Automatic Object Segmentation on RGB-D Data using Surface

Normals and Region Similarity

Hamdi Yalin Yalic and Ahmet Burak Can

Computer Engineering Department, Hacettepe University, Ankara, Turkey

Keywords: Object Segmentation, RGB-D Data, Region Growing, Surface Normals.

Abstract: In this study, a method for automatic object segmentation on RGB-D data is proposed. Surface normals

extracted from depth data are used to determine segment candidates first. Filtering steps are applied to depth

map to get a better representation of the data. After filtering, an adapted version of region growing

segmentation is performed using surface normal comparisons on depth data. Extracted surface segments are

then compared with their spatial color similarity and depth proximity, and finally region merging is applied

to obtain object segments. The method is tested on a well-known dataset, which has some complex table-top

scenes containing multiple objects. The method produces comparable segmentation results according to

related works.

1 INTRODUCTION

With the easy accessibility of powerful depth

sensors, depth maps have become available as well

as RGB images for any indoor scene. Along with

these developments, scientific studies on 3D object

recognition, segmentation, and tracking using RGB-

D data have become popular in the recent years. In

the field of robotic vision, depth sensors are used

due to their low costs and mobility features.

Object detection and segmentation are very

useful tasks for robotic grasping, as well as for the

recognition and classification of the objects.

Therefore, segmentation is an important step for

robotic vision applications and used as an input for

the next steps. In situations where segmentation is

difficult with RGB information, more successful

results can be achieved with RGB and depth

information used together. Depth information

enables to develop image processing approaches

without being affected by color, texture and lighting

features of the objects. Furthermore, depth

information gives more cues about shape structures

and contributes to a better understanding of the

scene.

This paper introduces an automatic object

segmentation method for tabletop scenes, which may

contain mixed and complex sets with multiple

objects in it. In this method, the segmentation

process is performed in two main steps. Firstly,

image segmentation is performed on depth data

using an adapted version of region growing

algorithm, based on surface normal similarity and

smoothness. Then segmented regions are grouped

using depth proximity and their RGB features based

on color correlogram (Huang et al., 1997) and

histogram similarity to obtain objects. In order to

obtain better and accurate results, point cloud

filtering steps are applied and smoothed depth data

are obtained. The proposed method works unsuper-

vised and automatically segments different objects

in the scene. It produces meaningful segmentation

results comparable with previous studies.

The outline of the paper is as follows: Section 2

gives the related works in the literature about object

segmentation on RGB-D data. Section 3 introduces

the proposed method. Section 4 gives the

experimental results. Finally, Section 5 concludes

with the given results.

2 RELATED WORK

Studies on image and object segmentation using

RGB-D data has begun in the last seven years. Liu

(Liu et al., 2013) studied automatic object segmenta-

tion using RGB-D cameras. In their camera with

depth sensor (like Microsoft Kinect), there were

Yalic, H. and Can, A.

Automatic Object Segmentation on RGB-D Data using Surface Normals and Region Similarity.

DOI: 10.5220/0006617303790386

In Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2018) - Volume 4: VISAPP, pages

379-386

ISBN: 978-989-758-290-5

379

some gaps and absences in the depth map. This

situation resulted in object boundaries not being

properly determined. To solve this problem, they

developed a three-way (trilateral) filter that includes

distance, RGB values, and boundary information.

They apply warping, error cleaning, and affine

mapping to eliminate holes caused by depth sensor

data acquisition. After this step, they used the

probabilistic boundary detector (Pbd) component. Pb

(probability of boundary) is a method that uses only

color information and extracts the boundary-priority

map of the object according to color differences. By

adding depth attribute as a parameter, it calculates

the probability that a pixel is at the boundary and its

orientation. In the last step, they segmented the

object using graph-cut and separated it from the

scene. During the evaluation of their work, they

measured the contribution of the depth information

that they added to the basic methods. They have

increased the performance of the segmentation.

