CoChat
Sign Up
Cited by 26Face recognition and analysisFace Recognition and PerceptionForensic Anthropology and Bioarchaeology Studies

The DOI matches its Crossref record; 108 of the 108 references it cites were found in a registry or an index; 23 of the 123 sources its claims cite were checked for support.

Forensic Facial Comparison: Current Status, Limitations, and Future Directions

Open access
Bacci, Nicholas, Davimes, Joshua G., Steyn, Maryna, Briers, Nanette·2021·Biology
DOI 10.3390/biology10121269PMC PMC8698381
Open this paper in CoChat — free accountOpen PDF

Abstract

Global escalation of crime has necessitated the use of digital imagery to aid the identification of perpetrators. Forensic facial comparison (FFC) is increasingly employed, often relying on poor-quality images. In the absence of standardized criteria, especially in terms of video recordings, verification of the methodology is needed. This paper addresses aspects of FFC, discussing relevant terminology, investigating the validity and reliability of the FISWG morphological feature list using a new South African database, and advising on standards for CCTV equipment. Suboptimal conditions, including poor resolution, unfavorable angle of incidence, color, and lighting, affected the accuracy of FFC. Morphological analysis of photographs, standard CCTV, and eye-level CCTV showed improved performance in a strict iteration analysis, but not when using analogue CCTV images. Therefore, both strict and lenient iterations should be conducted, but FFC must be abandoned when a strict iteration performs worse than a lenient one. This threshold ought to be applied to the specific CCTV equipment to determine its utility. Chance-corrected accuracy was the most representative measure of accuracy, as opposed to the commonly used hit rate. While the use of automated systems is increasing, trained human observer-based morphological analysis, using the FISWG feature list and an Analysis, Comparison, Evaluation, and Verification (ACE-V) approach, should be the primary method of facial comparison.

Overview

AI snapshot

Written from the paper’s full text.

TL;DR

This review summarizes validation studies of forensic facial comparison (FFC) by morphological analysis (MA) using the FISWG feature list and ACE-V, finding high accuracy under optimal and eye-level CCTV conditions but poor performance with analogue CCTV. It recommends trained human observer-based MA as the primary method, with strict/lenient iterations and chance-corrected accuracy, while noting CCTV image quality and disguises as major limitations.

Key findings

  • MA using the FISWG feature list was highly accurate and reliable for optimal standardized photographs versus unstandardized photographs: chance-corrected accuracy 99.1% and reliability 92.1% (75 face sets, n = 750 comparisons). Standard digital CCTV was lower: 82.6% accuracy and 74.3% reliability (100 face sets, n = 1000), while eye-level digital CCTV was nearly as good as photographs: 97.3% accuracy and 77.3% reliability (95 face sets, n = 950).
  • Analogue CCTV produced much poorer results: accuracy as low as 33.1%, false negative rate 75.2%, and reliability 37.8% (130 face sets, n = 1300). A strict iteration performed worse than a lenient iteration only for analogue CCTV, which the authors propose as a marker that a CCTV installation is below a usable threshold for FFC.
  • Disguise effects were mixed: sunglasses yielded high performance at 90.4% (n = 390), surpassing standard CCTV without sunglasses (82.6%), whereas brimmed caps yielded a low chance-corrected accuracy of 68.1% (n = 420). Low false positive rates (<1.6%) were reported for optimal photographs and standard CCTV installations.
  • Chance-corrected accuracy (weighted Cohen’s kappa with squared weighting) was the most representative accuracy measure, unlike hit rate or balanced accuracy, and both strict and lenient iterations should be conducted; FFC should be abandoned when a strict iteration performs worse than a lenient one.

Methods

  • Review and summary of the authors’ validation studies plus relevant technical literature on FFC and CCTV. The Wits Face Database included 622 unique African male individuals aged 18–35, each with 10 photographs (6220 facial photographs) across controlled and uncontrolled conditions; 337 participants (54.2%) were also recorded under CCTV conditions, including standard digital IP, eye-level digital IP, analogue CCTV, and with caps or sunglasses.
  • MA was conducted using the FISWG morphological feature list and an ACE-V approach. Comparisons were arranged as face sets containing no-match and positive-match trials, with reported sample sizes of 75 sets (n = 750), 100 sets (n = 1000), 130 sets (n = 1300), 95 sets (n = 950), sunglasses n = 390, and brimmed caps n = 420.
  • Main analysis compared strict versus lenient iterations and used chance-corrected accuracy via weighted Cohen’s kappa against true match-up information. The paper also reviews CCTV installation standards, image quality factors, compression, and pixel-density thresholds.

