Double Trouble? Impact and Detection of Duplicates in Face Image

Datasets

Torsten Schlett

, Christian Rathgeb

, Juan Tapia

and Christoph Busch

da/sec - Biometrics and Security Research Group, Hochschule Darmstadt, Germany

Keywords:

Biometrics, Face Images, Dataset Cleaning, Mislabeling, Image Hash, Face Recognition, Quality Assessment.

Abstract:

Various face image datasets intended for facial biometrics research were created via web-scraping, i.e. the

collection of images publicly available on the internet. This work presents an approach to detect both exactly

and nearly identical face image duplicates, using ﬁle and image hashes. The approach is extended through

the use of face image preprocessing. Additional steps based on face recognition and face image quality as-

sessment models reduce false positives, and facilitate the deduplication of the face images both for intra- and

inter-subject duplicate sets. The presented approach is applied to ﬁve datasets, namely LFW, TinyFace, Adi-

ence, CASIA-WebFace, and C-MS-Celeb (a cleaned MS-Celeb-1M variant). Duplicates are detected within

every dataset, with hundreds to hundreds of thousands of duplicates for all except LFW. Face recognition and

quality assessment experiments indicate a minor impact on the results through the duplicate removal. The ﬁnal

deduplication data is made available at https://github.com/dasec/dataset-duplicates.

1 INTRODUCTION

Face recognition or other facial biometrics research

often involves web-scraped face image datasets, e.g.

to train or evaluate face recognition models. Scraping

face images from the web can accidentally lead to the

inclusion of mislabelled or duplicated images. This

paper presents a duplicate detection approach that

searches for both exact and near duplicates, which is

applied to a selection of ﬁve web-scraped face image

datasets: LFW (Huang et al., 2007), TinyFace (Cheng

et al., 2018), Adience (aligned) (Eidinger et al., 2014),

CASIA-WebFace (Yi et al., 2014), and C-MS-Celeb

(aligned) (Jin et al., 2018). Additional steps are pre-

sented to mitigate false positives and to systematically

deduplicate the datasets through the use of face recog-

nition and face image quality assessment models. The

ﬁnal deduplication data is made publicly available

and the general methodology should be applicable to

any typical web-scraped face image datasets beyond

the ones examined in this paper. The effect of the du-

plicate removal on face recognition and face image

https://orcid.org/0000-0003-0052-2741

https://orcid.org/0000-0003-1901-9468

https://orcid.org/0000-0001-9159-4075

https://orcid.org/0000-0002-9159-2923

https://github.com/dasec/dataset-duplicates

quality assessment experiments is examined as well.

The rest of this paper is structured as follows:

• Related work is discussed in section 2.

• The fundamental duplicate detection approach is

presented in section 3.

• Preservative deduplication steps after the initial

duplicate detection are described in section 4.

• Effects of duplicate removal on face recognition

and face image quality assessment are investi-

gated in section 5.

• Findings are summarized in section 6.

2 RELATED WORK

To the best of our knowledge there is no closely re-

lated facial biometrics work that proposes a simi-

lar face image duplicate detection based on ﬁle and

image hashes (section 3), with subsequent preser-

vative deduplication based on face recognition and

face image quality assessment (section 4), including

an examination of multiple existing separate web-

scraped face image datasets. There are however var-

ious more loosely related works on automated face

image dataset label cleaning.

Schlett, T., Rathgeb, C., Tapia, J. and Busch, C.

Double Trouble? Impact and Detection of Duplicates in Face Image Datasets.

DOI: 10.5220/0012422500003654

In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 801-808

ISBN: 978-989-758-684-2; ISSN: 2184-4313

801

In (Jin et al., 2018) a graph-based label cleaning

method is presented and applied to the MS-Celeb-

1M (Guo et al., 2016) dataset, resulting in the C-MS-

Celeb dataset that is further examined in this work.

This approach ﬁrst extracts face image feature vec-

tors using a pretrained model, based on which simi-

larity graphs are constructed between the images of

each original subject label. The cleanup then consists

of the deletion of insufﬁciently similar graph edges

and the application of a graph community detection

algorithm. Only sufﬁciently large communities are re-

tained. An additional step computes the similarity of

remaining images to the retained community feature

centers, which may lead to the assignment of these

images to a sufﬁciently similar community.

