Obstacle and Planar Object Detection using Sparse 3D Information for a

Smart Walker

S´everine Cloix

1,2

, Viviana Weiss

, Guido Bologna

, Thierry Pun

and David Hasler

Vision Embedded Systems, CSEM SA, Jaquet Droz 1, Neuchˆatel, Switzerland

Computer Science Department, University of Geneva, Route de Drize 7, Carouge, Switzerland

Keywords:

Binocular Stereo Vision, Sparse 3D Map, Obstacle Detection, Object Detection, Boosting, Elderly Care.

Abstract:

With the increasing proportion of senior citizens, many mobility aid devices have been developed such as the

rollator. However, under some circumstances, the latter may cause accidents. The EyeWalker project aims

to develop a small and autonomous device for rollators to help elderly people, especially those with some

degree of visual impairment, avoiding common dangers like obstacles and hazardous ground changes, both

outdoors and indoors. We propose a method of real-time stereo obstacle detection using sparse 3D information.

Working with sparse 3D points, in opposition to dense 3D maps, is computationally more efﬁcient and more

appropriate for a long battery-life. In our approach, 3D data are extracted from a stereo-rig of two 2D high

dynamic range cameras developed at the CSEM (Centre Suisse d’Electronique et de Microtechnique) and

processed to perform a boosting classiﬁcation. We also present a deformable 3D object detector for which the

3D points are combined in several different ways and result in a set of pose estimates used to execute a less

ill-posed classiﬁcation. The evaluation, carried out on real stereo images of obstacles described with both 2D

and 3D features, shows promising results for a future use in real-world conditions.

1 INTRODUCTION

To help in their mobility, millions of senior citizens

use mobility aids such as rollators. But these devices

may fail to help or, even worse, can cause accidents.

This occurs when the user misjudges the nature or

the extent of some obstacles, which can happen in

any kind of familiar or unknown environments. Vari-

ous prototypes of “intelligent walkers”, developed to

answer these issues, are usually motorized and pro-

grammed to plan routes and to detect obstacles with

active or passive sensors. However, such aids are usu-

ally unaffordable or at a prototype level, hence, the

user might be reluctant to such complex systems. Fi-

nally, their use is often limited to indoor situation due

to their weight and their short battery life.

Unlike the current trend, the EyeWalker project

aims to develop a low-cost, ultra-light computer

vision-based device for users with mobility problems.

It is meant to be an independent accessory that can be

easily ﬁxed on a standard rollator and with a daylong

autonomy. Our device will warn users of potentially

hazardous situations or help to locate a few particu-

lar objects in miscellaneous environments and under

widely varying illumination conditions. The users we

initially target are elderly persons that still live rela-

tively independently. To meet the requirement of de-

signing a device that helps rollator users in their daily

activities, we work on ground change and obstacle de-

tection. The former is built on (Weiss et al., 2013) and

the latter is based on (Cloix et al., 2013). We present

obstacle and object detection methodsbased on boost-

ing classiﬁcation using sparse 3D data. Our 3D sparse

maps are built from stereo images captured from two

high dynamic range cameras that can handle bright

illumination, indoors or outdoors.

This paper is organized as follows: Section 2 de-

scribes relevant examples related to the state-of-the-

art in stereo computer vision; Section 3 describes our

approach; the hardware setup used for the evaluation

of our detectors, the discussed results and the future

work are detailed in Section 4 followed by the con-

clusion in Section 5.

2 RELATED WORK

The deﬁnition of an obstacle highly depends on the

domain the detector is developed for. For driving as-

sistance, an obstacle will be any object standing on

292

Cloix S., Weiss V., Bologna G., Pun T. and Hasler D..

Obstacle and Planar Object Detection using Sparse 3D Information for a Smart Walker.

DOI: 10.5220/0004661602920298

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 292-298

ISBN: 978-989-758-004-8

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

a dominant ground surface (Broggi et al., 2011). In

the ﬁeld of health care rehabilitation, it would rather

be a static or moving object in the walking path at

any height from the ground to head-level (Ong et al.,

2013).

The domain of application also leads to the choice

of the sensors. In robotics, devices mainly devel-

oped for an outdoor usage to scan the frontal scene

are equipped with ultrasound, laser or radar (Lacey

and Rodriguez-Losada, 2008). Range ﬁnders allow

extracting a dense depth map of the captured scene.

