Finding Similar non-Collapsed Faces to Collapsed Faces Using Deep

Learning Face Recognition

Ashwinee Mehta

1 a

, Maged Abdelaal

2 b

, Moamen Sheba

2 c

and Nic Herndon

1 d

Department of Computer Science, East Carolina University, Greenville, U.S.A.

School of Dental Medicine, East Carolina University, Greenville, U.S.A.

Keywords:

Similar, non-Collapsed Face, Face Recognition, Classiﬁcation, Collapsed Face, Reconstruction.

Abstract:

Face recognition is the ability to recognize a person’s face in a digital image. Common uses of face recognition

include identity veriﬁcation, automatically organizing raw photo libraries by person, tracking a speciﬁc person,

counting unique people and ﬁnding people with similar appearances. However, there is no systematic and

accurate study for ﬁnding a similar non-collapsed face to a given collapsed face. In this paper we focus on the

use case of ﬁnding people with similar appearances that will help us to ﬁnd a similar face without a collapse

to a collapsed face for dental reconstruction. We used Python’s Open-CV for age and gender classiﬁcation

and face recognition for ﬁnding similar faces. Our results provide a set of similar images that can be used for

reconstructing the collapsed faces for creating dentures. Thus with the help of a similar non-collapsed face,

we can reconstruct a collapsed face for designing effective dentures.

1 INTRODUCTION

Images contain a lot of crucial information that can be

used for a variety of applications. The importance of

search applications that closely match facial features

in image-based searches is increasing day by day. We

all have wondered if there was someone out there who

looked just like us. We are all aware of the urban

legend that there are six people out there who look

just like us. The problem lies in ﬁnding those similar

people. Face matching and retrieval can be used in

forensics applications in matching forensic sketches

to face photograph databases. Face matching also has

its applications in recommender systems for glasses,

hairstyle, jewellery, etc., and in dentistry for ﬁnding a

similar face for facial reconstruction and denture de-

signing.

A lot of work has been performed in the ﬁeld of

facial similarity where different methods and met-

rics have been proposed for ﬁnding similar faces to

a given input face like Scale-Invariant Feature Trans-

form (SIFT) descriptors, Lookalike networks, Local

Binary Pattern (LBP), Doppelganger lists, etc. There

https://orcid.org/0000-0002-7167-2563

https://orcid.org/0000-0002-7414-423X

https://orcid.org/0000-0003-1188-2080

https://orcid.org/0000-0001-9712-148X

are situations in real life where it is required to ﬁnd

a similar face to a face which is collapsed or has

structural deformities. However, none of the exist-

ing methods have focused on faces with deformities

or a collapse. All the existing methods have the in-

put and output similar images of normal faces with-

out collapse or deformities where the facial similarity

is calculated and a similar face is returned from a tar-

get dataset.

Missing human teeth cause the body to reabsorb

the bone that supported the teeth. Over the course of

about 10-20 years the jawbone shrinks signiﬁcantly

which results in a condition known as facial collapse.

Humans with facial collapse appear much older than

they are. The facial collapse not only alters the hu-

man’s facial appearance, but also affects their dental

health. This facial collapse can be prevented with the

placement of dental implants. The implant sends a

piezoelectric signal to the bone which prevents the

bone from reabsorption. Dentures are designed by us-

ing the denture impressions of the jaw and mouth af-

ter which the dentist will create models usually from

wax or plastic, based off the impression. The pa-

tient then tries the model several times to check for

ﬁt, shape, and even color before the denture is made.

The current denture design workﬂow does not have a

systematic approach to include the aesthetic factors,

patient’s pre-treatment facial shape and in-progress

Mehta, A., Abdelaal, M., Sheba, M. and Herndon, N.

Finding Similar non-Collapsed Faces to Collapsed Faces Using Deep Learning Face Recognition.

DOI: 10.5220/0011696900003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 4: VISAPP, pages

897-904

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

897

denture design visualizations, instead relying on dis-

cussing mockups with the patients during appoint-

ments. This results into waiting for the ﬁnal denture

ﬁtting on the patient to evaluate the ﬁnal denture aes-

thetics.