This work in contrast aims to detect face image

duplicates across the examined datasets using gen-

eral ﬁle and image hashes, without any initial reliance

on subject labels or facial feature vectors. The latter

are however used in an additional but technically op-

tional preservative deduplication phase (in addition to

face image quality assessment). As this work exam-

ines C-MS-Celeb itself, it is shown that the original

MS-Celeb-1M (Guo et al., 2016) collection and the

C-MS-Celeb cleaning approach (Jin et al., 2018) did

not sufﬁce to remove a substantial number of appar-

ent duplicates (over 500,000 are removed at the end

of this work in section 4).

Another cleaning approach applied to i.a. MS-

Celeb-1M (Guo et al., 2016) is presented in (Jaza-

ery and Guo, 2019), albeit without a comparison to

(Jin et al., 2018). This other approach ﬁrst selects

a “reference set” of images for each subject, based

on face image quality assessment (and on web search

engine rankings for the images, if available). More

speciﬁcally the images with quality scores above the

mean quality score are selected within the subject

sets. The following cleaning steps then either keep or

discard the other images of each subject’s set, based

on whether each image’s mean similarity score to the

“reference set” images is above a threshold that varies

depending on the image quality score and parameters

conﬁgured by the researchers (through the examina-

tion of a data subset).

This is again in contrast to this work’s duplicate

detection objective, which this label cleaning doesn’t

conceptually address. A relation is the use of both

quality assessment and similarity scores. But in this

work, besides the use of different models, these are

used in the aforementioned preservative deduplication

phase, which builds upon the independent ﬁle/image

hash duplicate detection. The way in which the qual-

ity and similarity scores are considered differs as well,

and e.g. includes the resolution of inter-subject du-

plicates (instead of only ﬁltering within each isolated

image set per subject).

In (Zhang et al., 2012) the “Celebrities on the

Web” dataset was constructed, utilizing names found

in textual data collected alongside the images to estab-

lish initial subject label candidates, with the ﬁnal sub-

ject label assignment being facilitated through the use

of a face image similarity graph based on the initial

labels. This is more distantly related work as it per-

tains to the collection of dataset images from scratch,

instead of duplicate detection or label cleaning for ex-

isting face image datasets, which may not typically

provide the textual data that is used by this approach.

Nevertheless, part of this approach involves the inten-

tional web search for near duplicate images with the

help of image hashing, to gather further text data. The

approach can also handle multiple faces in collected

images.

In this work the base duplicate detection in con-

trast only requires images (without subject labels or

text), which likewise do not necessarily have to be

images that show a single face. The additional preser-

vative deduplication however expects existing sub-

ject labels and assumes that each image is supposed

to show exactly one face, both of which should be

the case for typical already constructed face image

datasets.

3 DUPLICATE DETECTION

This section describes the fundamental duplicate de-

tection approach, which considers exact and near du-

plicates.

3.1 Exact Duplicates

Exact duplicates can technically be found purely via

exact data comparisons. For the sake of computa-

tional efﬁciency we use BLAKE3 (O’Connor et al.,

2021) hashes of the ﬁle data to collect the initial du-

plicate sets. File hash false negative duplicates are

impossible, since the same data always results in the

same hash value, but false positives are technically

possible (hash collisions). Therefore an additional

step which checks for fully identical ﬁle data is em-

ployed within each duplicate set found by the ﬁle hash

step, to ensure that there are no false positives.

Although this kind of exact duplicate check is ar-

guably simple to conduct, exact duplicates were nev-

ertheless found in all of the examined datasets, imply-

ing that no such check was performed as part of the

datasets’ creation.

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

802

Figure 1: Three near duplicate pair examples, all detected

by pHash. From left to right, the ﬁrst example is from

CASIA-WebFace, the second from C-MS-Celeb, and the

third from Adience.