But this equipment is expensive and power consum-

ing, which is of main concern in health care assistance

ﬁeld, not to mention the weight. It justiﬁes why stereo

vision (Rodr´ıguez et al., 2012) or IR sensors are more

employed (Ong et al., 2013).

As far as computer vision is concerned, multi-

view vision is a topic in which intensive research

has been conducted for the last half century (Seitz

et al., 2006). It allows 3D reconstruction thanks to

the hardware performance enabling real-time applica-

tions. Commercial cameras like the Bumblebee2

and

the Microsoft Kinect

are also the main catalysers of

this research growth by providing dense 3D maps.

Focusing on binocular vision, several methods

were developed to detect obstacles such as digital ele-

vation map (DEM) (Oniga and Nedevschi, 2010) and

occupancy grids (Rodr´ıguez et al., 2012) (Perrollaz

et al., 2010). Stereo information is also often used for

distance computation after an obstacle is detected in

one of the two images (Jos´e et al., 2011). Moreover a

majority of the most recent stereo vision-based obsta-

cle detectors use dense 3D maps to cope with stereo

mismatch. Regarding sparse 3D maps, (Toulminet

et al., 2006) extract the desired 3D features to detect

vehicles in a very constrained environment given by

the application domain. This lets the sparse 3D in-

formation usage barely exploited to detect everyday

obstacles with various sizes and shapes.

The contribution of this paper lies in the use of

sparse 3D information to develop a novel strategy of

obstacle and object detection by boosting classiﬁca-

tion. Extracting a few stereo data, computationally

more efﬁcient, allows the obstacle detector to run in

real-time in any illumination conditions and for a long

time thanks to our low power-consuming cameras.

3 METHODS

We want to implement two binary classiﬁcation-based

detectors, one for general obstacles and one for spe-

http://ww2.ptgrey.com/stereo-vision/bumblebee-2

http://www.xbox.com/en-US/kinect

ciﬁc objects. The former is described in Section 3.2

and the latter is explained in Section 3.3. Since a 3D

point cloud lets a priori suppose the presence of an

obstacle, the detectors use information from left and

right pictures to compute sparse 3D maps. The imple-

mented classiﬁcation algorithm is the AdaBoost (Fre-

und and Schapire, 1997) that generates a linear com-

bination of weak classiﬁers.

3.1 Stereo Correspondence

In sparse 3D maps, only very few points have their

distance known. Those points are usually with spe-

ciﬁc characteristics like feature points (Bay et al.,

2008) or corners (Harris and Stephens, 1988). The

stereo correspondence can be carried through various

methods among which there are cross-correlation or

descriptor matching, the latter being the one we chose

to meet real-time and battery-life requirements.

Knowing the intrinsic and extrinsic parameters of

the stereo-rig obtained at the calibration phase and the

disparity, i.e. x

− x

, with x

and x

the respective x-

coordinate of the observed point in the left and right

undistorted and rectiﬁed pictures, the three space-

coordinates of a point can be computed. To extract

a 3D point cloud from each pair of pictures, we pro-

ceed as follows: (i) corners are detected (Harris and

Stephens, 1988) in both left and right pictures, pre-

liminarily equalized and Gaussian blurred; (ii) they

are described with BRIEF descriptor (Calonder et al.,

2010) and matched using the Hamming distance; (iii)

lens distortion is taken into account in an intermediate

step to compute the 3D coordinates of each corner.

3.2 Obstacle Detection

The ﬁrst family of features employed to describe the

cloud of points is a 3D box (Figure 1, top left). This

feature, deﬁned by six parameters, the three location

coordinates and the dimensions, returns the number

of points inside the box.

The second family of features computes the dif-

ference of luminance in deﬁned areas between the left

picture and the right picture (Figure 1, top right). The

parameters are the size of the region, the location in

the left image and the shift, ∆x, on the x-axis of the lo-

cation in the right image. The epipolar-constrained ar-

eas thus deﬁned allow computing the sum of the pixel

values.

The last family of features, depicted in Figure 1

(bottom), is similar to the Haar ﬁlters used by (Viola

and Jones, 2004). In our case, the ﬁlters are deﬁned

by a black zone around the centre of the left image.