With a collapsed face, only the bottom one third

of the face is affected and needs to be restored be-

fore denture designing. This facial restoration needs a

similar face which can be used as a reference for com-

paring the facial shape proportions to convert the col-

lapsed face into non-collapsed facial shape. Inclusion

and evaluation of facial aesthetics is important while

planning for facial or dental reconstruction treatment.

Our proposed method focuses towards the goal

of ﬁnding a similar non-collapsed face to a given

collapsed face for reconstructing the lower third of

the collapsed face before designing the denture. We

have used Python’s Deep Learning Face Recognition

(Geitgey, 2018) which is built using Dlib (King, 2009)

for ﬁnding similar non-collapsed faces to a given col-

lapsed face from a diverse target dataset of human

population. By using the proposed method, one can

ﬁnd multiple similar images and use them for facial

reconstruction for automatic denture designing.

2 MATERIALS AND METHODS

We have worked on ﬁnding similar non-collapsed

faces for a given collapsed face by using the Hu-

man Faces dataset (kaggle, 2020). The Human Faces

Dataset contains 224,500 images of human faces. The

image data in the dataset has been generated using

StyleGAN2 (Karras et al., 2020) which focuses on

improving the resolution and quality of images. The

StyleGAN2 was trained by using the Flickr Faces HQ

(FFHQ) and Large-scale Scene UNderstanding Chal-

lenge (LSUN) datasets. The Human Faces Dataset

contains nine directories and each such directory con-

tains multiple folders that contain the image ﬁles.

Each image ﬁle contains an image of a single human

face. The dataset is diverse in a way that it contains

images of human faces of different age groups begin-

ning from toddlers to old humans and different gen-

ders, i.e., male and female, yet there’s no metadata

(e.g., age, race, etc.). All the image ﬁles in the dataset

have a resolution of 1024 × 1024.

OpenCV’s gender and age classiﬁcation is based

on a convolutional neural network architecture with

a total of 3 convolutional layers, 2 fully connected

layers and a ﬁnal output layer. “Conv1” is the ﬁrst

convolutional layer that has 96 nodes of kernel size

7. “Conv2” is the second convolutional layer that has

256 nodes of kernel size 5. “Conv3” is the third con-

volutional layer that has 384 nodes of kernel size 3.

The two fully connected layers have 512 nodes each.

Gender prediction is framed as a classiﬁcation prob-

lem. The output layer in the gender prediction net-

work is of type softmax with 2 nodes indicating the

two classes “Male” and “Female”. Age Prediction

should be approached as a regression problem since

we are expecting a real number as the output. How-

ever, it is difﬁcult to accurately estimate the age of a

person and even humans ﬁnd it challenging. Hence,

age prediction was framed as a classiﬁcation problem

where we try to estimate the age group the person is

in. The age prediction network has 8 nodes in the ﬁ-

nal softmax layer indicating the age ranges 0 to 2, 4 to

6, 8 to 12, 15 to 20, 25 to 32, 38 to 43, 48 to 53 and 60

to 100 years old. The different model ﬁles used and

loaded for performing the age and gender detection

task are listed below:

• gender net.caffemodel: It is the pre-

trained model weights for gender detection.

• deploy gender.prototxt: It is the model

architecture for the gender detection model.

• age net.caffemodel: It is the pre-trained

model weights for age detection.

• deploy age.prototxt: It is the model archi-

tecture for the age detection model.

• res10 300x300 ssd iter 140000 fp16.

caffemodel: It is the pre-trained model

weights for face detection.

• deploy.prototxt.txt: It is the model ar-

chitecture for the face detection model.

Our proposed workﬂow as shown in Figure 2 had

the following steps:

1. Image Classiﬁcation: In order to ﬁnd a sim-

ilar non-collapsed face, it was important to ﬁnd the

similar face around the same age-group and gender.