3.2 Near Duplicates

There can also be images that differ slightly, yet are

arguably so similar that they should be counted as

near duplicates for typical facial biometrics research.

While it is well-deﬁned what exact duplicates are,

as ﬁles are either exactly identical or not, there are

many different possible ways in which near duplicates

can be deﬁned. In this paper we employ two image

hash implementations with default settings from the

“ImageHash” Python package at version 4.3.1

, based

on cursory manual examinations of found duplicates

prior to the main experiments: “pHash” (for “percep-

tual hashing”) and “crop-resistant hashing”. See Fig-

ure 1 for near duplicate examples.

The pHash method resulted in more detected du-

plicates across all examined datasets than the crop-

resistant hashing method, and so the latter may be

comparatively less suitable for this particular use case

(i.e. at least with the used default conﬁguration).

As the duplicate sets found by different image

hash functions can overlap, and since image hashes

may also lead to false positive duplicates, additional

set merging and false positive correction should be

employed as part of the ﬁnal preservative deduplica-

tion (i.e. deduplication that aims to keep one image

per duplicate set, instead of simply removing all pos-

sible duplicates). This is further described in subsec-

tion 4.2.

3.3 Preprocessed Face Images

Various face recognition and face image quality as-

sessment models rely on face image preprocessing,

which typically crops and aligns the original face im-

age based on detected facial landmarks. The just

https://pypi.org/project/ImageHash/

Original Face detection Preprocessed

Figure 2: Preprocessing example using an image from the

CASIA-WebFace dataset.

described duplicate detection can optionally be car-

ried out using such preprocessed face images. This

should be done as an additional step after the dupli-

cate detection on the unmodiﬁed original images, as

the facial landmark detection required for preprocess-

ing may fail for some images, and because the pre-

processed/original image variants can both yield du-

plicate sets that the other image variant does not.

In this paper the similarity transformation as em-

ployed for ArcFace (Deng et al., 2019) is used for face

image preprocessing. The SCRFD-10GF model buf-

falo l from InsightFace (Guo et al., 2022)

is used to

obtain the facial landmarks that are required by this

preprocessing step. For images in which multiple

faces are detected, a primary face is selected based

on the detection’s bounding box width and height, its

proximity to the image center, and the detector con-

ﬁdence score. The preprocessed images all have the

same 112×112 width and height. Figure 2 shows a

preprocessing example.

Face images for which the landmark detection

failed amount to 0 for LFW, 859 for TinyFace (only

4 of which are duplicates based on the detection on

the original images), 79 for Adience (9 of which are

duplicates), 129 for CASIA-WebFace (2 of which are

duplicates), and 6,179 for C-MS-Celeb (351 of which

are duplicates). These images are simply not consid-

ered by this additional duplicate detection step.

3.4 Examined Dataset Duplicates

Table 1 shows an overview of the examined datasets,

including the total image counts and the duplicate

counts. Intra-subject duplicates are duplicates found

within the same subject. I.e. the identical or very sim-

ilar images belong to the same subject (are stored as

mated samples), and are presumed to stem from dif-

ferent capture attempts or sessions, but in fact stem

from the very same capture attempt. Inter-subject du-

plicates on the other hand are found across different

subjects.

LFW, TinyFace, and Adience each have under

20,000 face images. CASIA-WebFace is substantially

https://github.com/deepinsight/insightface/tree/master

/python-package

Double Trouble? Impact and Detection of Duplicates in Face Image Datasets

803

larger with 494,414 face images, which is however

still closer to the three smaller datasets than to C-MS-

Celeb with its 6,464,016 face images. The number of

different subjects does not increase as sharply as the

number of total face images for the larger datasets.

Among the datasets, LFW is the only one with a

very low number of duplicates. The evaluation in sec-

tion 5, which will investigate the impact on facial bio-

metrics, consequently omits the LFW dataset.

The TinyFace dataset includes 153,428 non-face

images, which are not examined in this paper. Only

the remaining 15,975 face images are considered.

For the Adience dataset we more speciﬁcally use

the “aligned” version of the face images.

