECOGNITION FROM SINGLE SAM-

PLE

Omelina L.

Abstract: An accurate face recognition from a single sample is a challenging task and

number of issues, like changing expression, remain to be addressed. Facial expression, changing

face geometry, has a strong influence on the face recognition accuracy. This negative influence

is even higher when only a single training sample is available. We extend the idea of a rule

based approach to generate a synthetic facial expression for the recognition purpose.




Time: 17:50 - 19:30

Room: PS1

56

RECONSTRUCTING A 3D TRAJECTORYUN-

DER PERSPECTIVE PROJECTION

Park H.S., Shiratori T., Matthews I., Sheikh Y.

Abstract: We present an algorithm for reconstructing the 3D trajectory of a moving point

from its correspondence in 2D images, given the 3D pose and time of capture of the cameras

that produced each image. We solve for the trajectory parameters using linear least squares

followed by nonlinear optimization and study a geometric analysis of the problem. This enables

us to reconstruct 3D motion from videos or images and to characterize the cases when accurate

reconstruction is possible.




Time: 17:50 - 19:30

Room: PS1

57

AUTOMATIC LATENT FINGERPRINTMATCH-

ING

Paulino A. A., Jain A. K., Feng J.

Abstract: Identifying suspects based on impressions of fingers lifted from crime scenes

(latent prints) is extremely important to law enforcement agencies. Latents are usually poor

quality images blurred, smudgy, small area. Our goal is to improve latent matching accuracy

by using few manually marked features. We propose to align the latent with the fingerprint in

the database using a descriptor-based Hough Transform, establish minutiae correspondences,

and compute a global similarity score. Our experimental results show improvement in the

performance compared to a commercial matcher.




Time: 17:50 - 19:30

Room: PS1

58

UNMANNEDAERIAL SYSTEMS FORWILDLIFE

MONITORING

Perez F., Albo C., Viguria A., Ollero A.

Abstract: Recent advances in key technologies have boosted the development of Unmanned

Aerial Systems (UAS) for many civilian applications regarding inspection, surveillance or mon-

itoring. The overall aim of this research is to investigate the use of UAS for making population

census and biological sample collection, as well as evaluating the environmental impact of infras-

tructures on the natural environment. The implementation of such system intends to develop

new data acquisition methods using advanced technologies that have already shown their use-

fulness in other areas, but have been rarely implemented in a realistic way as a technological

tool for environmental preservation.




Time: 17:50 - 19:30

Room: PS1

59

ROBUST PEOPLE DETECTION BASED ON

APPEARANCE AND SHAPE

Pishchulin L., Jain A., Wojek C., Andriluka M., Thormaehlen T., Schiele B.

Abstract: In this work, we investigate how 3D shape models from computer graphics can

be leveraged to ease training data generation. In particular we employ a rendering-based

reshaping method to generate thousands of synthetic samples from only a few persons and

views. Experiments on a challenging multi-view dataset indicate that the data from just eleven

persons suffices to achieve good performance, while combination of our synthetic data with real

data outperforms even the state of the art.




Time: 17:50 - 19:30

Room: PS1

60

SUPER-RESOLUTION OF BRAINMR IMAGES:

A SPARSE REPRESENTATION APPROACH

Rueda A., Malpica N., Romero E.

Abstract: Spatial resolution of Magnetic Resonance (MR) imaging is limited by diverse

physical, technological and patient safety considerations. These factors together affect the

precision of brain tissue segmentations, producing voxel misclassifications and distorting mor-

phometry results. This work presents the application of sparse representations to generate

high-resolution versions of brain MR images, by mixing up high and low frequency information

with prior knowledge.




Time: 17:50 - 19:30

Room: PS1

61

DEPTH ACCURACY GAINS FOR 3D RECON-

STRUCTION FROM MULTIPLE VIEWS

Rumpler M., Irschara A., Bischof H.

Abstract: This work investigates the influence of using multiple views for 3D reconstruction

with respect to depth accuracy and robustness. We perform synthetic experiments on a typical

aerial photogrammetric camera network and investigate how baseline (i.e. triangulation angle)

and redundancy affect the depth uncertainty of triangulated scene points. Furthermore, we

propose an efficient dense matching algorithm that utilizes pairwise optical flow followed by a

robust correspondence chaining approach.




Time: 17:50 - 19:30

Room: PS1

62

PARALLELISING BUNDLE ADJUSTMENT

Salas-Moreno R.