The ﬁrst step was to classify the image according to

gender into male and female categories. We used

Python’s OpenCV for age and gender detection of hu-

man faces in the image ﬁles. The images were classi-

ﬁed into two genders, i.e., male and female, and fur-

ther classiﬁed into different age groups under each

gender. The different age classes included 0 to 2, 4

to 6, 8 to 12, 15 to 20, 25 to 32, 38 to 43, 48 to 53 and

60 to 100 years old.

The different steps involved in classiﬁcation of

gender and age include:

1. Read the image using the cv2.imread() method.

2. After the image is resized to the appropriate size,

use the get faces() function to get all the de-

tected faces from the image.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

898

3. Iterate on each detected face image and

call our get age predictions() and

get gender predictions() to get the pre-

dictions.

4. Print the age and gender along with the conﬁdence

levels.

2. Generating Image File Metadata: After de-

tection of gender and age, the next step was to store

the the output of classiﬁcation as metadata for every

image ﬁle. The ﬁle name of every image ﬁle was re-

named to include the metadata of age and gender and

their respective conﬁdence levels.

3. Sorting Image Files Into Folders: After gen-

erating the image ﬁle metadata, the next step was

to sort the classiﬁed data into different folders based

on the gender and age for searching the similar non-

collapsed face to a collapsed face in the folder that

corresponds to the collapsed face’s detected gender

and age (e.g., male 0to2, male 4to6, etc.).

4. Generate Similar Images: After having all the

images sorted in different folders as per their classi-

ﬁcation result, the next step was to look for a similar

non-collapsed face to the given input collapsed face

based on the gender and age of the input collapsed

face by using Python’s Deep Learning Face Recogni-

tion as shown in Figure 1. We found out that the pro-

cess of generating similar images is slow. To speed up

the process we created small samples of each folder

mentioned in Step 3 that contained 500 images and

generated similar images by searching only the sam-

ple folder.

3 RESULTS AND DISCUSSION

Figure 3 shows the query image that contains a col-

lapsed face of a male in the age group 60 to 100 and

the top ten similar images to the query image ranked

by their similarity scores. Figure 6 shows the query

image of a collapsed face of a female in the age group

48 to 53 and the top ten similar images to the query

image ranked by their similarity scores. In Figure 3,

the most similar face has a similarity score of 0.6050

followed by the other less similar faces with an in-

creasing similarity scores. In Figure 6, the most sim-

ilar face has a similarity score of 0.6210 followed by

other less similar faces with an increasing similarity

scores. Out of all the images involved in the experi-

ment, we can say that, using the proposed method, the

image with the lowest score is the most similar image

to the given query image. When we visually evalu-

ate the results, we can conﬁrm that this is the case for

most of the query images.

In Figure 3, Figure 5 and Figure 6, the gender of

all the similar images in the output matches the gen-

der of the query image that contains a collapsed face.

In Figure 3 the query image is of a male and all the

similar images are male and in Figures 5 and 6, the

query images are of a female and all the similar im-

ages are female. Also, the age of most of the simi-

lar images matches with the age of the query image

in Figure 3 and Figure 5. Having ten similar images

gives the ﬂexibility to choose from the set that might

have some images that do not satisfy some criteria.

The proposed method also works for faces in all age

groups that do not have a collapsed face.

The dataset had some duplicate images which

were returned in the similarity results. In this situa-

tion, only one such image was selected and the next

similar image was considered. Some images in the

dataset did not have clear face visibility due to which

there were errors. After excluding such images, we

were able to ﬁnd similar non-collapsed images to a

given collapsed face. Since the dataset used in this

paper did not have labelled data, we used machine

learning to classify the images into different age and

gender categories. During this process of classiﬁca-

tion, there were images that were not classiﬁed into

correct folders. Thus some folders contained images

that belong to other age and gender category. This re-

sulted into missing out on some images that were po-

tential candidates for similar images. During the step

of generating image ﬁle metadata i.e., gender and age

labels, Python’s OpenCV based age and gender de-

tection also displays a conﬁdence score for each la-

bel. This conﬁdence score can be used for ﬁltering

out incorrectly labelled images from the age and gen-

der category folders so that each folder contains only

correctly labelled images.