Although the absolute number of duplicates for

CASIA-WebFace is higher than for the smaller

datasets, the duplicate percentage with respect to the

face image total is relatively low (excluding LFW).

The C-MS-Celeb dataset is a subset of MS-Celeb-

1M (Guo et al., 2016) with cleaned subject labels. In

this paper the “aligned” image variants are used. Both

the absolute number of duplicates as well as the du-

plicate percentage with respect to the face image to-

tal is the highest among all examined datasets. For

this dataset 33,918 inter-subject duplicates simulta-

neously are intra-subject duplicates. The combined

count of images that are part of some duplicate set is

thus 885,476. This intra-/inter-subject duplicate over-

lap did not occur for the other datasets, so the com-

bined count for each of these simply is the sum of the

intra- and inter-subject counts.

Besides duplicate detection within each dataset,

the approach can also be applied to check for overlap

between the datasets. No such inter-dataset duplicate

cases were detected using the original images. Some

cases were however found using the preprocessed im-

ages, and a subset has been manually conﬁrmed to be

true positives (predominantly but not exclusively be-

tween CASIA-WebFace and C-MS-Celeb). Although

a more extensive dataset overlap investigation is out-

side the scope of this work, this does indicate that the

approach of this paper could be used as a part of fu-

ture work on this topic.

4 PRESERVATIVE

DEDUPLICATION

This section describes additional preservative dedu-

plication steps that continue from the fundamental du-

plicate detection of the prior section. These steps keep

one selected image per duplicate set (subsection 4.1,

subsection 4.3) and mitigate false positive duplicate

detections (subsection 4.2). They also move the kept

image of an inter-subject duplicate set to the most ﬁt-

ting subject within that set, based on similarity scores

(subsection 4.4). The ﬁnal image removal and im-

age subject-move counts for the examined datasets are

listed in Table 2. These counts correspond to the pub-

licly available deduplication lists

4.1 Exact Intra-Subject Deduplication

The deduplication of duplicate sets that consist of ex-

actly identical images all within the same subject is

trivial, as any one of the identical images could be

randomly selected. For the sake of reproducibility, we

more speciﬁcally sort the images in ascending lexical

ﬁle path order and select the ﬁrst one to be kept as the

deduplicated image.

The deduplication of exact inter-subject duplicate

sets is less straightforward, as there are multiple dif-

ferent subject candidates to which the deduplicated

image could be assigned to. This is thus addressed

separately in subsection 4.4.

4.2 False Positive Correction

The image hashes used for near duplicate detection

as introduced in subsection 3.2 can lead to false pos-

itives, and these false positive duplicate sets can even

consist of images that a manual inspection can eas-

ily identify as different faces. To avoid the unde-

sired exclusion of these false positive images from the

cleaned datasets, an additional false positive correc-

tion based on face recognition is applied.

The complete near duplicate detection thus works

as follows: Near duplicate sets are ﬁrst found sepa-

rately by the image hash functions. As different image

hash functions can yield overlapping duplicate sets,

the next step is to merge any overlapping sets, the re-

sult being disjunct near duplicate sets.

Then the false positive correction is carried out on

each near duplicate set. This correction step forms all

comparison pairs between the images within a dupli-

cate set, and then ﬁlters out any images of pairs for

which the similarity score is below a set threshold.

This work employs the publicly available

Mag-

Face (Meng et al., 2021) model with iResNet100

backbone trained on MS1MV2 (another variant of

MS-Celeb-1M (Guo et al., 2016)) for face recognition

(similarity scores), as well as for quality assessment

in the following parts, as this model can be employed

for both use cases. The similarity score threshold for

the false positive correction is set to 0.40 based on the

https://github.com/dasec/dataset-duplicates

https://github.com/IrvingMeng/MagFace

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

804

Table 1: Image, subject, and duplicate counts for the examined datasets prior to the preservative deduplication in section 4.

Intra: Intra-subject duplicates (all images in duplicate sets belonging to only one subject). Subjects-w.-intra: Subjects with at

least one intra-subject duplicate. Inter: Inter-subject duplicates (all images in duplicate sets belonging to multiple subjects).