Limitations

  • Stated: the studies simulated real-world conditions in a select number of scenarios, with pre-set conditions limited to the broad questions investigated; it was not possible to isolate which specific limiting factor (resolution, angle of incidence, color, lighting, subject-to-camera distance, etc.) contributed most to poor MA performance. The database currently includes only males, no female database exists, only two CCTV camera specifications were tested, and infrared CCTV conditions were not tested.
  • Stated: image quality depends on multiple interacting factors, and in some circumstances these factors cannot be isolated from one another; future targeted testing is needed to determine exact contributions and thresholds. No formal standardized training or certification platforms for FFC were identified.
  • Evident from the text: the paper is a narrative review of the authors’ own studies and selected literature rather than a systematic review or meta-analysis, so findings may not generalize to other populations, camera systems, or disguises; face masks, for example, are noted as not yet tested for human observer-based MA.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/biology10121269

Cites

108 of 108 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

23 of 123 cited sources checked · 4 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Facial comparison accuracy and data quality are directly correlated.

    “Facial comparison accuracy and data quality are, thus, directly correlated [95,96], especially in terms of individual accuracy variation across multiple analysts [97] and individual analyst ability overestimation [98].”
    Quote in the paperSubstantial claim3 of 4 sources checked
    • cites [95] Fysh, 2018, The Kent Face Matching Test, Br. J. Psychol. source

      The Accuracy section (page unavailable) reports 80% accuracy on the GFMT’s optimised images versus 66% on the more variable KFMT images. The General Discussion attributes the KFMT’s greater difficulty to variability in its face photographs. This supports an association between image conditions and accuracy, but the paper does not establish a direct correlation between data quality and accuracy. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [96] Kramer, 2019, Unfamiliar Face Matching With Driving Licence and Passport Photographs, Perception source
    • cites [97] Burton, 2010, The Glasgow Face Matching Test, Behav. Res. Methods source

      The General Discussion (page unavailable) says the test uses high-quality, well-lit images to avoid making matching artificially difficult and to optimize accuracy. It does not compare levels of data quality or establish a direct correlation between quality and facial-comparison accuracy. (LLM, conf=0.76)

      Experimental — accuracy not yet measured
    • cites [98] Bindemann, 2014, Perceived ability and actual recognition accuracy for unfamiliar and famous faces, Cogent Psychol. source

      The citation [98] is attached to the point about individual ability overestimation, not the earlier claim about data quality. The paper’s Discussion (§5; page unavailable) says observers’ initial judgments of their ability to recognize unfamiliar faces poorly predict their accuracy. It does not establish that the observers were analysts or that they specifically overestimated their ability rather than misjudged it in either direction. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Premisecitation

    There is a concerning lack of standardization of required CCTV installation, recording conditions, and image quality.

    “While this global increase in CCTV data is beneficial to criminal investigation and facial comparison, there is a concerning lack of standardization of required installation, recording conditions, and image quality [20,82,89,90,91,92].”
    Quote in the paperSubstantial claim3 of 6 sources checked
    • cites [20] Stephan, 2019, An overview of the latest developments in facial imaging, Forensic Sci. Res. source

      The Face mapping (photo-comparison) section says differences in recording conditions and image resolution are common and complicate forensic comparisons; it also says ideal image-quality and photographic conditions are rarely met. It does not address a lack of standards for CCTV installation requirements specifically. No page number is available in our copy. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [82] Lee, 2009, Matching unfamiliar faces from poor quality closed-circuit television (CCTV) footage: An evaluation of the effect of training on facial identification ability, Axis Online J. CAHId
    • cites [89] Kleinberg, 2007, Failure of anthropometry as a facial identification technique using high-quality photographs, J. Forensic Sci. source
    • cites [90] Keval, 2008, Can we ID from CCTV? Image quality in digital CCTV and face identification performance, Mob. Multimedia Image Process. Secur. Appl. source

      Section 1.2 says there were no tested guidelines for digital CCTV video quality and storage, and that existing analogue guidance did not address digital recording and transmission. Section 5.4 describes varying system configurations and recommends testing them for their intended use. The paper does not establish a lack of required standards for CCTV installation or recording conditions generally. (LLM, conf=0.82)

      Experimental — accuracy not yet measured
    • cites [91] Smith, 2008, Colour analysis and verification of CCTV images under different lighting conditions, Image Qual. Syst. Perform. V source
    • cites [92] Bindemann, 2013, The effect of image pixelation on unfamiliar-face matching, Appl. Cogn. Psychol. source

      The Introduction (page not available) says CCTV image quality is affected by factors such as lighting and contrast, and that the factors and levels of degradation are often difficult to define. The paper does not establish a lack of required standards for CCTV installation or recording conditions. (LLM, conf=0.78)

      Experimental — accuracy not yet measured
  • Premisecitation

    There are large discrepancies between informal training courses in improvement of facial examiner expertise.