4 RELATED WORK

In (Torun et al., 2009), a method is developed to

match similar faces in scattered datasets and to rec-

ognize images of a person in different states. A solu-

tion that works more effectively than traditional face

recognition methods, which works with a high error

rate according to different exposure values, is pre-

sented. By using SIFT descriptors that are used in ob-

ject recognition problems, ﬁrst, the face’s points of in-

terest are identiﬁed followed by matching points of in-

terests between two faces. Then, the similarity ratios

of two faces to each other are calculated by looking at

the distances. The average distances of the points of

interest assumed to be correctly mapped determined

this ratio.

Finding Similar non-Collapsed Faces to Collapsed Faces Using Deep Learning Face Recognition

899

Figure 1: Face Recognition Pipeline.

In (Schroff et al., 2011) a new method is intro-

duced for comparing pairs of face images that allows

recognition or veriﬁcation between images where the

mutually visible portion of face is small. They have

explored the idea of using a sorted Doppelganger

list as a signature and evaluate distance functions for

comparing the signatures. Each probe in a query pair

is compared to all members of the Library. The com-

parison results in a ranked list of look-alikes, the ﬁrst

being the most similar to the query. Then, a similar-

ity measure between the two probes is computed by

comparing the ranked lists.

In (Sadovnik et al., 2018), the emphasis is on pre-

senting evidence that ﬁnding facial look-alikes and

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

900

Figure 2: Proposed Workﬂow.

recognizing faces are two distinct tasks. They ex-

pected that many features which are useful for face

recognition will also be useful for face similarity.

They have used the pre-trained VGG face CNN de-

scriptor network for face-recognition and then ﬁne-

tuned the weights on a new dataset that is targeted at

capturing preceived facial similarity to perform well

at the face similarity task by proposing the lookalike

network. The results show that the proposed method

outperforms the face recognition baseline at the task

of predicting which faces will appear more similar to

human.

In (Ramanathan et al., 2004), the focus is on deriv-

ing a measure of similarity between faces. They men-

tion that illumination, pose variations, disguises, ag-

ing effects and expression variations are some of the

key factors that affect the performance of face recog-

nition systems. They have suggested a framework to

compensate for pose variations and introduce the no-

tion of ’Half-faces’ to circumvent the problem of non-

uniform illumination and used the similarity measure

to retrieve similar faces from a database containing

multiple images of individuals. They concluded that

the similarity measure helps in studying the signiﬁ-

cance facial features play in affecting the performance

of face recognition systems.

In (McCauley et al., 2021a), they have used the

twin database to calculate a baseline measure of

the worst-case scenario of facial similarity in Face

Recognition using a deep CNN. They have carried out

a performance analysis of two Face Recognition tools

presented with highly similar faces and using an ex-

perimental twin threshold, potential look-alikes were

extracted from the datasets for further analysis. The

similarity measure presented in this paper can be used

to compare facial similarity to a comparison score

from a Face Recognition system in order to better un-

derstand the impact that facial similarity has on Face

Recognition and to identify potential look-alikes from

large datasets.

In (Alasadi et al., 2019), they have showed that

adopting an adversarial deep learning-based approach

allows for the model to maintain the accuracy at face

matching while also reducing demographic dispar-

ities compared to a baseline (non-adversarial deep

learning) approach at face matching and paves way

for more accurate and fair face matching algorithms.

They have proposed a deep-learning adversarial ap-

proach for reducing bias in face-matching algorithms.

The proposed GAN-based framework consisted of

two parts- one which tries to maximize the face

matching quality and the other which tries to min-

imize the ability of the network to infer the demo-

graphic properties of the individual whose facial im-

age is under consideration.