Mishra (Mishra and Aloimonos, 2012, Mishra et

al., 2012) developed a segmentation strategy that

separates simple objects from the scene by using

color, texture and depth knowledge. They defined

the simple object as a compact zone surrounded by

depth and contact boundary. They improved on the

fixation-based segmentation method (Mishra et al.,

2009), which they found the most suitable closed

contour around the given point in the scene. In their

recent work, they have proposed a fixing strategy

that selects points from within an object, as well as a

method that allows the selection of closed curves

only for objects. The visual cues (color, depth)

around each edge pixel indicate whether that pixel is

at the object boundary. This information is kept in

the probabilistic boundary edge map. An edge finder

(Martin et al., 2004) with local brightness, texture

and color cues, was used when calculating this map.

If the segmentation process is examined

sequentially, it is seen that first a probabilistic

boundary edge map is obtained. By selecting the

most probable edge pixels in this map, the object

side is determined. By considering the object sides

of the boundaries in this map, fixation points are

selected on that side. Closed curves are obtained

from the determined points by the fixed-based

segmentation method and the resultant object

segments are obtained. They have studied the

experimental results in a comprehensive dataset (Lai

et al., 2011), quantitatively and qualitatively. During

quantitative analysis, they measured the

segmentation accuracy of segmented objects as a

single closed region and achieved a success rate of

over 90%.

Richtsfeld and colleagues have performed

numerous studies (Richtsfeld et al., 2012a),

(Richtsfeld et al., 2014) for object segmentation on

RGB-D data. In their study on the implementation of

Gestalt principles for object segmentation

(Richtsfeld et al., 2012b), they have defined the

relationships between surface patches on a 3D image

based on Gestalt principles in order to build a

learning-based structure. The scene structure is

rapidly abstracted by plane fitting on a 3D point

cloud. The fast and commonly used RANSAC

method is used for this purpose. But for curved

objects, it is necessary to soften and bend the

surfaces. At this point, they used the mathematical

construct called NURBS (non-uniform rational B-

splines), which is widely used in the field of

graphics. This structure enables to display all kinds

of conic sections (spherical, cylindrical, ellipsoid,

etc.). The plane placed on the point cloud is matched

to the cloud by minimizing the length of the nearest

point. On the geometric structure, a final model

selection is applied to determine the surface patches

to be used in object segmentation.

In their article describing perceptual grouping for

object segmentation on RGBD data (Richtsfeld et

al., 2014), they presented a comprehensive study

combining previous works. Once the surfaces have

been determined, the relationships between adjacent

surfaces, which are based on the Gestalt principles

mentioned above, are calculated to group them.

These features are surface color, the similarity of

relative size and texture amount, color similarity at

3D surface boundaries, average curvature and

curvature variance at 3D surface boundaries, average

depth and variance at 2D surface boundaries. Two

attribute vectors have been defined from the

previous relations: neighbouring and non-

neighbouring. These two vectors were used for

training SVM on hand-labeled RGB-D image sets.

The surface pairs of the same object are selected as

positive samples, while the surface samples of two

different objects or the object-background pairs are

negative samples. In the decision making, SVM

gives the probability value of each vector, as well as

producing binary results. They defined a graph

where surface patches represent the nodes and

probability values obtained from SVM represents

the edges. Finally using the graph-cut, object

segmentation was performed. As a result of the

detailed analysis of the segmented objects in

different complex scenes separated by different

categories, they achieved a success rate of over 90%.

They stated that on the same data sets, Mishra et al.

remained around 65%.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

380

3 AUTOMATIC OBJECT

SEGMENTATION

In this paper, an automatic method for segmenting

objects in complex scenes on RGB-D data is

proposed. To accomplish this, an image

segmentation process is performed on the geometric

structure of the objects using the depth information.

Subsequently, object segmentation was performed

using neighbourhood information based on the depth

cue and color properties of the obtained regions.

The overall design of the system is illustrated in

the Figure-1. Details of all steps will be given in the

next sections of this paper.

3.1 Filtering

Depth data have some different features that are not

present in its RGB image, depending on the depth

acquisition device. Irregular density (describes

number of points in a given area), noise, and outliers

are some of the important ones. Previous studies

(Liu et al., 2013) show that point cloud data needs to

be filtered for better segmentation.

In our study, down-sampling like voxelgrid or

morphological operations are not preferred because

they modify the important properties of the point

cloud set. We used filters to eliminate noises and

remove outliers. The pass-through filter in Equation

(1) used for eliminating the areas where the depth

information is missing or larger than a pre-defined

threshold. P denotes the point cloud set for x,y,z

coordinates.