    “A recent study on the performance of informal training courses on facial comparison suggested that there are large discrepancies between courses in improvement of facial examiner expertise [134].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [134] Towler, A., Kemp, R.I., Mike Burton, A., Dunn, J.D., Wayne, T., Moreton, R., and White, D. (2019). Do professional facial image comparison training courses work?. PLoS ONE, 14. source

      The Abstract and General discussion report that short courses did not improve identification accuracy, while the three-day course produced small, inconsistent improvements. They do not establish large discrepancies between courses in improving examiner expertise; the three-day course did not produce the qualitative changes associated with that expertise. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
  • Premisecitation

    Classification schemes are viewed as prone to high inter-observer error.

    “This descriptive approach is preferred as classification schemes are viewed as prone to high inter-observer error [55,59,109].”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [55] Milborrow, S., Morkel, J., and Nicolls, F. (2008). The MUCT Landmarked Face Database. Pattern Recognit. Assoc. S. Afr., Available online: http://www.milbo.org/muct/.
    • cites [59] Sagonas, 2015, 300 Faces In-The-Wild Challenge: Database and results, Image Vis. Comput. source

      Section 3.5 (Discussion; page not available) reports high variation among annotators for facial landmarks without clear semantic meaning, particularly the chin, but high agreement for eye and mouth landmarks. It does not assess classification schemes or say that such schemes generally have high inter-observer error. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [109] Blunden, B. (2009, January 25–30). Anti-Forensics: The Rootkit Connection. Proceedings of the Black Hat USA 2009, Las Vegas, NV, USA.
  • Premisecitation

    A minimum horizontal pixel count of 10-16 per face for a known face and 20 pixels for an unknown face is considered the bare minimum for successful identification in frontal view.

    “A minimum horizontal pixel count of 10–16 per face for a known face [121,123,124] and 20 pixels for an unknown face [92] is considered the bare minimum for successful identification in frontal view.”
    Quote in the paperNot a substantial claim2 of 4 sources checked
    • cites [121] Bachmann, 1991, Identification of spatially quantised tachistoscopic images of faces: How many pixels does it take to carry identity?, Eur. J. Cogn. Psychol. source
    • cites [123] Costen, 1994, Spatial Content and Spatial Quantisation Effects in Face Recognition, Perception source
    • cites [124] Costen, 1996, Effects of high-pass and low-pass spatial filtering on face identification, Percept. Psychophys. source

      The Results report that identification of pixelized faces declined below 23 horizontal pixels per face, while the Discussion places the useful range at 16–32 pixels per face. The paper therefore supports the importance of resolution around 16 pixels, but not 10–16 pixels as a general bare minimum for identifying a known face. The separate 20-pixel claim for unknown faces is attached to citation [92], not this source. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [92] Bindemann, 2013, The effect of image pixelation on unfamiliar-face matching, Appl. Cogn. Psychol. source

      For this source, the citing sentence attributes the 20-pixel figure for unknown faces. In Experiment 1, the Results report that matching accuracy exceeded chance at 20 horizontal pixels per face but not at 14 or 8. That supports 20 pixels as a threshold under those test conditions, not as a general bare minimum for successful identification: overall accuracy remained above chance at lower resolutions, and the paper studied matching an unfamiliar face against a high-quality photograph. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • It may state two separate claims as one. 30
  • Probably not a substantial claim. 17
  • The quote only partly states this claim. 1

  • It may state two separate claims as one.

    Findingcitation

    Approximately 21.2% loss of IP CCTV data and approximately 3.6% loss of analogue CCTV data occurred during the establishment of the Wits Face Database.

    “Our studies [41] found approximately 21.2% loss of IP CCTV data and approximately 3.6% loss of analogue CCTV data during the establishment of the Wits Face Database (Table 1).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [41] Bacci, 2021, Development of the Wits Face Database: An African database of high-resolution facial photographs and multimodal closed-circuit television (CCTV) recordings, F1000Research source

  • It may state two separate claims as one.