In (Bicego et al., 2005), they have presented a

novel approach for extracting characteristic parts of

a face. Instead of ﬁnding a priori speciﬁed features

such as nose, eyes, mouth or others, the proposed ap-

proach is aimed at extracting from a face the most

distinguishing or dissimilar parts with respect to an-

other given face, i.e. at “ﬁnding differences” between

faces by feeding a binary classiﬁer by a set of image

patches, randomly sampled from the two face images,

and scoring the features by their mutual distances.

In (Ruys et al., 2006), they have investigated the

role of differences between people in the comparison

process. They have proposed a dual process model in

which dissimilarity can function in two different ways

with opposite effects on social judgement. They have

discussed two different arguments- 1. Dissimilarity

between people may decrease the likelihood of plac-

ing them in the same category. 2. Activated dissimi-

larity may determine the detection of feature overlap

between people during comparison and therefore in-

ﬂuence the holistic similarity assessment.

In (O’Toole et al., 2007), they have compared

seven state-of-the-art face recognition algorithms

with humans on a facematching task where humans

and algorithms determined whether pairs of face im-

ages, taken under different illumination conditions,

were pictures of the same person or of different peo-

ple. It was found out that three algorithms surpassed

human performance matching ”difﬁcult” face pairs

and six algorithms surpassed humans on ”easy” face

pairs. They mention that although illumination vari-

Finding Similar non-Collapsed Faces to Collapsed Faces Using Deep Learning Face Recognition

901

Figure 3: The query image is a collapsed face of a male in the age group 60 to 100. Top ten non-collapsed faces ranked

by their similarity are found that can be used for reconstructing the bottom third of the collapsed face. The next two ranked

images has the similarity of 0.7569 and 0.7574.

Figure 4: The query image is a collapsed face of a female in the age group 60 to 100. Top ten non-collapsed faces ranked

by their similarity are found that can be used for reconstructing the bottom third of the collapsed face. The next two ranked

images has the similarity of 0.6223 and 0.6243.

Figure 5: The query image is a collapsed face of a male in the age group 48 to 53. Top ten non-collapsed faces ranked by their

similarity are found that can be used for reconstructing the bottom third of the collapsed face. The next two ranked images

has the similarity of 0.7033 and 0.7079.

ation continues to challenge face recognition algo-

rithms, current algorithms compete favorably with

humans.

(Kramer and Reynolds, 2018) have focused on

face matching using proﬁle images. They have com-

pared face matching accuracy when both frontal and

proﬁle image of each face were presented, with accu-

racy using each view alone. Surprisingly, they found

no beneﬁt when both views were presented together.

Also, they found out that there was no difference in

performance when front and proﬁle views were used

suggesting that both views were similarly useful for

face matching. Overall, these results suggest that

either frontal or proﬁle views provide substantially

overlapping information regarding identity or partici-

pants are unable to utilize both sources of information

when making decisions.

In (Oron et al., 2018), they have proposed a new

method called Best-Buddies Similarity for template

matching using mutual nearest neighbors. The pro-

posed method follows the traditional sliding window

approach and by computing the Best-Buddies Simi-

larity between the template and every window of the

size of the template in the image. Best-Buddies Simi-

larity is calculated by leveraging statistical properties

of mutual nearest neighbors and was shown to be use-

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

902

Figure 6: The query image is a collapsed face of a female in the age group 48 to 53. Top ten non-collapsed faces ranked

by their similarity are found that can be used for reconstructing the bottom third of the collapsed face. The next two ranked

images has the similarity of 0.65763 and 0.65766.

ful for template matching in the wild. Key features of

BBS were identiﬁed and analyzed demonstrating its

ability to overcome several challenges that are com-

mon in real life template matching scenarios. They

also mention that the proposed method might have

additional applications in computer-vision or other

ﬁelds that could beneﬁt from its properties.