  







    





(1)

After that, radius outlier removal is used to erase

outlier points. Let    be the number of

neighbours for 



   in a radius of r. We

remove the points from the cloud with less than a

given number of neighbours 



within a radius r.

In Equation (2),  denotes the outlier points set and

 denotes the filtered points.

 









  







    





  



 

(2)

Pre-processing step is completed with filtering

and suitable data is obtained for the segmentation

step.

Point Cloud Filtering

RGB Depth

Image Segmentation by

Region Growing using

Surface Normals

Color Similarity and Depth

Proximity Features

Segmented Regions

Segmented Objects

Object Segmentation by

Region Merging

Figure 1: Overall system diagram for automatic object

segmentation.

Automatic Object Segmentation on RGB-D Data using Surface Normals and Region Similarity

381

3.2 Region-growing based

Segmentation using Normals

In our study, we used region-growing based

segmentation to cluster depth data. Since we only

deal with the depth data at this stage, we consider

the angle difference between the surface normals of

the points when growing the regions. Thus, while

smooth surfaces are treated as a single region,

regions with dramatically changing normals are

separated into different segments.

3.2.1 Surface Normal Estimation

Figure 2: (a) Sample RGB image from the dataset

(Richtsfeld et al., 2012a), (b) rotated depth image with

surface mapping, (c) computed surface normals using

depth map.

Surface normals are one of the important attributes

of a geometric surface and widely used in computer

graphics and vision applications. When a geometric

surface is considered, it is usually insignificant to

find a normal direction at a certain point on the

surface, as a perpendicular vector to the surface at

that point. However, since our acquired point cloud

data form a set of point samples on the real surface,

there are two common approaches for calculating the

surface normal. The underlying surface can be

obtained from the point cloud using surface mapping

techniques and compute normals from the mesh.

Another one is using approximations to derive the

surface normal from the point cloud directly. We

used the second approach since it gives sufficient

surface normal vector of the points about belonging

to the same surface.

We used widely known method (Hoppe et al.,

1992) for surface normal estimation. Briefly, it

selects a subset of points Φ inside point cloud P

from the neighbourhood of p. Then it fits local plane

through Φ and finally computes the normal vector N

of the plane. Our method computes the normal

vectors locally using six neighbouring points. The

number of neighbouring points is optimal for our

problem, and incrementing neighbour amount does

not have any positive affect on segmentation

according to our experiments. Figure-2 shows the

illustrations of computed surface normals from the

depth data of adjacent objects in the dataset.

3.2.2 Region-growing Segmentation

The aim of the algorithm is to merge similar points

that are similar enough in terms of smoothness.

Inputs of the algorithm are point cloud data after

filtering, computed surface normal and their

curvature value for each point in the cloud, normal

angle threshold and curvature threshold. Outputs are

the set of regions where each region is a set of points

that are marked as a part of the same surface.

The algorithm is based on region-growing

segmentation that is widely known image

segmentation method. But the main difference is that

it uses angle comparison between the point normals

instead of intensity. A seed point is randomly

selected and added to the set of seeds. The algorithm

finds neighbouring points for every seed point.

Every neighbour point is compared to the angle

between its normal and current seeds normal. If it is

less than the threshold (10/180* in radians, equals

to 10 degrees), then the current point is added to the

current segmented region. After that, every

neighbour is compared for the curvature. If

curvature value is less than the determined threshold

(1.0), then this point is added to the set of seeds.

Current seed is removed from the set. If the set of

seeds becomes empty, that means the algorithm has

grown the region and it is repeated from the

beginning. As a result, we obtain object surfaces in

the depth data that can be an important input for

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

382

object segmentation. Thresholds are determined

according to data set calculations and tuned for

optimal values. Pseudocode of the algorithm for

region growing segmentation based on surface

normal is shown below:

Inputs:

P // Points in the point cloud

N // Surface normal calculated from each point

C // Curvature value of the points





// Normal angle threshold





// Curvature threshold

Set SgtRgn = { } // Segmented region list

Set PointsToGrow = P // Points available for growing

Algorithm:

