Comparison of Sparse Image Descriptors for Eyes Detection in Thermal

Images

Mateusz Knapik and Bogusław Cyganek

Department of Electronics, AGH University of Science and Technology, Kraków, Poland

Keywords:

Eye Detection, Sparse Descriptors, Surf Descriptor, Daisy Descriptor, Thermal Imaging, Operator Monitoring.

Abstract:

Eye detection and localization are basic steps in many computer systems aimed at human fatigue monitoring.

In this paper we evaluate performance of two sparse image descriptors for eye detection in the long-range IR

spectrum. In the training phase, sparse descriptors of the training images are computed and used to create fea-

tures vocabulary. Final detections are done using bag-of-words approach and additional heuristic for geometric

constraints. Several thermal video sequences were recorded to allow for quantitive analysis of this approach.

Experimental results show that our method achieves high accuracy in real conditions.

1 INTRODUCTION

Behavior and alertness level of an operator play es-

sential role in both people’s safety as well as in avoi-

ding economic losses. Real-time monitoring of indi-

viduals operating machinery and various vehicles is

a subject of constant innovation in recent years. Such

systems must cope with wide range of difﬁcult con-

ditions like noise, temperature, day and night illumi-

nation (Cyganek, 2007). They must be also able to

provide correct response in presence of variability in

appearance and behavior of human subjects.

For camera-based solutions, thermal long range

IR imaging is gaining attraction in recent years, spur-

ring both industrial solutions and research projects (Ji

and Yang, 2002; Saradadevi and Bajaj, 2008; Azim

et al., 2009; Fan et al., 2009; Cyganek and Gruszczy

ski, 2014; Shah et al., 2014; Wang et al., 2013; Str ˛a-

kowska and Str ˛akowski, 2015; Marzec et al., 2016;

Ribeiro et al., 2017). Unlike visible light sensors, long

range IR cameras can operate with no external lig-

hting even in total darkness. On the other hand, they

are insensitive to even extreme visible lighting condi-

tions frequently encountered in car conditions. Other

challenging factors for visible light cameras, like, for

example, skin color or camouﬂage makeup, are eli-

minated with thermal imaging. This properties makes

them a viable alternative for classic cameras. Also, a

supporting factor is growing availability of low-cost

long-infrared spectrum sensors.

However, proper image processing and analysis

methods are needed to create end-to-end solution. A

system for tiredness level monitoring, which is based

on facial features recognition, is composed of many

smaller subsystems such as data acquisition, image

processing and analysis for behavioral modeling. Al-

most always, one of the task involved in detection of

the facial landmarks is the eye detection step. This

paper evaluates performance of two sparse image des-

criptors when used for human eyes detection and loca-

lization in thermal images. The data processing path

is composed of computation of either the Speeded Up

Robust Features (SURF) (Bay et al., 2008) or DAISY

(Tola et al., 2010) descriptors on a dense grid. Then

clustering to achieve distinctive prototype patterns is

applied. Further on there are modules for matching

to vocabulary, similarity distance computation and ﬁ-

nally geometric constraints veriﬁcation. Contrary to

other techniques, visibility of both eyes in an image

is not required. The proposed method can be used in

many domains, ranging from surveillance, entertain-

ment or medical imaging.

Rest of this paper is organized as follows:

Section 2 describes the existing state-of-the-art works

that are related to the proposed system. Section 3 is

an overview of the system architecture and theory be-

hind methods used throughout this paper. Section 4

presents the results of conducted experiments to as-

certain scientiﬁc validity of presented system. Finally,

Section 5 presents our conclusions.

638

Knapik, M. and Cyganek, B.

Comparison of Sparse Image Descriptors for Eyes Detection in Thermal Images.

DOI: 10.5220/0007576506380644

In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 638-644

ISBN: 978-989-758-354-4

2 RELATED WORKS

In this section general overview of the state-of-the-art

works related to localization of facial areas in infrared

spectrum is presented and discussed.

For near range infrared (NIR) spectra, system for

eye detection was proposed by Cyganek and Grusz-

czy

nski (Cyganek and Gruszczy

nski, 2014). In order

to detect and localize eyes, NIR images are fed into

cascade of classiﬁers. Specialized iris-pupil model is

used to determine initial regions of interest. Tensor

classiﬁer is then used to select the most probable pair

of ROIs. Authors reported accuracy of over 96%.