In (Santini and Jain, 1996), focus is on the prob-

lem of similarity matching. They state that similar-

ity matching is the single most important operation in

image databases. They have discussed some of the

models of human similarity like Euclidean Distance,

City-Block Distance, Thurstone-Shepard Models, the

Feature Contrast Model and they Fuzzy Feature Con-

trast Model. They have used images from the MIT

face images dataset for measuring similarity based on

geometric measures. They have modiﬁed the Fea-

ture Contrast Model into the Fuzzy Feature Contrast

Model. From the experimental results, they prove that

the Fuzzy Feature Contrast Model is almost symmet-

ric in case of human faces.

In (McCauley et al., 2021b), they have discussed

the problem of distinguishing identical twins and non-

twin look-alikes. They address two challenges- de-

termining a baseline measure of facial similarity be-

tween identical twins and applying this similarity

measure to determine the impact of doppelgangers,

or look-alikes, on FR performance for large face

datasets. The facial similarity measure is determined

via a deep convolutional neural network. The pro-

posed network provides a quantitative similarity score

for any two given faces and has been applied to large-

scale face datasets to identify similar face pairs.

In (R

ottcher et al., 2020), they have proposed a

new method for efﬁciently select similar pairs. They

have compared the proposed method with the adapted

version of a random selection process which is of-

ten found in state-of-the-art morphing attack research.

The conducted experiment proves that appropriate

pair selection not only increases the morph quality but

also substantially decreases the standard deviation be-

tween many morphing techniques. An effective pre-

selection reduces the need for a perfect, low-artefact

producing, morphing algorithm. This is important as

automated morphing is still error-prone with the dif-

ﬁculty to fully remove all the artefacts. Automated

face recognition systems that operate on the purpose

of determining the similarity between two facial im-

ages are not only vulnerable to morphed faces but they

can also contribute to a morphing attack by ﬁnding

optimal pairs of data subjects in a sufﬁcient manner.

In (Lamba et al., 2011), they have prepared a look-

alike database and analyzed human performance with

the help of 50 volunteers. They have used both sub-

space and texture descriptor based algorithms for au-

tomatic algorithms. Their experimental results sug-

gest that, for look-alikes, humans and automatic al-

gorithms do not perform better than random guess.

They have also also proposed an algorithm that sig-

niﬁcantly improves the performance compared to ex-

isting algorithms. They state that it is important to

start considering complex covariates including look-

alikes and develop advanced algorithms.

5 CONCLUSIONS AND FUTURE

WORK

The bottom third of the face is adversely affected

when a facial collapse due to missing teeth and jaw

bone loss occurs. It is important to properly re-

store this third of the face before designing dentures.

The lower third of a collapsed face can be restored

by using a similar non-collapsed face as a reference

face. We have proposed a method to ﬁnd a similar

non-collapsed face to a given collapsed face using

Python’s Deep Learning Face Recognition. Our re-

sults provide the similar non-collapsed images to the

given collapsed image which can be used for recon-

Finding Similar non-Collapsed Faces to Collapsed Faces Using Deep Learning Face Recognition

903

structing the bottom third of a collapsed face before

designing dentures.

When looking for a similar non-collapsed face

which can be used as a reference to the query im-

age that contains a collapsed face, it would be ideal

to have a reference face that is a closest match to the

query image as this would help in more accurate re-

construction procedure of the query image that con-

tains a collapsed face. Since this proposed method is

about ﬁnding similar images from the target dataset,

the bigger and diverse the dataset is, the more is the

probability of ﬁnding a closer and a similar match to

the query image. However, it should be made sure

that the target dataset does not contain images of peo-

ple with a collapsed face. In addition to age and gen-

der, there can be other factors like race, ethnicity, skin

color, height, weight, etc. that can affect the similar-

ity score and help to ﬁnd the closest match. It is worth

exploring the similar images in a more wide and di-

verse dataset to check for the similarity scores and

similar images. Also, we would like to collect images

of patients that have collapsed faces to evaluate the

proposed method.

REFERENCES

Alasadi, J., Al Hilli, A., and Singh, V. K. (2019). To-

ward fairness in face matching algorithms. In Pro-

ceedings of the 1st International Workshop on Fair-

ness, Accountability, and Transparency in MultiMe-

dia, FAT/MM ’19, page 19–25, New York, NY, USA.