    Premisecitation

    Keval and Sasse found that the number of correct identifications of faces by untrained viewers decreased by 12-18% as MPEG-4 quality decreased and by 4-6% as Wavelet quality decreased.

    “Keval and Sasse [90] found that the number of correct identifications of faces by untrained viewers decreased by 12–18% as MPEG-4 quality decreased and by 4–6% as Wavelet quality decreased (92–32 Kbps for both compression formats).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [90] Keval, 2008, Can we ID from CCTV? Image quality in digital CCTV and face identification performance, Mob. Multimedia Image Process. Secur. Appl. source

      The Abstract reports decreases from 92 to 32 Kbps for both formats, but its figures are 12 correct identifications (about 18%) for MPEG-4 and 4 correct identifications (about 6%) for Wavelet—not percentage ranges of 12–18% and 4–6%. Section 4.7 also says the Wavelet change was not statistically significant. (LLM, conf=0.89)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    Automated facial identification systems are increasing in performance.

    “such as the increasing performance of automated systems [31,75,76]”
    Quote in the paperNot a substantial claim2 of 3 sources checked
    • cites [31] Kaur, 2020, Facial-recognition algorithms: A literature review, Med. Sci. Law source
    • cites [75] Ngan, M., Grother, P., and Hanaoka, K. (2020). Ongoing Face Recognition Vendor Test (FRVT) Part 6A: Face Recognition Accuracy with Masks Using Pre-COVID-19 Algorithms. source
    • cites [76] Ngan, M., Grother, P., and Hanaoka, K. (2020). Ongoing Face Recognition Vendor Test (FRVT) Part 6B: Face Recognition Accuracy with Face Masks Using Post-COVID-19 Algorithms. source

  • It may state two separate claims as one.

    Premisecitation

    Most systems are notably still lacking in applicability for facial comparison and are primarily disadvantaged not by installation and recording conditions but by image quality.

    “Most systems are notably still lacking in applicability for facial comparison and are primarily disadvantaged not by installation and recording conditions but by image quality [20,82,89,90,92].”
    Quote in the paperNot a substantial claim3 of 5 sources checked
    • cites [20] Stephan, 2019, An overview of the latest developments in facial imaging, Forensic Sci. Res. source

      The Face mapping (photo-comparison) section describes weak validation and practical limits on facial comparisons, supporting the claim that methods have limited applicability. But it identifies both image quality and photographic conditions as important, explicitly calling replication of camera conditions crucial. It does not support the claim that image quality is the primary disadvantage rather than recording conditions. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [82] Lee, 2009, Matching unfamiliar faces from poor quality closed-circuit television (CCTV) footage: An evaluation of the effect of training on facial identification ability, Axis Online J. CAHId
    • cites [89] Kleinberg, 2007, Failure of anthropometry as a facial identification technique using high-quality photographs, J. Forensic Sci. source
    • cites [90] Keval, 2008, Can we ID from CCTV? Image quality in digital CCTV and face identification performance, Mob. Multimedia Image Process. Secur. Appl. source

      The Introduction (page unavailable) says poor-quality CCTV images make unfamiliar-face identification extremely difficult, and §5.2 reports that identification performance deteriorated as video quality fell. But the Introduction also says systems are often poorly set up, producing those images. The paper does not establish that most systems are unsuitable for facial comparison or that image quality, rather than installation and recording conditions, is the primary disadvantage. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [92] Bindemann, 2013, The effect of image pixelation on unfamiliar-face matching, Appl. Cogn. Psychol. source

      The General Discussion reports that pixelation substantially impairs unfamiliar-face matching and makes comparison of CCTV footage with suspect photographs difficult. It does not assess whether most systems are unsuitable for facial comparison or establish that image quality, rather than installation and recording conditions, is their primary disadvantage; it links pixelation to camera resolution and subject distance. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • It may state two separate claims as one.

    Findingcitation

    Analogue CCTV data captured at an angle of incidence of 22° and height of 2.5 m performed worse overall, although it is unclear if this was because of camera position or poor image quality.

    “In contrast, our analogue CCTV data, captured at an angle of incidence of 22° and height of 2.5 m, performed worse overall, although it is unclear if this was because of camera position or poor image quality [48].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [48] Bacci, 2021, Performance of forensic facial comparison by morphological analysis across optimal and suboptimal CCTV settings, Sci. Justice source

  • It may state two separate claims as one.