For thermal imaging, currently existing works are

mostly suited for a constrained imaging conditions,

i.e. both eyes have to be present in the frame for algo-

rithm to work correctly.

Wang et al. (Wang et al., 2013) proposed sy-

stem for automatic eyes localization. Haar-like featu-

res boosted with Adaboost algorithm are used for de-

termining 15 subregions of face area, determining the

eyes position. Additionally, for situations when sub-

ject wears glasses, their system ﬁrst tries to localize

centers of each lens. However, face region is obtai-

ned by binarization and pixel summation, so presence

of any objects with temperature similar to the human

subject can affect the process and decrease its perfor-

mance.

Face tracking method in long-range IR based sy-

stems was proposed by Str ˛akowska et al. (Str ˛akow-

ska and Str ˛akowski, 2015). It is based on a principle,

that corners of the human eyes are the hottest regi-

ons of the face. In their approach, face area is lo-

calized using modiﬁed binarization algorithm. Then

H-maxima transformation is used to preprocess the

image and select regions of interests. Final result is

based on calculation of geometrical distances bet-

ween proposed regions.

In 2016, Marzec et al. in (Marzec et al., 2016)

presented new approach for fast eyes localization in

thermal images. Two-stage system consists of face

area detector and classiﬁer based on multilayer per-

ceptron neural network. Authors claims high accu-

racy of their method, however similar constraints as in

Wang’s methods are present, for example, both eyes

have to be visible to obtain regions of interest for furt-

her processing.

On the other hand, face detection in thermal ima-

ging using template matching was recently presented

by Ribeiro et al. (Ribeiro et al., 2017). Their propo-

sed method achieves good accuracy and low compu-

tational complexity. Authors also report that the met-

hod based on the Haar cascades, often used for visi-

ble light spectrum, can yield better results, but trai-

ning process of such classiﬁer requires large amount

of data and can be time consuming.

For features detection, Lowe presented algorithm

called Scale-invariant Feature Transform (SIFT)

(Lowe, 2004). It is one of the most well-known al-

gorithms for detection and descriptor generation of

local features in images. Due to it’s high invariance

to uniform scaling, orientation and illumination chan-

ges it is used in many applications including image-

stitching, video tracking and object detection, to name

a few.

Bay et al. proposed scale- and rotation-invariant

feature descriptor and complementary keypoint detec-

tor called Speeded Up Robust Features (SURF) (Bay

et al., 2008) as much faster alternative to SIFT. For de-

tection of points of interest SURF uses an approxima-

tion of the determinant of the Hessian blob detector.

Authors claims that SURF is more robust than SIFT

against different image transformations while being

several times faster to compute at the same time.

Tola et al. introduced feature descriptor called

DAISY (Tola et al., 2010), designed speciﬁcally to be

efﬁcient when computed densely, i.e. for 3D recon-

struction from the wide-baseline image pairs. It requi-

res less computational power than SIFT. In compari-

son to SURF, it is claimed not to introduce artifacts

that degrade matching performance when computed

on a dense grid.

Bag-of-words image categorization methods were

presented in a past by several researchers (Csurka

et al., 2004; Lazebnik et al., 2006; Fulkerson et al.,

2008) colorredboth for image classiﬁcation as well

as feature detection. Only recently BoW method was

also applied for feature detection in other spectra

(Malpani et al., 2016). In their work, Malpani et al.

presented system for human detection and localiza-

tion in thermal images.

To the best of our knowledge, the presented paper

is a ﬁrst comparison of SURF and DAISY descriptors

for eye detection in thermal images.

3 EYE LOCALIZATION IN

THERMAL IMAGES

An architecture of the eye detection module descri-

bed in this paper is presented in Fig. 1. Further secti-

ons present detailed description of each of the visible

modules. In our method, SURF and DAISY feature

descriptors are used for feature extraction. Either fea-

ture description technique was used separately on the

same set of images and the results have been com-

pared. Facial-features localization is later done using

sparse representation of a thermal image. Training

Comparison of Sparse Image Descriptors for Eyes Detection in Thermal Images

639

Figure 1: Eye detection system architecture.

images of eyes are used to create vocabulary of des-

criptors in a training phase. Later, descriptors extrac-

ted from new images are compared to the dictionary

to compute the similarity measure. Statistical model-

ling of a training data is used to remove features that

are dissimilar from the trained patterns. High con-

ﬁdence features are gathered into clusters and ﬁnal

eye candidates are chosen using series of geometrical

checks.