Abstract: Bundle Adjustment (BA) is one of the final steps in feature-based 3D reconstruc-

tion with a moving camera. By optimising the estimated set of point and camera parameters,

it helps to increase trajectory accuracy and reduce error-buildup. Aiming to optimise dense

reconstruction results in real-time [1], we identified sub-steps suitable for parallel computations

and implemented a hybrid GPU/CPU solution with speed-ups of up to 10 times compared to

a recent CPU-only version [2].




Time: 17:50 - 19:30

Room: PS1

63

SPATIO-TEMPORAL REGISTRATION FORDIG-

ITAL ROBUSTWATERMARKING IN VIDEOS

Schaber P., Kopf S., Effelsberg W.

Abstract: Digital Robust Watermarking is a technique to irreversibly embed information

into media such as digital video. For extraction, most schemes require the unmarked original

video. As the marked copy can be significantly distorted, a precise spatio-temporal registration

is required. Thus, feature analysis and registration is a very important task. We present some

of our current work, e.g., a temporal registration with sub-frame precision, and a watermarking

approach based on geometric modifications.




Time: 17:50 - 19:30

Room: PS1

64

BODY HEIGHT ESTIMATION USING A SIN-

GLE CAMERA

Scharfenberger C., Chakraborty S., Faerber G.

Abstract: Due to increasing interest in maximizing passenger’s comfort, body height esti-

mation aims to facilitate ingress by means of individually adjusted seat positions. Our algorithm

robustly extracts approaching drivers in panoramic images for various parking scenarios. Based

on extracted head and foot points, our approach estimates the camera pose relative to a ground

plane. Thus, it enables absolute height estimation using a single omnidirectional camera at-

tached to the car.




Time: 17:50 - 19:30

Room: PS2

65

DETECTION OF PEOPLE ANDMOTION ANAL-

YSIS IN AERIAL IMAGE SEQUENCES

Schmidt F.

Abstract: The subject of the presented research project is the development of suitable

methods for the detection of people and the analysis of their motion in aerial image sequences.

In this poster we give an overview of the overall system and the available data. Furthermore we

deliver insights into the modules for object detection, object tracking and density estimation

and show some qualitative results.




Time: 17:50 - 19:30

Room: PS2

66

A COMPUTATIONAL METHOD FOR QUAN-

TITATIVE ASSESSMENT OF THEMOVEMENT

OF THE THORACIC AORTA

Schwartz E., Holfeld J., Czerny M., Langs G.

Abstract: The causal relationship between strain on the vessel wall and the development

of vascular pathologies is a common hypothesis in vascular medicine. However, the study of

these effects is mostly limited to static vessel models. Limited by the intricate fluid-structure

interactions, state-of-the art hemodynamic methods work under the assumption that motion

of the vessel wall not due to blood flow can be neglected. While this simplification allows

for trustworthy measurements of forces in vessels further from the heart, it does not hold for

the thoracic aorta, which is strongly influenced by the motion of the myocardium. Pathologies

arising in this part of the vascular system are amongst the most hazardous and difficult to treat.

Nonetheless, their pathogenesis, ramifications and the precise effects of the various treatment

options available are only poorly understood. The purpose of this study is the development and

application of methods capable of describing the motion of the thoracic aorta from ECG-gated

CT sequences during the cardiac cycle. The obtained measurements are used for deepening

the understanding of patterns of movement in the thoracic aorta as well as for the evaluation

of the effects of interventions such as stent-grafting. Results of the proposed method on 4

examplary cases are presented. In all of these, a pathology in the aortic arch has been treated

by supra-aortic rerouting followed by stent-graft placement. ECG-gated CT sequences are

aquired after rerouting and after stent-graft placement. The movement during one sequence

is computed using an automatic registration method that results in a dense deformation field.

An automatic segmentation procedure is used to extract the region of interest - the wall of the

aorta. Comparing these pre- and post- interventional models allows for new insights into the

detailed effects of the performed interventions.




Time: 17:50 - 19:30

Room: PS2

67

FROM LDA TO VISION VIA POPULATION

STRUCTURE

Sharmanska V., Lampert C.H.

Abstract: STRUCTURE is a model-based clustering method, which infers population

structure and assigns individuals to populations; the model considers each individual as a mix-

ture of a few source populations. Latent Dirichlet allocation (LDA) is a generative approach for

topic modeling tasks in text processing; the model finds the thematic structure of a collection

of documents. We show that these two models describe the same generative process.