While PointsToGrow != null do

| SgtRgn

current

= { } // Region segment

| Seed

current

={ } // Set of current seeds

| NextPoint = rand (PointsToGrow)

| Seed

current

= Seed

current

NextPoint

| SgtRgn

current

= SgtRgn

current

NextPoint

| PointsToGrow = PointsToGrow NextPoint

| For i = 0 to size(Seed

current

) do

| | SeedCurrNeig = FindNearNeigbours (Seed

current

[i])

| | For j = 0 to size(SeedCurrNeig) do

| | | P

= SeedCurrNeig[j]

| | | If PointsToGrow.contains ( P

) AND

| | | cos-1( N[Seed

current

[i]] , N[Seed

current

[j]] ) < 



then

current

= SgtRgn

current

 P

] < 



then

current

= Seed

current

P

| | | end if

| | end for

| end for

| SgtRgn = SgtRgn SgtRgn

current

end while

Return SgtRgn

As a result of the algorithm, all the regions are

obtained from the point cloud which are on the same

object surface. Figure-3 shows the segmentation

results after region growing which is mapped and

compared with its original RGB image. As seen

from the output, each surface of the 3D objects

segmented as different regions. So, this yields to an

over-segmentation problem that is one object is

represented with multiple region segments. These

regions need to be merged to obtain single object

instances. There are also small regions segmented

mostly in object boundaries because of surface

normal values. These regions are also handled in the

next region-merging section.

Figure 3: Region growing segmentation result.

3.3 Object Segmentation based on

Spatial Color Similarity of the

Regions

In order to achieve better segmentation performance,

regions have to be merged appropriately for

extracting objects from the scene. Only depth cue

cannot be sufficient at this point. Because, as it can

be seen in Figure 3, it is difficult to decide whether

the boundary points belong to the same object or

different object in complex scenes where there are

multiple adjacent objects. Segmenting an object that

is partially occluded by another object is relatively

easy if there is a distance between them. For these

reasons, in addition to depth information, we have

proposed an approach that adds color information of

regions to the process. Since the surfaces of the

objects in the datasets have complex texture and

color properties, using histograms did not yield

successful results in our experiments.

To extract global color information of segments,

we used Huang's approach (Huang et al., 1997),

which is mostly known for image indexing or

retrieval. It is a widely used image feature called

color correlogram. It parses the spatial correlation of

colors and tolerates large changes in appearance and

shape caused by viewing directions. This property

makes this feature convenient for our work because

different surfaces of the same object have similar

color characteristics.

In our method, the color image obtained from the

region is firstly quantized as 64 colors (4x4x4) in

RGB space. Then, using the 4 pre-defined distance

values as in Huang's work, the neighborhoods at

these distances are determined for each pixel. By

taking the color values at these points, the

correlogram is calculated and transformed into the

1x64 feature vector, resulting in a histogram-like

color Auto-correlogram.

The color correlogram is calculated (denoted by

) and sorted for all segmented regions. At the

final stage, the color correlograms extracted from

the regions and the contact boundaries of these

regions have been taken into consideration in order

to merge the regions. Starting from the first region,

the correlogram vectors are compared ( is the

histogram comparison function, squared Euclidean

distance is used). And it is also checked whether

these regions have contact boundaries. In doing so,

all boundary points (denoted by  function) of

the region are compared to all boundary points of the

other regions, and these two regions are merged if

they share common boundary.  operator specifies

that two regions have mutual boundaries i.e.

Automatic Object Segmentation on RGB-D Data using Surface Normals and Region Similarity

383

boundary points are closer in means of depth

distance. Below algorithm explains region merging

process. These steps are performed iteratively until

all regions are checked.

Algorithm:

Object

= Segment

For i = 0 to size(Region_segments) do

| For j = i + 1 to size(Region_segments) do

| | If ( Bound (Segment

)  Bound (Segment

)   )

AND

| | dist (corr (Segment

), corr (Segment

)) < thresh then

| | | Object

= Object of Segment

| | | Object

Segment

 Segment

| | end if

| end for

end for

After the object segmentation process is

completed, there are still some areas which have few

points, especially in the object boundaries. The

correlogram extracted from these regions does not

give meaningful results. Therefore, these regions are

merged with another surface/region adjacent to

them. This process, which has no effect on the

segmentation quantitative results, has the effect on

quality of the results and it provides a better

representation of the segmented object. The region

with the maximum number of points is not treated as

an object because it represents the plane on which

the objects are standing (mostly surface of the table).