Thus the parameters to deﬁne a feature are its shape

ObstacleandPlanarObjectDetectionusingSparse3DInformationforaSmartWalker

293

Figure 1: Visualisation of the obstacle detector features.

(top left): a 3D box in which we count the number of points;

(top right) a pair of areas for the computation of difference

of luminance; (bottom) Haar ﬁlter with the third shape de-

scribed in the sketch on the right.

(black square, horizontal or vertical band), the size of

the patch, the ratio of the central zone to the patch and

the location of the patch in the image.

3.3 Planar Object Detection

We aim at detecting planar objects in a 2D image by

combining the 3D points in several different ways, re-

sulting in a set of pose estimates for a less ill-posed

classiﬁcation. We assume the object is represented by

a 3D generic and normalized model M. From the 3D

point cloud, we estimate a list, Θ, of poses, θ ∈ Θ, of

the object. A feature family, F, is deﬁned in the model

space: the pose estimate allows the features f ∈ F de-

ﬁned on the model to “stick” to objects of any shape

and position in the scene, as depicted on Figure 2.

The planar object model used is a front facing

square with a deﬁned side size, ζ, and described

by a triplet of ordered corners {P

(ζ,0,0); P

(0,0,0);

(0,ζ,0)}. For N matched corners extracted as de-

scribed in Section 3.1, the number of triplets is O(N

To reduce this number of pose estimates, we use the

FLANN algorithm (Muja and Lowe, 2009) to keep

triplets that have two points horizontally aligned and

two vertically aligned. The pose estimation matrix is

extracted by matching P

to p

, i = [0;1;2] like shown

in Figure 2.

Three families of features were deﬁned as follows:

(i) the cosine of the angle deﬁned by the three points;

(ii) the scale, over X, Y or the ratio X/Y; (iii) the

Haar-like ﬁlters. The latter is deﬁned by the follow-

Figure 2: Model correspondence to a triplet of points that

represents a pose of a cabinet door, with a deformable Haar-

like feature, the model being a front facing square.

Figure 3: Haar like feature for the detection of the cabinet

door.

ing parameters, (a) the ﬁlter: eight different shapes

that represent the edges and the corners, shown in Fig-

ure 3; (b) the size of the ﬁlter; (c) the position in the

model (X,Y) in order to localize the scanning window

around the estimated pose. The return value uses the

pixel values of the ﬁlter corners from the left picture:

H =

∑

0≤i≤3

× p

(1)

where s

is the sign deﬁned by the shape of the ﬁl-

ter (+ if black, − if white) and p

the pixel value of

corner i.

4 EXPERIMENTS & DISCUSSION

The evaluation of the methods are performed on static

stereo pictures acquired with an experimental setup

detailed in Section 4.1 in order to measure the perfor-

mance of our general obstacle and object detectors.

4.1 Hardware Setup

The rollator is equipped with a calibrated stereo-rig

of two cameras developed by the CSEM, called icy-

Cams. The 20 cm-baseline rig is tilt at 20 degrees

from the horizon and ﬁxed on the rollator at 67 cm

from the ground. These cameras are characterized by

their high dynamic range and their logarithmic com-

pression that allows coping with bright illumination.

Their power consumption is very low, which is of

great importance since the ﬁnal device has to have a

battery life of one day. The focal lengths are equal

to 3.8 mm and the resolution is 320x240 pixels, 14e-

6 m per pixel, resulting in a vertical view angle of

47.7 degrees (Figure 4). The cameras were connected

Figure 4: Actual hardware setup allowing viewing obstacles

at less than two meters and at the waist-level up to head-

level.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

294

to a laptop computer through a usb-powered Ethernet

switch for the acquisition of stereo frames in various

environments.

4.2 Obstacle Detection

To evaluate the performance of our binary classiﬁer,

experiments were conducted on a data set of eleven

types of obstacles (cf. Figure 5). The training set

was composed of 3107 pairs of pictures. The test

set had 732 pairs of frames of similar obstacles taken

in places different from the training set. The images

were taken from real scenes, both indoors and out-

doors, with a frame rate of approximately 7 f ps and

walk speed of about 0.5 m/s. The two classes are de-

ﬁned as follows: a frame is labelled positive if there

is an obstacle on the pathway at less than two meters

from the cameras; the frame is negative otherwise.