3.1 Contrast Enhancement

One of the main characteristics of thermal images is

their low contrast. Low variance in temperature bet-

ween facial landmarks can decrease the performance

of feature detection modules. In our method, stream

of the thermal images from the acquisition module are

preprocessed using multi-scale local contrast enhan-

cement method proposed by Cvetkovic et al. (Cvet-

kovic et al., 2009).

Every pixel of the input image is processed with

local-spatial ﬁlter computed as follows:

= i

+ (i

− m(i

)) (1)

where i

is the input pixel intensity and m is the mean

value around the pixel i

in a window of size 3 × 3

pixels calculated using Equation (2). Size of the win-

dow has been chosen empirically to suite the low re-

solution of the input images.

m(i

(x, y)) =

∑

s=−1

∑

p=−1

(x + s, y + p) (2)

Images before and after contrast enhancement can

be seen in Fig. 2.

(a)

(b)

Figure 2: Thermal image before (upper) and after contrast

enhancement (lower).

3.2 Vocabulary Extraction

Bag-of-words dictionary is created in the training

phase that starts with a computation of a dense grid

of feature descriptors on the training images of eyes

taken in long-range infrared spectrum. Dimensions

of the grid have been selected empirically and set to

desc

. Feature descriptors obtained in such way could

be used directly to form a feature vocabulary, howe-

ver to increase the generalization of our model feature

reduction technique called k-means clustering have

been applied. This iterative process groups together

similar descriptors to form k visual words from the

clusters centroids. Because the size of the dictionary

impacts the detection effectiveness, it was chosen em-

pirically, as will be shown later.

To measure the distance of a single feature from

the dictionary, an Euclidian metric is used. It is calcu-

lated as follows:

d( f , v) = min

( f

− x

)

+ ··· + ( f

n−1

− x

n−1

)

··· ,

( f

− x

k−1

)

+ ··· + ( f

n−1

− x

k−1

n−1

)

(3)

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

640

Table 1: Parameters used in the presented algorithms.

Parameter Value Used in Comment

desc

1 to 3 pixels Dense grid sampling

Distance in vertical and horizontal direction between

feature descriptors

, w

2 by 2 Distance comparison

How many neighboring features distances are used

for comparison

size

100 to 200 words Distance comparison Number of visual words in the dictionary

10 or 15 pixels Algorithm 1

Maximal distance in pixels between

neighboring points

ecc

0.85 Algorithm 2 Eccentricity limit for cluster

min

, g

max

2x2 pixels,

50x50 pixels

Algorithm 2 Cluster length and width constraints

min

, e

max

0.8,

Algorithm 2 Distance ratio between two eyes

max

60 deg Algorithm 2 Maximal angle between two clusters

where f is new feature, x

, x

, ··· , x

k−1

are the ele-

ments of ith visual word from the vocabulary, n deno-

tes the size of a single word and k denotes the size of

the whole vocabulary.

3.3 Distance Thresholding

Because human eyes will occupy relatively small area

of the whole image and the fact, that feature descrip-

tors are computed on a dense grid, our system em-

ploys features reduction method based on a statistical

modelling. To reduce number of potential regions of

interest, we compare the distance d( f , v) to a thres-

hold value obtained in a training phase. A mean plus

two standard deviations method described in (Miller,

1991) has been used in experiments presented in this

paper and was computed as follows

= D + 2 ·

∑

i=1



d( f

, v) − D



(4)

where N is the number of features in a training set,

d( f

, v) is the distance of an ith training feature (be-

fore clustering) from the vocabulary v and D is an

average distance of all training features from voca-

bulary v.

In detection phase, feature or group of features

will not be passed to further processing blocks if it’s

distance from vocabulary is bigger than the t

thres-

hold. To improve robustness, in our experiments we

used sum of distances of w

× w

features, both to

compute threshold as well as in detection phase.

Visualization of thresholding process can be seen

in Fig. 3(b).