4 EXPERIMENTAL RESULTS

In this section, experimental results of the proposed

method are discussed. Experiments are performed on

four datasets. First one is (Lai et al., 2011) RGB-D

Objects Dataset. It consists of 300 objects in 51

categories. This set is widely used for object

recognition, but we used it only for evaluation of our

object segmentation method on single instance

objects, not for comparison. They also have another

dataset (Lai et al., 2012) called RGB-D Scenes

Dataset which consists of 14 scenes and many

objects with multiple views. The Object

Segmentation Dataset (Richtsfeld et al., 2012a)

contains 111 scenes which consist of boxes,

cylindrical objects, occluded and stack objects with

complex scenes. Willow garage dataset (Aldoma and

Richtsfeld, 2012) contains approximately 160

frames includes mixed objects. All datasets are

moderately realistic and contain table-top scenes

with objects stand on it.

Table 1: Object segmentation accuracy on different scene

categories compared with related works.

Segmentation accuracy %

Scene category

Ours

Richtsfel

Mishr

Various single

objects (dataset

Lai,2011)

Single boxes

Cylindrical objects

Stacked boxes

Occluded objects

Mixed objects

Complex scenes

During quantitative analysis, we considered the

number of objects segmented correctly from the total

number of objects. We used sub-categories such as

single object instances (contains various type of

objects like apple, ball, cap, banana etc.), and also

different categories described in (Richtsfeld et al.,

2014) like single boxes, cylindrical objects, stacked

boxes, occluded objects, mixed objects and complex

scenes. Table-1 shows the object segmentation

accuracy of our method on different scene categories

compared with related works. Scene category of the

various single objects are from dataset (Lai et al.,

2011) and other categories are from dataset

(Richtsfeld et al.). As expected, our method gives

the best results in itself at the scenes where there is

only one object. The segmentation of simple objects

like boxes is more successful than objects with

complex surfaces. Multiple stacked objects or

occlusions do not have a negative affect on our

segmentation performance. Also, the number of

objects in a scene, would not affect performance, but

accuracy is reduced in more complex scenes

(Figure-4). When we examine fault cases of our

method, we can see that different surfaces of the

object are grouped incorrectly in cases where they

contain different color information. Richtsfeld’s

method seems better than our method for many

scene categories in Table-1, because they used

learning techniques for relations between object

surfaces, and that leads to better segmentation

results in complex scenes.

One of the most important factors affecting the

segmentation performance is the image resolution.

Images in the dataset (depth and RGB) have a

maximum resolution of 640x480 pixels. Due to the

small size of the objects in the scene, the properties

of the surface of the objects (like color, surface

normals) cannot be extracted precisely. As a result,

incorrect segmentation occurs.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

384

Stack of boxes

Various objects

Mixed & complex scene

RGB image

Depth data

Extracted

surfaces after

region growing

Segmented

Objects

Figure 4: Sample results from the datasets (First column and last column images (Richtsfeld et al., 2012a), middle image

(Aldoma and Richtsfeld)). Images in first two rows are from the datasets and last two rows are produced from our method.

When comparing the performance of our method

with the related works, we considered those who

experimented on the same dataset with us. In doing

so, we have included the best segmentation accuracy

of those studies that they mentioned. As it can be

seen in Table 2, our method is competing with or

even exceeding the relevant works in the field of

RGB-D object segmentation. Our comparison also

includes the results obtained by using only depth

information, where the colors on the surfaces are not

taken into consideration. Although we can achieve a

certain performance using only depth information,

we get the best results with color + depth data.

We compared the number of points difference

between ground truth data and our segmented object

while performing qualitative analysis. The ratio of

the number of points in the object segment to the

difference gives the incorrectly segmented points.

The average value is 1.23% through all scenes in all

of the datasets, that means segmented objects have

98.77% correct points compared to ground truth.

Table 2: Comparison of our method with the related works

(overall percentage of objects segmented correctly).