The indoor obstacles are: (i) a cupboard; (ii) a desk

chair; (iii) a corridor wall; (iv) a white board; (v) a

doorframe not centered on the path; (vi) a coffee ma-

chine in a kitchen; (vii) a dish-washer. The outdoor

obstacles are: (i) a bush; (ii) a street lamp pole; (iii)

a parking fence; (iv) a road sign. The non-obstacle

data are made of the pictures of long corridors, street,

ofﬁce and kitchen views.

For the sake of comparison with our approach, we

implemented a baseline detector that counts the num-

ber of points into a warning area. The decision was

made according to the number of points present into a

frontal cube centered on the path at 0.5 meter from

the cameras, 1.5 meter deep, 0.8 meter high and a

width varying between 0.5 and 0.8 meter. The results

demonstrate it is not possible to have more than 75%

true-positive rate without less than 30% of false alarm

Figure 5: Samples of left-camera pictures for the obstacle

data set:(ﬁrst two rows) pictures labelled as negative (no

obstacle or above two meters from the cameras); (last two

rows) pictures labelled as positive,(obstacle is at less than

two meters from the cameras).

(cf. Figure 6).

The obstacle boosting classiﬁer was evaluated by

generating four kinds of classiﬁers: three were trained

with only one of the feature families described in Sec-

tion 3.2 and one with all the families. The Figure 7

shows the average performance of each feature fam-

ily independently and of the three families together.

The 3D box family presents a better performance

than the other features and the simple detector. Be-

sides, the addition of the other features enables a sig-

niﬁcant improvement (true-positive rate from 69.8%

to 75.4% at 10% false-positive rate). Even some of

the classiﬁers generated with all the feature families

give more than 83% true-positive rate for less than

10% false-positive rate. The 3D boxes return value

actually represents a density of points, which is ex-

pected to be high for positive frames, i.e. with an

obstacle. The resulting weak classiﬁer being char-

acterised by its threshold, optimized at the training

phase, it explains why this boosting classiﬁer per-

forms better than the basic detector.

For the second family of features, the return value

is the difference of luminance. Such features should

return small scores when the two patches cover the

same object. The variability in the distances of the

training obstacles could explain the difﬁculty to the

training framework to choose the best luminance fea-

tures.

Finally, the Haar ﬁlter score should be high when

the ﬁlter is centered on a dark or light patch of colour

representing an obstacle. Here as well, the actual per-

formance of such a detector can be explained by the

variability of the patch size and shape in the train-

ing set. As a result, the classiﬁers composed of only

one of the two latter families perform worse than the

one of 3D boxes. A solution would be to grow the

database for a better distribution, especially when we

see that the classiﬁer built with all the feature families

performs better than the one of 3D boxes.

The tests were carried out ofﬂine on a 2.93

GHz/Linux x86-64 desk computer. The actual pro-

gram uses a single processor. In these conditions the

detection took 244 ms. This computation time cor-

responds to 4 f ps and can be improved by code op-

timisation and by the use of parallelization for com-

puting each weak classiﬁer score. On an embedded

system the use of a GPU would improve the time per-

formance.

4.3 Cabinet Door Detection

After having detected obstacles with a certain degree

of accuracy, another user requirement is to detect spe-

ciﬁc objects with the use of vision features provided

ObstacleandPlanarObjectDetectionusingSparse3DInformationforaSmartWalker

295

to a boosted classiﬁer. To evaluate the approach, the

planar object chosen is a kitchen cabinet door. Pre-

liminary experiments are performed on a data set of

stereo pictures taken in a kitchen show room (Fig-

ure 8). The positives are annotated with triplets. Un-

known zones are also marked. From this annotation,

negatives per pair of images are extracted by the

triplet detector and according to the distance from the

positive samples and the unknown zones. The train-

ing is done with 80% of the data set, the remaining

20% being reserved for the test, randomly selected.

The annotation is rigid in a sense that the triplet has

to represent the door so that the vertical segment is on

the hinge.

Figure 9 shows the performance (66.7% true-

positive rate at 10% false-positive rate) and high-

lights several limitations of the pose estimator chosen.