Time: 17:50 - 19:30

Room: PS2

68

THE IMPORTANCE OF SHAPE TO ECHOCAR-

DIOGRAM SEGMENTATION

Stebbing R.

Abstract: Echocardiography is an important tool in the diagnosis of heart-disease. Auto-

matic segmentation of the endocardium is desirable to provide physicians with a fast method

for measuring various parameters of heart function. While the segmentation problem appears

straightforward, methods relying only on local image features perform poorly because of the

variation in patient anatomy and acquisition variables. Incorporating global shape information

is therefore required.




Time: 17:50 - 19:30

Room: PS2

69

DYNAMIC TEXTURE PREDICTION FORH.264/AVC

INTER CODING

Stojanovic A., Ohm J.-R.

Abstract: We propose an extension to a H.264/AVC encoder that improves the coding

performance for sequences containing dynamic textures. Based on a model introduced in [1],

we compute a prediction frame from already encoded frames. The encoder decides whether

to use the synthesized content using an RD decision. A new macroblock mode along with an

additional reference frame with flexible position are introduced for the purpose. The latter are

used by means of signaling to the decoder to synthesize certain regions in an efficient way. A

bitrate reduction of up to 30% over H.264/AVC at equal PSNR is achieved.




Time: 17:50 - 19:30

Room: PS2

70

A FREE-VIEWPOINT VIRTUALMIRRORWITH

MARKER-LESS USER INTERACTION

Straka M., Hauswiesner S., Rther M., Bischof H.

Abstract: We present a Virtual Mirror system which is able to simulate a physically correct

full-body mirror on a monitor. In addition, users can freely rotate the mirror image which allows

them to look at themselves from the side or from the back, for example. This is achieved through

a multiple camera system and visual hull based rendering. A real-time 3D reconstruction and

rendering pipeline enables us to create a virtual mirror image at 15 frames per second on a

single computer.




Time: 17:50 - 19:30

Room: PS2

71

FUNCTIONAL ANALYSIS OF NEURAL NET-

WORKS: NEURONS SEGMENTATION AND

ACTIVATION MAPS

Ullo S., Del Bue A., Murino V., Maccione A., Berdondini L.

Abstract: In the last few decades, neuroscience has tried to go beyond the analysis of

single isolated neuronal structures, toward the study of their mutual functional interaction

over time. In this work we propose a new approach for studying in vitro neuronal networks,

by understanding their morphology (detecting neurons nuclei) and providing activation maps,

obtained by exploiting the information of the neurons electrophysiological signals.




Time: 17:50 - 19:30

Room: PS2

72

A PORTABLE LOW VISION AID

Urena R., Morillas C., Pelayo F.

Abstract: This work proposes a customizable and portable aid system for low vision im-

paired. The system aims to transform images taken from the patient’s environment and tries to

convey the best information possible through his visual rest, applying various transformations

to the input image and projecting the processed image on a head-mounted-display, HMD. The

main transformations are contrast and edge enhancement and tone-mapping. To achieve real

time performance (over 25 frames per second) on a lightweight platform we use portable devices

based on a GPU NVIDIA ION 2 and on a XILINX SPARTAN III FPGA




Time: 17:50 - 19:30

Room: PS2

73

RANDOM DEPTH: THREE FOR THE PRICE

OF ONE

Velasco-Forero S.

Abstract: Random projection depth provides a center-outward ordering of vectors in vector

space. We explore some theoretical properties and practical application in salient detection,

multivariate mathematical morphology and robust descriptor of non-rigid 3D shapes.




Time: 17:50 - 19:30

Room: PS2

74

MACHINE LEARNING FOR TARGET DETEC-

TION

Vink J.P.

Abstract: Deterministic approaches to detect targets (e.g. compression artifacts, skin or

nuclei) generally fail to cope with the highly varying appearance of targets. Machine learning

strategies like AdaBoost offer more flexibility and have the ability to generalize. We developed

a framework to train a detector with high performance and low computational cost. To this

end, we modified AdaBoost to include awareness of computational feature cost by introducing

a bias towards previously selected features.




Time: 17:50 - 19:30

Room: PS2

75

CAMERA-BASED ANALYSIS OF ROTARYKILNS

Waibel P., Matthes J., Keller H.B.