    Findingcitation

    In 130 face sets (n = 1300 comparisons) recorded on a low-resolution suboptimal analogue CCTV system, MA had accuracies as low as 33.1%, false negative rates of 75.2%, and reliability of 37.8%.

    “The results were found to be much poorer, with accuracies as low as 33.1% with extremely high false negative rates (75.2%) and questionable reliability (37.8%) [48] (Figure 2).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [48] Bacci, 2021, Performance of forensic facial comparison by morphological analysis across optimal and suboptimal CCTV settings, Sci. Justice source

  • It may state two separate claims as one.

    Findingcitation

    In an analysis of 100 face sets (n = 1000 comparisons) with standard digital CCTV recordings as the target image, MA had lower accuracy (82.6%) and reliability (74.3%).

    “In the analysis of 100 face sets (n = 1000 comparisons) with standard digital CCTV recordings as the target image, a lower accuracy (82.6%) and reliability (74.3%) were noted [21] (Figure 2).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [21] Bacci, 2021, Validation of forensic facial comparison by morphological analysis in photographic and CCTV samples, Int. J. Legal Med. source

  • It may state two separate claims as one.

    Findingcitation

    Faces disguised by brimmed caps yielded an exceedingly low CCA (68.1%), yet not as low as the analogue CCTV comparisons (33.1%).

    “In contrast, faces disguised by brimmed caps yielded an exceedingly low CCA (68.1%) [49], yet not nearly as low as the analogue CCTV comparisons (33.1%) [48] (Figure 2).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [49] Bacci, 2021, Assessing the effect of facial disguises on forensic facial comparison by morphological analysis, J. Forensic Sci. source
    • cites [48] Bacci, 2021, Performance of forensic facial comparison by morphological analysis across optimal and suboptimal CCTV settings, Sci. Justice source

  • It may state two separate claims as one.

    Findingcitation

    In an analysis of 75 face sets (n = 750 comparisons), chance-corrected accuracy and reliability of MA on optimal photographs were 99.1% and 92.1%, respectively.

    “In an analysis of 75 sets of faces (each containing nine no match comparisons and one positive match comparison or 10 no match comparisons—compared to a target image, total n = 750 comparisons), the chance corrected accuracy and reliability were found to be almost perfect in optimal photographs (99.1% and 92.1%, respectively) [21].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [21] Bacci, 2021, Validation of forensic facial comparison by morphological analysis in photographic and CCTV samples, Int. J. Legal Med. source

  • It may state two separate claims as one.

    Findingcitation

    The performance of MA in faces disguised with sunglasses was markedly high (90.4%), surpassing performance under the same standard CCTV conditions without sunglasses (82.6%), but not better than photographic (99.1%) or eye-level CCTV data (97.3%).

    “Overall, the performance of MA in faces disguised with sunglasses was markedly high (90.4%) [49], in fact surpassing the performance of facial comparison under the same standard CCTV conditions without sunglasses (82.6%), but not better than the photographic (99.1%) [21] or eye-level CCTV data (97.3%) [48] (Figure 2).”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [49] Bacci, 2021, Assessing the effect of facial disguises on forensic facial comparison by morphological analysis, J. Forensic Sci. source
    • cites [21] Bacci, 2021, Validation of forensic facial comparison by morphological analysis in photographic and CCTV samples, Int. J. Legal Med. source
    • cites [48] Bacci, 2021, Performance of forensic facial comparison by morphological analysis across optimal and suboptimal CCTV settings, Sci. Justice source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Findingown data

    In the current work, color was found to not be of much importance.

    “although in the current work, color was found to not be of much importance.”
    Quote in the paperNot a substantial claim

  • It may state two separate claims as one.

    Findingown data

    Certain descriptors were found cumbersome to utilize; skin color and luminance were often ignored due to a mismatch, despite confirmation that two faces were indeed the same.

    “Certain descriptors were found cumbersome to utilize; for example, as mentioned above, skin color and luminance were often ignored due to a mismatch, despite confirmation that two faces were indeed the same.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Findingcitation

    Faces in the standard CCTV were composed of approximately double the number of pixels (41 × 52 pixels) and a much higher accuracy and reliability were obtained.