Firstly, the actual pose estimator is not robust to the

lack of robustness of the corner detection-matching

step. The insufﬁcient precision of the corner detector

and its parametrization depends on the distance to the

object and also on the picture quality. The latter was

lowered by the poor illumination of the scenes and the

cameras resolution. When the desired 3D corners are

not detected (because not detected on one or both pic-

tures or not matched and thus discarded), the object

is not localized. As a result, about 80% of the tested

doors were detected prior classiﬁcation. Secondly, we

have to look at the way negative samples used for the

classiﬁer training were extracted. Any triplet that has

at least one point far from the ones of a positive triplet

is considered negative. As a consequence, a negative

triplet can have up to two points that also belong to a

positive triplet. Thus the feature vector of the nega-

tive sample can be partially similar to the one of the

positive sample.

The ambiguity on closed doors is important since

kitchens include several doors on the same plane. To

reduce this ambiguity, new experiments were con-

ducted on opened doors. The annotation is made

broader to make a rotation-invariant classiﬁer: each

door is annotated once and the other seven triplets

deﬁning the door are generated and added into the

training data set. By focusing on opened doors,

Figure 6: Performance of the basic detector: (left) ROC

curve (right); precision-recall curve.

Figure 7: Performance of the boosting obstacle detector

built with the 3D box features only (red),with the Haar ﬁl-

ter features only (green), with the luminance difference fea-

tures only (blue), with all the features (purple): (left) ROC

curves; (right) Precision-recall curves.

the performance improved: the localisation rate went

from 85.4% to 96.1% and the true-positive rate from

66.5% to 83% at 10% false-positive rate.

From a point cloud, the complexity makes the

number of hypotheses too big, leading to an unsat-

isfying false-positive rate and a very low precision.

Despite the actual pose estimator, the classiﬁer gives

good results. However, the problem is complex be-

cause of the number of tests (about 150’000 per pair

of images) required for the localisation of a potential

door, the number of degrees of freedom regarding the

orientation estimation and the signiﬁcant appearance

variation (size, presence of handle or not, texture).

4.4 Future Work

The main focus is to improve the current general ob-

stacle detector recall and precision. To do so, we will

optimize the stereo matching by improving the image

pre-processing in order to reduce the noise stereo mis-

matching introduces. The actual experiments being

carried out on static stereo images, we intend to intro-

duce temporal information such as Kalman ﬁlters or

optical ﬂow to ﬁlter the 3D information. Indeed, we

suppose the knowledge of past frames can improve

the actual results. The data set size revealing the ef-

fect of the variability issue on the classiﬁcation, it has

to be increased.

Last but not least, the object detector evaluation

results suggest reducing the complexity of the prob-

lems encountered with the cabinet doors. The main

issues to solve are twofold: (i) to be able to detect all

Figure 8: Samples of the data set of cabinet doors.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

296

Figure 9: Performance of the boosting door detector built

with the Haar-like features: (blue) Performance of on all

doors; (red) Performance of on opened doors; (left) ROC

curves; (right) Precision-recall curves.

the desired corners among a few to limit the number

of tests required to meet real-time conditions; (ii) re-

view the hypotheses to restrict the variability of the

category deﬁning cabinet doors.

5 CONCLUSIONS

To help rollator users to avoid common dangers with a

computer vision-based device, we introduced two de-

tectors depending on features including 3D and stereo

data: one for general obstacles located at waist-level

and above, and a second one for speciﬁc objects. Both

detectors are based on boosting classiﬁcation. The

obstacle detector mixes three kinds of features among

which two require stereo information: 2D Haar ﬁl-

ters, 3D boxes and luminance comparison between

the stereo pictures. The experiments show promis-

ing results that can be improved by the future work

mentioned in Section 4.4. The deformable 3D object

detector, mainly composed of Haar-like features, re-

mains an interesting strategy, despite the evaluation

results, the actual pose estimator having to be thought

over to make the binary classiﬁer more robust and

faster.

ACKNOWLEDGEMENTS

This project is supported by the Swiss Hasler Foun-

dation Smartworld Program, grant Nr. 11083. We

thank our end-user partners: the FSASD, Fondation

des Services d’Aide et de Soins Domicile, Geneva,

Switzerland; EMS-Charmilles, Geneva, Switzerland;

and Foundation “Tulita”, Bogot´a, Colombia. The cab-

inet door database was made in Schaer Cuisines SA,

Neuchˆatel, Switzerland.