Abstract: Rotary kilns are industrially used in processes with high energy consumption

such as cement production or metal recycling. Our goal is to optimize the process control by

utilizing new camera-based features out of the inside of the kilns. Here, a method to detect the

repose and the filling angle of the solid bed is presented. At first, the inner kiln wall is spatially

transformed to a rectangle. The segmentation is accomplished with a level set based algorithm

by using intensity, motion and shape information. Both angles can be extracted directly.




Time: 17:50 - 19:30

Room: PS2

76

INVESTIGATIONS ONACTIVE SENSORS FOR

COMPUTER VISION

Weinmann M.

Abstract: The perception of 3D information about the local environment is still of great

interest. Recent developments show that new types of active sensors allow for a simultaneous

capturing of both range and image data, and they are not restricted on the assumption of a

static scene. Thus, these active sensors offer new possibilities for applications like surveillance,

navigation of autonomous vehicles, object tracking, object recognition, scene reconstruction or

scene interpretation.




Time: 17:50 - 19:30

Room: PS2

77

VISUAL LANDMARK-BASED OUTDOOR LO-

CALIZATION FOR MAVS

Wendel A., Irschara A., Bischof H.

Abstract: Highly accurate localization of a micro aerial vehicle (MAV) with respect to a

scene is important for a wide range of applications, in particular surveillance and inspection. We

introduce approaches for robust reconstruction of suitable visual landmarks, for the alignment

in a world coordinate system, and for fast monocular visual localization based on the concept

of virtual views in 3D space. Our system outperforms not only state-of-the-art visual SLAM

but also consumer-grade GPS systems.




Time: 17:50 - 19:30

Room: PS2

78

STRUCTURE BASEDMOSAICKING OF AERIAL

IMAGES FROM LOW ALTITUDE OF NON-

PLANAR SCENES

Wischounig-Strucl D., Quartisch M., Rinner B.

Abstract: We estimate the depth structure of sceneries in aerial images captured by small-

scale UAVs to improve the mosaicking of an orthographic overview image. Initial image trans-

formations derived from inaccurate position and orientation data are enhanced by the camera

pose obtained using Bundle Adjustment. Corresponding points are then selected on a common

ground plane to find accurate image transformations. The resulting mosaick preserves distances

and minimizes distortions and is immediately presented and optimized incrementally if more

images are considered.




Time: 17:50 - 19:30

Room: PS2

79

FACADE EXTRACTION FROMOBLIQUE AIR-

BORNE IMAGES

Xiao J., Gerke M., Vosselman G.

Abstract: Facades are seldomly used for building detection due to lack of information. In

this research a new approach on building faade extraction using airborne oblique images solely is

developed. Facades are detected using edge direction and height gradient. The former is carried

out on single image while the latter is extracted from an image pair using dense matching. One

image pair is sufficient to extract faade facing their viewing direction. Experiments on four

directions have shown that the extracted faade are quite reliable, and it can be used for further

building detection.




Time: 17:50 - 19:30

Room: PS2

80

TRACKING VIA LOCAL PATCHES AND HI-

ERARCHICAL SAMPLING

Yi K. M., Kim S. W., Jeong H., Choi J. Y

Abstract: To track objects showing partial occlusions and non-rigid deformations in real-

time, we propose a tracking method based on sequential Bayesian inference. The proposed

method is consisted of two parts: (1) modeling the target object using elastic structure of local

patches for robust performance; and (2) efficient hierarchical sampling method to obtain an

acceptable solution in real-time. The method is tested on a number of image sequences with

occlusion and non-rigid deformation.




Time: 17:50 - 19:30

Room: PS2

81

LATENT FINGERPRINT ENHANCEMENT

Yoon S., Jain A. K.

Abstract: Fingerprints have been used to identify persons for almost 100 years. One of

the irreplaceable functionality of fingerprint recognition is its capability to link partial prints,

called latents, found at crime scenes to suspects whose fingerprints are previously enrolled in

a fingerprint database. Due to their poor quality, automatic feature extraction and matching

of latents are challenging problems. We have proposed a semi-automatic latent enhancement

algorithm to provide visually enhanced latents to examiners for manual markups and improve

automatic matching performance.




Time: 17:50 - 19:30

Room: PS2

82

POSTER SESSION BOOKLET - iplab.dmi.unict.it · latent topic model based on nonnegative matrix...

Documents

Transcript of POSTER SESSION BOOKLET - iplab.dmi.unict.it · latent topic model based on nonnegative matrix...