3.4 Detection

Detection and localization tasks can be started after

the training phase is done. For every new image dense

grid of feature descriptors with the same grid dimen-

sions as in a training phase is computed and distance

3 is calculated for every feature. This creates a map of

distances between local descriptors and trained voca-

bulary. A sliding window of size w

× w

is then used

to compute average distance of neighboring features.

This is compared to the t

threshold and inliers are put

into a priority queue of length L

, which is ordered by

the distance value. The lower the value, the higher the

position in the queue.

To further decrease the amount of features, clus-

tering method called Salient Point Clustering (SPC)

is employed and every cluster is additionally approxi-

mated by an ellipse using central moments technique.

Both of these were presented and described in (Cyga-

nek, 2013).

Algorithm 1 shows the implementation details of

the SPC method using following deﬁnition of close

neighboring points: two points a and b are considered

as close neighbors if following inequality holds:



a, b



≤ d

(5)

where N denotes the Euclidean distance and d

stands

for an arbitrary chosen threshold.

A center point of a cluster c is computed using the

weighted average of every point, using it’s distance

value as a weight, using the following formulas.

max

= max

(6)

= w

max

− d

(7)

W =

∑

(8)

∑

· w

∑

· w

(9)

Comparison of Sparse Image Descriptors for Eyes Detection in Thermal Images

641

where D

denotes set of distances of features in a clus-

ter c, d

is a distance of ith feature in a cluster, w

a weight for ith feature, W is the sum of weights, N

is a number of features in a cluster c, x

and y

denote

the position of an ith feature of cluster c in a dense

features grid, and ﬁnally p

is the center point of a

cluster c.

Then the inertia tensor can be computed in terms

of the central moments, as follows.

∑

i=1

∑

j=1

− p

)

− p

)

(10)

T =

−c

(11)

To obtain length l

and width w

of the cluster c,

following formulas are computed:

, w

, λ

≥ λ

(12)

where λ

and λ

are eigenvalues of the inertia tensor

T .

Clusters are then inserted into another priority

queue of length L

that is ordered by ascending va-

lue of the median of cluster’s weights w

Algorithm 1: Salient Point Clustering.

Input: Set of points p

. Distance threshold d

Output: Clusters of points

1: while p

has elements do

2: pop last element from p

as s

3: for all clusters c do

4: c

is a ﬁrst point in cluster c

5: if inequality (5) holds for s and c

then

6: add s to cluster c

7: break

8: if s is still not clustered then

9: create new cluster, add s as a ﬁrst point

As a last step, simple geometric constraints checks

are made. Clusters must form ellipses of eccentricity

lower than g

ecc

and their length and width must be

within chosen limits. Finally, if more than two clusters

pass this criteria, their relative position, both the dis-

tance and the angle, is checked. These form the pro-

posed eye regions. Detailed description of this process

is shown in Algorithm 2.

4 EXPERIMENTAL RESULTS

The methods proposed in this article were implemen-

ted using Python language and well-known and publi-

cly available libraries like numpy, scipy, scikit-learn

and OpenCV. Experiments presented in this section

Algorithm 2: Geometric veriﬁcation.

Input: Clusters of points. Constants g

ecc

, g

min

max

, e

min

, e

max

, θ

max

Output: Proposed eye regions

1: for every cluster c do

2: if c eccentricity ≥ g

ecc

then

3: continue

4: if c width ≥ g

max

or length ≥ g

max

then

5: continue

6: if c width ≤ g

min

or length ≤ g

min

then

7: continue

8: add c to correct clusters list

9: if correct clusters length == 0 then

10: return empty list

11: else if correct clusters length == 1 then

12: return one proposed eye region

13: else

14: get ﬁrst two clusters from correct clusters list

as c

and c

15: set p

, p

to respective clusters centroids

16: if atan2(p

, p

) > θ

max

deg then

17: return empty list

18: D denotes the Euclidian distance between

points

19: if D(p

, p

) < e

min

or D(p

, p

) > e

max

then

20: return empty list

were run on a laptop computer equipped with 16GB

of RAM, and 4-core processor i7-6700HQ with the

2.6GHz clock. The 64-bit Windows



10 operating

system was used. Future implementations aimed at

speedup are also possible, e.g. by utilizing parallel

computing platforms, like CUDA

 or OpenCL



Quantitive evaluation was done using our own

thermal database that consist of over 3500 images

of human faces from three participants with manu-

ally labeled ground-truth frames. Database was cre-

ated using the FLIR



A35 camera that produces ima-

ges of resolution 320 × 256 pixels. For the purpose

of our comparison, number of experiments were con-

ducted. The following research questions were stated:

• What is the accuracy of the proposed method?