REFERENCES

Bay, H., Ess, A., Tuytelaars, T., and Van Gool, L. (2008).

Speeded-up robust features (surf). Computer vision

and image understanding, 110(3):346–359.

Broggi, A., Buzzoni, M., Felisa, M., and Zani, P. (2011).

Stereo obstacle detection in challenging environ-

ments: the VIAC experience. In International Con-

ference on Intelligent Robots and System, pages 1599–

1604, San Francisco, CA, USA. IEEE Computer So-

ciety.

Calonder, M., Lepetit, V., Strecha, C., and Fua, P. (2010).

Brief: binary robust independent elementary features.

In Proceedings of the 11th European conference on

Computer vision: Part IV, pages 778–792, Berlin,

Heidelberg. Springer-Verlag.

Cloix, S., Weiss, V., Guido, B., Pun, T., and Hasler, D.

(2013). Object detection and classiﬁcation using

sparse 3d information for a smart walker. In Swiss Vi-

sion Day 2013, Poster Session, ETH Z¨urich, Switzer-

land.

Freund, Y. and Schapire, R. E. (1997). A decision-theoretic

generalization of on-line learning and an application

to boosting. Journal of Computer and System Sci-

ences, 55(1):119–139.

Harris, C. and Stephens, M. (1988). A combined corner and

edge detector. In Alvey vision conference, volume 15,

page 50. Manchester, UK.

Jos´e, J., Farrajota, M., Rodrigues, J. M., and du Buf,

J. (2011). The smartvision local navigation aid for

blind and visually impaired persons. Available at

http://hdl.handle.net/10400.1/892.

Lacey, G. and Rodriguez-Losada, D. (2008). The evolution

of guido. Robotics & Automation Magazine, IEEE,

15(4):75–83.

Muja, M. and Lowe, D. G. (2009). Fast approximate nearest

neighbors with automatic algorithm conﬁguration. In

International Conference on Computer Vision Theory

and Application, pages 331–340. INSTICC Press.

Ong, S. K., Zhang, J., and Nee, A. Y. C. (2013). Assistive

obstacle detection and navigation devices for vision-

impaired users. Disability and Rehabilitation: Assis-

tive Technology, pages 1–8.

Oniga, F. and Nedevschi, S. (2010). Processing dense stereo

data using elevation maps: road surface, trafﬁc isle,

and obstacle detection. Vehicular Technology, IEEE

Transactions on, 59(3):1172–1182.

Perrollaz, M., Spalanzani, A., and Aubert, D. (2010). Prob-

abilistic representation of the uncertainty of stereo-

vision and application to obstacle detection. In In-

telligent Vehicles Symposium (IV), 2010 IEEE, pages

313–318. IEEE.

Rodr´ıguez, A., Yebes, J. J., Alcantarilla, P. F., Bergasa,

L. M., Almaz´an, J., and Cela, A. (2012). Assisting

the visually impaired: Obstacle detection and warning

system by acoustic feedback. Sensors, 12(12):17476–

17496.

Seitz, S. M., Curless, B., Diebel, J., Scharstein, D., and

Szeliski, R. (2006). A comparison and evaluation of

multi-view stereo reconstruction algorithms. In Con-

ference on Computer Vision and Pattern Recognition,

volume 1, pages 519–528. IEEE Computer Society.

ObstacleandPlanarObjectDetectionusingSparse3DInformationforaSmartWalker

297

Toulminet, G., Bertozzi, M., Mousset, S., Bensrhair, A.,

and Broggi, A. (2006). Vehicle detection by means of

stereo vision-based obstacles features extraction and

monocular pattern analysis. Image Processing, IEEE

Transactions on, 15(8):2364–2375.

Viola, P. and Jones, M. J. (2004). Robust real-time face

detection. International journal of computer vision,

57(2):137–154.

Weiss, V., Cloix, S., Guido, B., Hasler, D., and Pun, T.

(2013). A ground change detection algorithm using

colour and texture for a smart walker. In Swiss Vision

Day 2013, Poster Session, ETH Z¨urich, Switzerland.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

298