    “This is a suggested preliminary lower threshold as despite conditions being mostly similar between standard CCTV and analogue CCTV, in terms of angle of incidence and SCD, faces in the standard CCTV were composed of approximately double the number of pixels (41 × 52 pixels) and a much higher accuracy and reliability were obtained [21].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [21] Bacci, 2021, Validation of forensic facial comparison by morphological analysis in photographic and CCTV samples, Int. J. Legal Med. source

  • It may state two separate claims as one.

    Premisecitation

    The current recommended angle of incidence limit is 15°, as any steeper angle would result in significant loss of facial detail.

    “The current recommended angle of incidence limit is 15°, as any steeper angle would result in significant loss of facial detail [100].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [100] Ward, D. (2021, November 25). Testing Camera Height vs. Image Quality; Pennsylvania, USA. Available online: https://ipvm.com/reports/testing-camera-height.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • It may state two separate claims as one.

    Interpretationown data

    Image quality had the most notable effect on facial comparison performance (analogue CCTV recordings), while brimmed caps were the second-most limiting condition.

    “Image quality had the most notable effect on facial comparison performance (analogue CCTV recordings), while brimmed caps were found to be the second-most limiting condition.”
    Quote in the paperSubstantial claim

  • It may state two separate claims as one.

    Premisecitation

    Camera placement, lens focal length, and sensor size and pixel count dictate monitoring area, field of view, image quality, optical distortion, and noise.

    “Camera placement, particularly in relation to positioned angle and mounted height, dictates the monitoring area, while the camera lens, its focal length, as well as the sensor, specifically its size and number of pixels, dictate the field of view, image quality, level of optical distortion, and noise present [99].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [99] Damjanovski, V. (2014). CCTV from Light to Pixels, Elsevier. [3rd ed.].

  • It may state two separate claims as one.

    Premisecitation

    Typical surveillance camera mount heights are between 2.5 and 3 m on building exteriors and ceiling height for indoor surveillance.

    “Typical surveillance camera mount heights are between 2.5 and 3 m on building exteriors and ceiling height for indoor surveillance [99].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [99] Damjanovski, V. (2014). CCTV from Light to Pixels, Elsevier. [3rd ed.].

  • It may state two separate claims as one.

    Premisecitation

    Monitoring a large crowd outside a building requires 5% of picture height, while detecting a specific target requires 10%.

    “Monitoring a large crowd outside a building, for example, requires 5% of picture height, while detecting a specific target requires 10% [99].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [99] Damjanovski, V. (2014). CCTV from Light to Pixels, Elsevier. [3rd ed.].

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Premisecitation

    A potential target must occupy more than 400% of screen height for facial examination, and a minimum of 1 mm must be represented per pixel of the whole image.

    “A potential target must occupy more than 400% of screen height in order to conduct facial examination, and a minimum of 1 mm must be represented per pixel of the whole image (ISO62676 recommendations) [99].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [99] Damjanovski, V. (2014). CCTV from Light to Pixels, Elsevier. [3rd ed.].

  • It may state two separate claims as one.

    Premisecitation

    Standardized mounting heights translate to a steep angle of incidence, which reduces image quality and obscures relevant facial detail due to increased subject-to-camera distance and loss of useable resolution.

    “The problem with these standardized mounting heights is that they translate to a steep angle of incidence. This in turn reduces image quality and obscures relevant facial detail as a result of the increased SCD and subsequent loss of useable resolution [48,49].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [48] Bacci, 2021, Performance of forensic facial comparison by morphological analysis across optimal and suboptimal CCTV settings, Sci. Justice source
    • cites [49] Bacci, 2021, Assessing the effect of facial disguises on forensic facial comparison by morphological analysis, J. Forensic Sci. source

  • It may state two separate claims as one. The quote only partly states this claim.

    Interpretationreasoning

    Across all of the studies, chance-corrected accuracy (CCA) was the best and most representative measure of accuracy, compared to hit rate and balanced accuracy.

    “Across all of our studies, the best measure of accuracy was found to be the chance corrected accuracy (CCA) [21,48,49].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [21] Bacci, 2021, Validation of forensic facial comparison by morphological analysis in photographic and CCTV samples, Int. J. Legal Med. source
    • cites [48] Bacci, 2021, Performance of forensic facial comparison by morphological analysis across optimal and suboptimal CCTV settings, Sci. Justice source
    • cites [49] Bacci, 2021, Assessing the effect of facial disguises on forensic facial comparison by morphological analysis, J. Forensic Sci. source

  • It may state two separate claims as one.