Subjects-w.-inter: Subjects with at least one inter-subject duplicate.

Dataset Images Subjects Intra Subjects-w.-intra Inter Subjects-w.-inter

LFW 13,233 5,749 6 3 6 6

TinyFace 15,975 5,139 662 286 53 52

Adience 19,370 2,284 1,609 274 4 4

CASIA-WebFace 494,414 10,575 9,614 2,677 288 186

C-MS-Celeb 6,464,016 94,682 753,277 76,841 166,117 27,895

Table 2: The number of images that the preservative dedu-

plication (section 4) removed from the dataset, and the num-

ber of images that were moved to a new subject (subsec-

tion 4.4). Note that the latter can coincide with the removal

of duplicates already present in the move’s target subject.

Dataset Removed Moved

LFW 9 0

TinyFace 354 0

Adience 912 0

CASIA-WebFace 5,032 37

C-MS-Celeb 531,018 13,175

manual inspection of a randomly selected duplicate

candidate subset.

4.3 Quality-Based Deduplication

After the false positive correction step described in

subsection 4.2, the images in each near duplicate set

are sorted in descending quality score order. The ﬁrst

image is then selected as the deduplicated image (i.e.

the one with the highest quality score).

These quality scores are obtained using the same

MagFace model as in subsection 4.2. A negative in-

ﬁnity quality score is used for any image for which

no quality score could be computed, meaning that all

images with computed scores are preferred. The as-

cending lexical ﬁle path order serves as a tie breaker

if there are no images with computed quality scores,

or if the best two quality scores are identical.

For intra-subject deduplication the selection of the

deduplicated image concludes the preservative dedu-

plication. For inter-subject deduplication see the fol-

lowing subsection 4.4, as additional work is required

to decide to which subject the image should be as-

signed to (if any).

4.4 Comparison-Based Inter-Subject

Deduplication

Inter-subject deduplication is less straightforward

than intra-subject deduplication, as there are multi-

ple possible subjects to which the deduplicated im-

age could be assigned to per duplicate set. To resolve

this issue, the deduplicated image of a duplicate set

is compared against the non-duplicate images of all

subjects that are involved in the duplicate set.

Here “non-duplicate images” refers to those im-

ages that were never assigned to any duplicate set by

either the exact or near duplicate detection. If no such

images remain for a candidate subject, or if no face

recognition feature extraction could be applied to any

of these images, then the subject is excluded from

consideration. If no candidate subjects remain, then

the deduplicated image is excluded from the dataset.

The comparison of the deduplicated image to each

candidate subject’s non-duplicate image set is again

using the same MagFace model as in subsection 4.2.

For each subject the mean similarity score is com-

puted across the comparisons to the deduplicated im-

age. The subject with the highest mean similarity

score is selected as the potential target subject for the

deduplicated image.

However, if the mean similarity score of the poten-

tial target subject is below a set threshold, the dedupli-

cated image is instead excluded from the dataset. This

is done to avoid the assignment of the deduplicated

image to an incorrect subject, even if that subject is

the best candidate among the duplicate set’s subjects.

For this the same 0.40 similarity score threshold as in

the false positive correction in subsection 4.2 is used.

Additionally, if the absolute difference between

the mean similarity score of the potential target sub-

ject and the second best candidate subject is below

another set threshold, the deduplicated image is also

excluded from the dataset, to avoid uncertain subject

assignments. For this we select the similarity score

threshold 0.20 based on manual sample observations

and based on the statistics of the computed mean sim-

ilarity scores.

Double Trouble? Impact and Detection of Duplicates in Face Image Datasets

805

5 EFFECTS ON FACIAL

BIOMETRICS

This section investigates how deduplication alters the

results for face recognition and face image quality as-

sessment experiments. Two forms of deduplication

are considered: One is the complete removal of all im-

ages that are involved in any duplicate set according

to the approach described in section 3, and the other is

the preservative deduplication described in section 4.

The used face recognition / quality assessment

models rely on the same face image preprocessing as

described in subsection 3.3 for the preprocessed im-

age duplicate detection step. Face images for which

the landmark detection failed are excluded from the

experiments in this section.