Dataset

Methods

Richtsfeld

Mishra

Ours

only depth

Ours

depth+color

Richtsfeld – Object

Segmentation

Dataset

(Richtsfeld et al.,

2012a)

Willow garage

dataset

(Aldoma and

Richtsfeld, 2012)

Lai – RGBD Scenes

Dataset

(Lai et al., 2012)

Automatic Object Segmentation on RGB-D Data using Surface Normals and Region Similarity

385

5 CONCLUSIONS

We proposed an automatic object segmentation

method using RGB-D data based on region growing

and region similarity. We apply point cloud filtering

for better abstraction of the depth data. After that

region growing segmentation is performed on point

cloud data based on surface normals. Since

segmentation with surface normal causes over-

segmentation, segmented surfaces of the objects

need to be grouped at the final step. Thus, we apply

a region merging method based on color

correlogram similarity and depth proximity

information of surface regions.

Experimental results show that our method has

comparable segmentation accuracy. The contribution

of our study is that primary features like surface

normals and color similarity of object surfaces can

be used for object segmentation using RGB-D data.

We explained a method that works unsupervised and

does not need to know how many objects exist in the

scene. Additionally, objects category is not an issue.

As a future work, machine learning methods can

be used to extract surface relations. Furthermore,

computation time is high and thus the proposed

method cannot be used in a real-time system for

now. If the algorithm is parallelized by taking

advantage of GPUs, the method can achieve real-

time video processing.

REFERENCES

Aldoma, A. & Richtsfeld, A. 2012. The willow garage

object recognition challenge [Online]. Available:

http://www.acin.tuwien.ac.at/forschung/v4r/software-

tools/object-instance-recognition-dataset/ [Accessed].

Hoppe, H., DeRose, T., Duchamp, T., McDonald, J. &

Stuetzle, W. 1992. Surface reconstruction from

unorganized points. SIGGRAPH Computer Graphics,

26, 71-78.

Huang, J., Kumar, S. R., Mitra, M., Zhu, W.-J. & Zabih,

R. 1997. Image indexing using color correlograms.

Computer Vision and Pattern Recognition, 1997.

Proceedings., 1997 IEEE Computer Society

Conference on, IEEE, 762-768.

Lai, K., Bo, L., Ren, X. & Fox, D. 2011. A large-scale

hierarchical multi-view rgb-d object dataset. Robotics

and Automation (ICRA), 2011 IEEE International

Conference on, IEEE, 1817-1824.

Lai, K., Bo, L., Ren, X. & Fox, D. 2012. Detection-based

object labeling in 3d scenes. Robotics and Automation

(ICRA), 2012 IEEE International Conference on,

IEEE, 1330-1337.

Liu, H., Philipose, M. & Sun, M.-T. 2013. Automatic

Objects Segmentation with RGB-D Cameras. Journal

of Visual Communication and Image Representation.

Martin, D. R., Fowlkes, C. C. & Malik, J. 2004. Learning

to detect natural image boundaries using local

brightness, color, and texture cues. Pattern Analysis

and Machine Intelligence, IEEE Transactions on, 26,

530-549.

Mishra, A., Aloimonos, Y. & Fah, C. L. 2009. Active

segmentation with fixation. Computer Vision, 2009

IEEE 12th International Conference on, IEEE, 468-

475.

Mishra, A. K. & Aloimonos, Y. 2012. Visual

segmentation of “Simple” objects for robots. Robotics:

Science and Systems VII, 217.

Mishra, A. K., Shrivastava, A. & Aloimonos, Y. 2012.

Segmenting “simple” objects using RGB-D. Robotics

and Automation (ICRA), 2012 IEEE International

Conference on, IEEE, 4406-4413.

Richtsfeld, A., Morwald, T., Prankl, J., Zillich, M. &

Vincze, M. 2012a. Segmentation of unknown objects

in indoor environments. Intelligent Robots and

Systems (IROS), 2012 IEEE/RSJ International

Conference on, IEEE, 4791-4796.

Richtsfeld, A., Mörwald, T., Prankl, J., Zillich, M. &

Vincze, M. 2014. Learning of perceptual grouping for

object segmentation on RGB-D data. Journal of visual

communication and image representation, 25, 64-73.

Richtsfeld, A., Zillich, M. & Vincze, M. 2012b.

Implementation of Gestalt principles for object

segmentation. Pattern Recognition (ICPR), 2012 21st

International Conference on, IEEE, 1330-1333.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

386