• Is there a measurable beneﬁt of using feature des-

criptor designed for dense sampling?

Average results of three repetitions of every experi-

ment are presented. Before every test round dataset

was randomly split in half to create training and vali-

dation subsets. Result from every frame was categori-

zed in terms of true-positive (TP), true-negative (TN),

false-positive (FP), and false-negative (FN) parame-

ters and then F value was computed. Each correctly

recognized eye position, that is, the difference bet-

ween region found by an algorithm and the ground

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

642

Table 2: Results of detection accuracy experiments.

F-value result

DAISY SURF

Vocabulary 100 150 200 100 150 200

Cluster size 10 15 10 15 10 15 10 15 10 15 10 15

Step size 1 0.808 0.817 0.802 0.795 0.785 0.796 0.784 0.801 0.790 0.791 0.776 0.788

2 0.755 0.931 0.747 0.942 0.747 0.943 0.688 0.846 0.671 0.827 0.677 0.861

3 0.789 0.912 0.786 0.908 0.772 0.908 0.727 0.844 0.735 0.857 0.733 0.858

truth is less than 10 pixels (around 3% given the

size of an input image), is reported as a true-positive.

Every incorrectly detected eye is counted as false-

positive. False-negative is increased when algorithm

fails to detect an eye. Eyes not present in the picture

and not reported by algorithm are counted as true-

negative. Summary of the parameters used in the con-

ducted experiments is presented in Table 1. Full re-

sults of the experiments are presented in Table 2.

To ﬁnd optimal parameters for each type of feature

descriptor many combinations of parameters were

tested. Three different grid spacing were considered,

from 1 to 3 pixels, alongside three different dictionary

sizes, from 100 to 200 visual words and two sizes of

clustering thresholds, 10 and 15 pixels. Bigger clus-

ters in case of both descriptors increased the perfor-

mance signiﬁcantly. More features allowed to reﬁne

clusters size and ﬁnd center point more accurately.

However, it should be noted that clustering threshold

is correlated with grid spacing. Very dense grid results

in small intra-class difference between words. Incre-

ase in spatial distance between features allows for bet-

ter utilization of vocabulary, therefore bigger clusters

can be localized with better precision. This results in

better detector performance. On the other hand, in-

creasing vocabulary size had little effect on the efﬁ-

cacy of detection, even decreasing F-value in some

cases. Low resolution and noise in the input images

decrease the intra-class difference between computed

descriptors, therefore expanding dictionary size does

not bring radical performance increase while being

computationally expensive. We checked the lowest

and highest F-value for both descriptor types. Even

in worst case scenario, the F-value is higher than 0.67

for SURF descriptor and 0.74 for DAISY descriptor

which is quite good considering the vagueness of eye

contours in thermal images. DAISY descriptor achie-

ves around 6% better results on average, reaching

over 0.94 F-value with ﬁne-tuned grid and vocabulary

size (0.86 for SURF). This results answer our second

research question. In our experimental setup, a dicti-

onary with 200 words, combined with cluster size of

15 pixels and grid step of 2 pixels, gave the overall

best results for both types of descriptors.

(a)

(b)

(c)

Figure 3: Distance map (top), corresponding map after

thresholding and clustering (middle) and ﬁnal detection re-

sult (bottom).

5 CONCLUSION

In this paper two types of local features descriptors

are compared for usage in eye detection in thermal

images system. Dense grid of feature descriptors is

used to create a dictionary of visual words, that are

Comparison of Sparse Image Descriptors for Eyes Detection in Thermal Images

643

then used for eye regions detection and localization.

As shown, the proposed system achieves high accu-

racy as well as proves the beneﬁts of using feature

descriptor designed with dense sampling in mind.

There is also room for improvements in implemen-

tation after which it can operate in real-time. The pro-

posed system can be also used in other domains, such

as medicine, surveillance or operators fatigue monito-

ring.

ACKNOWLEDGEMENTS

This work was supported by the AGH University

of Science and Technology under the grant no.