    Premisecitation

    Eye-level mounted cameras at 1.8 m ground height provide an approximate 0-15° angle of incidence with the subject and provide the most optimal capture of facial detail even with natural head tilt.

    “Eye-level mounted cameras at 1.8 m ground height provide an approximate 0–15° angle of incidence with the subject and provide the most optimal capture of facial detail even with natural head tilt [48].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [48] Bacci, 2021, Performance of forensic facial comparison by morphological analysis across optimal and suboptimal CCTV settings, Sci. Justice source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Premisecitation

    The FISWG feature list is the most exhaustive list available, including over 130 facial component characteristics and over 290 characteristic descriptors.

    “The FISWG feature list is also the most exhaustive list available, including over 130 facial component characteristics and over 290 characteristic descriptors [51].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [51] Phillips, 2000, The FERET evaluation methodology for face-recognition algorithms, IEEE Trans. Pattern Anal. Mach. Intell. source

  • It may state two separate claims as one.

    Premisecitation

    Poor quality CCTV recordings and extracted images have been shown to affect face matching ability in both novice and experts and lead to high overall false positive rates.

    “Poor quality CCTV recordings and extracted images have been shown to affect face matching ability in both novice and experts and leads to high overall false positive rates [82].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [82] Lee, 2009, Matching unfamiliar faces from poor quality closed-circuit television (CCTV) footage: An evaluation of the effect of training on facial identification ability, Axis Online J. CAHId

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    The FISWG feature list includes 18 facial components, each with associated descriptors, as well as a nineteenth descriptive component for use with uncategorized features.

    “This list includes 18 facial components, each with associated descriptors, as well as a nineteenth descriptive component for use with uncategorized features [51,54].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [51] Phillips, 2000, The FERET evaluation methodology for face-recognition algorithms, IEEE Trans. Pattern Anal. Mach. Intell. source
    • cites [54] Gao, 2008, The CAS-PEAL large-scale chinese face database and baseline evaluations, IEEE Trans. Syst. Man Cybern. source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    Stephan identified that discrepancy in camera-to-face distances between photographs to be compared presented with varying degrees of perspective distortion of facial features.

    “Stephan [103] identified that discrepancy in camera-to-face, or skull, distances between photographs to be compared, presented with varying degrees of perspective distortion of facial features.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [103] Stephan, 2015, Perspective distortion in craniofacial superimposition: Logarithmic decay curves mapped mathematically and by practical experiment, Forensic Sci. Int. source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Premisecitation

    Vitek et al. found correct identifications decreased from 88 to 48% as HEVC encoding quality decreased (30 kbps-15 kpbs) and recommend 20 Kbps as a minimum threshold value.

    “Vitek et al. [125] found correct identifications decreased from 88 to 48% as HEVC encoding quality decreased (30 kbps–15 kpbs) and they recommend 20 Kbps as a minimum threshold value.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [125] Viték, S., Klíma, M., and Krasula, L. (2014, January 13–16). Video compression technique impact on efficiency of person identification in CCTV systems. Proceedings of the Proceedings—International Carnahan Conference on Security Technology, Rome, Italy. source

  • It may state two separate claims as one.

    Premisecitation

    Most modern IP cameras are better suited to handling high-contrast environments, but older analogue systems generally provide either over- or under-exposed coverage with limited middle grounds.

    “Most modern IP cameras are better suited to handling high-contrast environments, but older analogue systems generally provide either over- or under-exposed coverage with limited middle grounds [99].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [99] Damjanovski, V. (2014). CCTV from Light to Pixels, Elsevier. [3rd ed.].

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    At shorter distances, particularly below 6 m, the distortion was found to be more pronounced.

    “At shorter distances, particularly below 6 m, the distortion was found to be more pronounced [103].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [103] Stephan, 2015, Perspective distortion in craniofacial superimposition: Logarithmic decay curves mapped mathematically and by practical experiment, Forensic Sci. Int. source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    Vitek et al. recalculate Utochkin's recommendations to 35 pixels for a known individual and 83 pixels for an unknown individual.

    “Based on relative subject size on screen, Vitek et al. [125] recalculate Utochkin’s [126] recommendations to 35 pixels for a known individual and 83 pixels for an unknown individual.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [125] Viték, S., Klíma, M., and Krasula, L. (2014, January 13–16). Video compression technique impact on efficiency of person identification in CCTV systems. Proceedings of the Proceedings—International Carnahan Conference on Security Technology, Rome, Italy. source
    • cites [126] Utochkin, S. (2021, October 30). The Principles of CCTV Design in VideoCAD Interframe Compression. Available online: https://cctvcad.com/Files/the_principles_of_cctv_design_in_videocad_part5.pdf.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    Super-recognizers are being deployed.