Association for Computing Machinery.

Bicego, M., Grosso, E., and Tistarelli, M. (2005). On ﬁnd-

ing differences between faces. In International Con-

ference on Audio-and Video-Based Biometric Person

Authentication, pages 329–338. Springer.

Geitgey, A. (2018). Face recognition library.

kaggle (2020). Human faces dataset.

Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen,

J., and Aila, T. (2020). Analyzing and improving

the image quality of stylegan. In Proceedings of the

IEEE/CVF conference on computer vision and pattern

recognition, pages 8110–8119.

King, D. E. (2009). Dlib-ml: A machine learning toolkit.

Journal of Machine Learning Research, 10:1755–

1758.

Kramer, R. S. and Reynolds, M. G. (2018). Unfamiliar face

matching with frontal and proﬁle views. Perception,

47(4):414–431.

Lamba, H., Sarkar, A., Vatsa, M., Singh, R., and Noore, A.

(2011). Face recognition for look-alikes: A prelimi-

nary study. In 2011 International Joint Conference on

Biometrics (IJCB), pages 1–6.

McCauley, J., Soleymani, S., Williams, B., Dando, J.,

Nasrabadi, N., and Dawson, J. (2021a). Identical

twins as a facial similarity benchmark for human fa-

cial recognition. In 2021 International Conference

of the Biometrics Special Interest Group (BIOSIG),

pages 1–5.

McCauley, J., Soleymani, S., Williams, B., Dando, J.,

Nasrabadi, N., and Dawson, J. (2021b). Identical

twins as a facial similarity benchmark for human fa-

cial recognition. In 2021 International Conference

of the Biometrics Special Interest Group (BIOSIG),

pages 1–5.

Oron, S., Dekel, T., Xue, T., Freeman, W. T., and Avi-

dan, S. (2018). Best-buddies similarity—robust tem-

plate matching using mutual nearest neighbors. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 40(8):1799–1813.

O’Toole, A. J., Jonathon Phillips, P., Jiang, F., Ayyad, J.,

Penard, N., and Abdi, H. (2007). Face recognition al-

gorithms surpass humans matching faces over changes

in illumination. IEEE Transactions on Pattern Analy-

sis and Machine Intelligence, 29(9):1642–1646.

Ramanathan, N., Chellappa, R., and Roy Chowdhury, A.

(2004). Facial similarity across age, disguise, illumi-

nation and pose. In 2004 International Conference on

Image Processing, 2004. ICIP ’04., volume 3, pages

1999–2002 Vol. 3.

ottcher, A., Scherhag, U., and Busch, C. (2020). Finding

the suitable doppelg

anger for a face morphing attack.

In 2020 IEEE international joint conference on bio-

metrics (IJCB), pages 1–7. IEEE.

Ruys, K. I., Spears, R., Gordijn, E. H., and de Vries, N. K.

(2006). Two faces of (dis) similarity in affective judg-

ments of persons: Contrast or assimilation effects re-

vealed by morphs. Journal of Personality and Social

Psychology, 90(3):399.

Sadovnik, A., Gharbi, W., Vu, T., and Gallagher, A. (2018).

Finding your lookalike: Measuring face similarity

rather than face identity. In Proceedings of the IEEE

Conference on Computer Vision and Pattern Recogni-

tion (CVPR) Workshops.

Santini, S. and Jain, R. (1996). Similarity matching. In

Asian conference on computer vision, pages 571–580.

Springer.

Schroff, F., Treibitz, T., Kriegman, D., and Belongie, S.

(2011). Pose, illumination and expression invariant

pairwise face-similarity measure via doppelg

anger list

comparison. In 2011 International Conference on

Computer Vision, pages 2494–2501.

Torun, R. B., Yurdakul, M., and Duygulu, P. (2009). Find-

ing similar faces. In 2009 IEEE 17th Signal Process-

ing and Communications Applications Conference,

pages 960–963.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

904