15/11/421.

REFERENCES

Azim, T., Jaffar, M., Ramzan, M., and Anwar, M. (2009).

Automatic fatigue detection of drivers through yaw-

ning analysis. In Signal Processing, Image Proces-

sing and Pattern Recognition, pages 125–132, Berlin,

Heidelberg. Springer Berlin Heidelberg.

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

Speeded-up robust features (surf). Comput. Vis. Image

Underst., 110(3):346–359.

Csurka, G., C., R., D., L., F., J., W., and C., B. (2004). Visual

categorization with bags of keypoints. workshop on

statistical learning in computer vision. International

Conference on Computer Vision (ECCV 2004), page

1–22.

Cvetkovic, S., Schirris, J., and With, de, P. (2009). Locally-

adaptive image contrast enhancement without noise

and ringing artifacts. In Proceedings of the IEEE

International Conference on Image Processing (ICIP

2007), October 16-19, 2007, San Antonio, Texas, pa-

ges 557–560, United States. Institute of Electrical and

Electronics Engineers (IEEE).

Cyganek, B. (2007). Soft System for Road Sign Detection,

pages 316–326. Springer Berlin Heidelberg, Berlin,

Heidelberg.

Cyganek, B. (2013). Object Detection and Recognition in

Digital Images: Theory and Practice. Wiley.

Cyganek, B. and Gruszczy

nski, S. (2014). Hybrid computer

vision system for drivers’ eye recognition and fatigue

monitoring. Neurocomputing, 126:78–94.

Fan, X., Yin, B.-C., and Sun, Y.-F. (2009). Yawning de-

tection based on gabor wavelets and lda. 35:409–

413+432.

Fulkerson, B., Vedaldi, A., and Soatto, S. (2008). Locali-

zing objects with smart dictionaries. In Forsyth, D.,

Torr, P., and Zisserman, A., editors, Computer Vision

– ECCV 2008, pages 179–192, Berlin, Heidelberg.

Springer Berlin Heidelberg.

Ji, Q. and Yang, X. (2002). Real-time eye, gaze, and face

pose tracking for monitoring driver vigilance. 8:357–

377.

Lazebnik, S., Schmid, C., and Ponce, J. (2006). Beyond

bags of features: Spatial pyramid matching for recog-

nizing natural scene categories. 2006 IEEE Computer

Society Conference on Computer Vision and Pattern

Recognition - Volume 2 (CVPR’06), page 1–22.

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. Int. J. Comput. Vision, 60(2):91–

110.

Malpani, S., S, A. C., and Narasimhadhan, A. V. (2016).

Thermal vision human classiﬁcation and localization

using bag of visual word. In 2016 IEEE Region 10

Conference (TENCON), pages 3135–3139.

Marzec, M., Lam

za, A., Wróbel, Z., and Dziech, A. (2016).

Fast eye localization from thermal images using neu-

ral networks. Multimedia Tools and Applications.

Miller, J. (1991). Short report: Reaction time analysis

with outlier exclusion: Bias varies with sample size.

The Quarterly Journal of Experimental Psychology

Section A, 43(4):907–912.

Ribeiro, R. F., Fernandes, J. M., and Neves, A. J. R. (2017).

Face detection on infrared thermal image. In SIGNAL

2017 : The Second International Conference on Ad-

vances in Signal, Image and Video Processing, pages

38–42. IARIA.

Saradadevi, M. and Bajaj, P. (2008). Driver fatigue de-

tection using mouth and yawning analysis. 8.

Shah, A., Kukreja, S., Shinde, P., and Kumari, A. (2014).

Yawning detection of driver drowsiness. 2.

Str ˛akowska, M. and Str ˛akowski, R. (2015). Automatic eye

corners detection and tracking algorithm in sequence

of thermal medical images. Measurement Automation

Monitoring, 61(6):199–202.

Tola, E., Lepetit, V., and Fua, P. (2010). Daisy: An efﬁ-

cient dense descriptor applied to wide baseline stereo.

IEEE TRANS. PATTERN ANALYSIS AND MACHINE

INTELLIGENCE, 32(5).

Wang, S., Liu, Z., Shen, P., and Ji, Q. (2013). Eye localiza-

tion from thermal infrared images. Pattern Recogni-

tion, 46(10):2613 – 2621.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

644