5.1 Face Recognition

The following face recognition experiments require

the selection of mated and non-mated (ISO/IEC JTC1

SC37 Biometrics, 2022) pairs of face images.

For the mated pairs, one approach could be to

select all possible pairs for each subject. This may

however result in a substantially increased number of

mated pairs for subjects with comparatively larger im-

age counts within a dataset, since a subject with N

images would yield (N · (N − 1))/2 mated pairs, and

each image would consequently be involved in N − 1

mated pairs.

All mated pairs

per subject:

“Circular” mated

pairs per subject:

Figure 3: Two mated pair selection approaches. The num-

bered graph nodes represent the ordered face images of a

single subject, with the graph edges representing the se-

lected mated pairs.

But since these experiments are supposed to com-

pare results for the datasets with and without dupli-

cates, a balanced number of mated pairs per image is

preferred. Mated pairs are thus instead selected in a

“circular” manner for each subject in a dataset, as il-

lustrated in Figure 3. This simply means that every

image with index i forms a mated pair with the next

image with index i + 1. The last image also forms a

pair with the ﬁrst image if the subject has more than

two images. A subject with more than two images

thus results in a number of mated pairs equal to the

number of images, and a subject with exactly two im-

ages results in a single mated pair. Each image is con-

sequently always involved in exactly two mated pairs

for subjects with more than two images, or one mated

pair for subjects with two images. The lexicograph-

ically ascending order of the image paths is used as

the image order. An additional beneﬁt of this “cir-

cular” mated pair selection is the reduction of the re-

quired computational resources for the experiments,

since the number of all possible mated pairs would be

substantially higher (over 300 million in total).

The following lists the mated pair count per

dataset, as well as the number of subjects implicitly

excluded due to only containing a single image:

• TinyFace: 11,881 (153 excluded subjects).

• Adience: 18,093 (815 excluded subjects).

• CASIA-WebFace: 494,284 (0 excluded subjects).

• C-MS-Celeb: 6,457,562 (123 excluded subjects).

A number of non-mated pairs equal to the number

of mated pairs is randomly selected per dataset.

The MagFace model previously described in sub-

section 4.2 is employed for face recognition. The sim-

ilarity scores are then used to assess the face recog-

nition performance in terms of the False Non-Match

Rate (FNMR), the False Match Rate (FMR), and the

Equal Error Rate (EER) (ISO/IEC JTC1 SC37 Bio-

metrics, 2022). Table 3 shows the results. There are

mostly minor differences between the variants with

and without duplicates, and duplicate removal both

increased and decreased error rates in different cases.

5.2 Quality Assessment

The effect of the duplicate removal on face image

quality assessment is examined in terms of the change

of partial Area Under Curve (pAUC) values for Error

versus Discard Characteristic (EDC) curves (Schlett

et al., 2023; Grother and Tabassi, 2007), using the

FNMR and FMR as the EDC errors. This evalua-

tion is based both on quality scores and the previ-

ously computed similarity scores. The quality scores

are computed by two state-of-the-art models, which

aim to assess the face image quality in terms of the

biometric utility for face recognition (ISO/IEC JTC1

SC37 Biometrics, 2022). The ﬁrst model is again the

MagFace model described in subsection 4.2, and the

second model is CR-FIQA(L) (Boutros et al., 2023),

which is likewise publicly available

. Table 4 and

Table 5 show the results using MagFace and CR-

FIQA(L), respectively. The pAUC values are com-

puted for the [0%,20%] discard fraction range instead

https://github.com/fdbtrs/CR-FIQA

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

806

Table 3: Face recognition performance in terms of the Equal Error Rate (EER) and the False Non-Match Rate (FNMR) at

ﬁxed approximated False Match Rate (FMR) values. The “Original” variant refers to the unmodiﬁed datasets, the “Full”

variant to the removal of all images involved in duplicate sets as per section 3, and the “Preservative” variant to the image

removal and subject-reassignment as per section 4.

Dataset Variant EER FNMR @ FMR ≈ 1e − 3 FNMR @ FMR ≈ 1e − 2

TinyFace Original 9.5783% 30.0564% 18.8115%

TinyFace Full 9.7627% 31.1831% 19.8787%

TinyFace Preservative 9.7489% 31.4489% 19.6862%

Adience Original 1.7576% 3.0122% 2.0063%

Adience Full 1.7527% 2.9770% 2.0120%

Adience Preservative 1.7539% 3.0017% 2.0172%

CASIA-WebFace Original 7.1955% 10.2724% 8.7041%

CASIA-WebFace Full 7.2278% 10.3316% 8.7493%

CASIA-WebFace Preservative 7.2100% 10.3130% 8.7240%

C-MS-Celeb Original 7.0506% 12.8246% 10.0817%

C-MS-Celeb Full 6.9599% 12.6121% 9.9401%

C-MS-Celeb Preservative 6.9094% 15.4464% 10.0007%

Table 4: FNMR and FMR EDC pAUC values for the

[0%,20%] comparison discard fraction range, using Mag-

Face for quality assessment.

Dataset Variant FNMR FMR

TinyFace Original 0.859% 0.977%

TinyFace Full 0.864% 0.972%

TinyFace Preserv. 0.865% 0.973%

Adience Original 0.708% 0.992%

Adience Full 0.731% 0.994%

Adience Preserv. 0.726% 0.996%

CASIA-WebFace Original 0.820% 1.002%

CASIA-WebFace Full 0.820% 1.002%

CASIA-WebFace Preserv. 0.820% 1.002%

C-MS-Celeb Original 0.814% 1.000%

C-MS-Celeb Full 0.804% 1.000%

C-MS-Celeb Preserv. 0.802% 1.002%

of the full EDC curves, since higher discard fractions

are not usually considered as operationally relevant

(Schlett et al., 2023). Similarly to the face recognition

results of the prior subsection, there are minor differ-

ences which include both increases and decreases in

the error values.

6 SUMMARY

The presented exact and near face image duplicate

detection approach based on ﬁle hashes and image

hashes (section 3), including the additional use of pre-

processed face images (subsection 3.3), found dupli-

cates in all ﬁve examined web-scraped datasets (sub-

section 3.4). With the exception of the LFW dataset,

over 1% of all images in each original dataset are

duplicate candidates, which ranges from hundreds to

hundreds of thousands of duplicate candidates in ab-

Table 5: FNMR and FMR EDC pAUC values for the

[0%,20%] comparison discard fraction range, using CR-

FIQA(L) for quality assessment.

Dataset Variant FNMR FMR

TinyFace Original 0.801% 0.890%

TinyFace Full 0.803% 0.890%

TinyFace Preserv. 0.809% 0.885%

Adience Original 0.865% 0.958%

Adience Full 0.880% 0.959%

Adience Preserv. 0.875% 0.957%

CASIA-WebFace Original 0.799% 1.000%

CASIA-WebFace Full 0.800% 0.999%

CASIA-WebFace Preserv. 0.799% 0.999%

C-MS-Celeb Original 0.812% 0.997%

C-MS-Celeb Full 0.800% 0.997%

C-MS-Celeb Preserv. 0.798% 0.999%

solute numbers (subsection 3.4). While most du-

plicate set images belong to a single dataset subject

(intra-subject duplicates), there also are some that be-

long to multiple (inter-subject duplicates), especially

in the C-MS-Celeb dataset (subsection 3.4).

Preservative deduplication steps (section 4) were

applied after the initial duplicate detection. This com-

prised the mitigation of false positive duplicate detec-

tion, the selection of the highest quality face images

per duplicate set (according to a face image quality as-

sessment model), and the assignment of deduplicated

face images to the most ﬁtting subject within inter-

subject duplicate sets (or none if uncertain). The ﬁnal

deduplication data for the examined datasets is pub-

licly available

Minor effects due to duplicate removal were ob-

served on face recognition and face image quality as-

https://github.com/dasec/dataset-duplicates

Double Trouble? Impact and Detection of Duplicates in Face Image Datasets

807

sessment results in the experiments (section 5).

Finding duplicates in all examined datasets does

indicate that such accidental inclusion of duplicates

could be a common occurrence for web-scraped face

image datasets, so that any potential future dataset

construction of this kind should consider implement-

ing a duplicate ﬁlter. It may further be sensible to

examine other existing web-scraped datasets, as they

could likewise contain duplicates.

ACKNOWLEDGEMENTS

This research work has been funded by the German

Federal Ministry of Education and Research and the

Hessian Ministry of Higher Education, Research, Sci-

ence and the Arts within their joint support of the

National Research Center for Applied Cybersecurity

ATHENE. This project has received funding from the

European Union’s Horizon 2020 research and inno-

vation programme under grant agreement No 883356.

This text reﬂects only the author’s views and the Com-

mission is not liable for any use that may be made of

the information contained therein.

REFERENCES

Boutros, F., Fang, M., Klemt, M., Fu, B., and Damer, N.

(2023). CR-FIQA: Face image quality assessment

by learning sample relative classiﬁability. In Conf.

on Computer Vision and Pattern Recognition (CVPR),

pages 5836–5845. IEEE.

Cheng, Z., Zhu, X., and Gong, S. (2018). Low-resolution

face recognition. In Asian Conf. on Computer Vision

(ACCV).

Deng, J., Guo, J., and Zafeiriou, S. (2019). ArcFace: Ad-

ditive angular margin loss for deep face recognition.

In Conf. on Computer Vision and Pattern Recognition

(CVPR).

Eidinger, E., Enbar, R., and Hassner, T. (2014). Age and

gender estimation of unﬁltered faces. IEEE Trans.

on Information Forensics and Security, 9(12):2170–

2179.

Grother, P. and Tabassi, E. (2007). Performance of biomet-

ric quality measures. IEEE Trans. on Pattern Analysis

and Machine Intelligence, 29(4):531–543.

Guo, J., Deng, J., Lattas, A., and Zafeiriou, S. (2022). Sam-

ple and computation redistribution for efﬁcient face

detection. In Intl. Conf. on Learning Representations

(ICLR).

Guo, Y., Zhang, L., Hu, Y., He, X., and Gao, J. (2016). MS-

Celeb-1M: A dataset and benchmark for large-scale

face recognition. Proc. 14th European Conf. on Com-

puter Vision.

Huang, G. B., Ramesh, M., Berg, T., and Learned-Miller,

E. (2007). Labeled faces in the wild: A database for

studying face recognition in unconstrained environ-

ments. Technical report, University of Massachusetts,

Amherst.

ISO/IEC JTC1 SC37 Biometrics (2022). ISO/IEC 2382-

37:2022 Information technology - Vocabulary - Part

37: Biometrics. International Organization for Stan-

dardization.

Jazaery, M. A. and Guo, G. (2019). Automated cleaning of

identity label noise in a large face dataset with quality

control. IET Biometrics, 9(1):25–30.

Jin, C., Jin, R., Chen, K., and Dou, Y. (2018). A com-

munity detection approach to cleaning extremely large

face database. Computational intelligence and neuro-

science.

Meng, Q., Zhao, S., Huang, Z., and Zhou, F. (2021). Mag-

Face: A universal representation for face recognition

and quality assessment. In Conf. on Computer Vi-

sion and Pattern Recognition (CVPR), pages 14225–

14234. IEEE.

O’Connor, J., Aumasson, J., Neves, S., and Wilcox-

O’Hearn, Z. (2021). BLAKE3. https://github.com

/BLAKE3-team/BLAKE3/.

Schlett, T., Rathgeb, C., Tapia, J., and Busch, C. (2023).

Considerations on the evaluation of biometric quality

assessment algorithms. IEEE Trans. on Biometrics,

Behavior, and Identity Science (TBIOM).

Yi, D., Lei, Z., Liao, S., and Li, S. Z. (2014). Learning face

representation from scratch.

Zhang, X., Zhang, L., Wang, X.-J., and Shum, H.-Y. (2012).

Finding celebrities in billions of web images. IEEE

Trans. on Multimedia, 14(4):995–1007.

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

808