    “and the deployment of super-recognizers [135,136]”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [135] Davis, J.P., and Tamonyte, D. (2017, January 6–8). Masters of disguise: Super-recognisers’ superior memory for concealed unfamiliar faces. Proceedings of the 2017 Seventh International Conference on Emerging Security Technologies (EST), Canterbury, UK. source

      Checked before — its earlier verdict is held back until a new check runs.

    • cites [136] Russell, 2009, Super-recognizers: People with extraordinary face recognition ability, Psychon. Bull. Rev. source

      Checked before — its earlier verdict is held back until a new check runs.

  • It may state two separate claims as one.

    Premisecitation

    For classification schemes to be effective, they need to be tailored to specific populations, which has only been considered by two studies to date.

    “In addition, for classification schemes to be effective, they need to be tailored to specific populations, which has only been considered by two studies to date [58,127].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [58] Kostinger, M., Wohlhart, P., Roth, P.M., and Bischof, H. (2011, January 6–13). Annotated Facial Landmarks in the Wild: A large-scale, real-world database for facial landmark localization. Proceedings of the 2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), Barcelona, Spain. source
    • cites [127] Stephan, 2017, Estimating the Skull-to-Camera Distance from Facial Photographs for Craniofacial Superimposition, J. Forensic Sci. source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Premisecitation

    European standards require a minimum of the top quarter of a subject on screen height and the face to represent a minimum of 1000 pixels per meter of screen height.

    “Considering the conservative European standards for facial image comparison [105], a minimum of the top quarter of a subject must be included on screen height and the face would need to represent a minimum of 1000 pixels per meter of screen height [99].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [105] British Security Industry (2021, October 30). Video Surveillance Systems for Use in Security Applications: System Requirements—General. Available online: https://www.bsia.co.uk/publications/video-surveillance/.
    • cites [99] Damjanovski, V. (2014). CCTV from Light to Pixels, Elsevier. [3rd ed.].

  • It may state two separate claims as one.

    Interpretationreasoning

    When conducting MA using the FISWG facial feature list, color is the first component, and therefore inaccurate image color data may lead to decreased accuracy in performance.

    “When conducting MA using the FISWG facial feature list [47], color is the first component, and therefore, inaccurate image color data may lead to a decreased accuracy in performance.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [47] Facial Identification Scientific Working Group (2021, November 25). Facial Image Comparison Feature List for Morphological Analysis. Available online: https://fiswg.org/FISWG_Morph_Analysis_Feature_List_v2.0_20180911.pdf.

  • Probably not a substantial claim.

    Premisecitation

    FISWG and ENFSI recommend trained human observer-based MA using the FISWG feature list and an ACE-V approach as the principal method of facial comparison for identification purposes.

    “As such, the authors strongly advise that trained human observer-based MA, using the FISWG feature list [47] and an ACE-V approach [27], should remain the principal method of facial comparison for identification purposes, as recommended by both the FISWG and ENFSI [27,114].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [27] Facial Identification Scientific Working Group (2021, November 25). Facial Comparison Overview and Methodology Guidelines. Available online: https://fiswg.org/fiswg_facial_comparison_overview_and_methodology_guidelines_V1.0_20191025.pdf.
    • cites [114] ENFSI (2018). Best Practice Manual for Facial Image Comparison, ENFSI. Available online: https://enfsi.eu/wp-content/uploads/2017/06/ENFSI-BPM-DI-01.pdf.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.

    Premisecitation

    FISWG advises against the use of photo-anthropometry for facial image comparison and recommends superimposition only to be utilized in conjunction with MA.

    “In addition, FISWG advises against the use of photo-anthropometry for facial image comparison and recommends superimposition only to be utilized in conjunction with MA [27].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [27] Facial Identification Scientific Working Group (2021, November 25). Facial Comparison Overview and Methodology Guidelines. Available online: https://fiswg.org/fiswg_facial_comparison_overview_and_methodology_guidelines_V1.0_20191025.pdf.

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    The FISWG has put forward a document describing guidelines for training and expertise requirements of FFC analysts and trainers.

    “The FISWG has put forward a document describing guidelines for training and expertise requirements of FFC analysts and trainers [133].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [133] FISWG (2020). Guide for Role-Based Training in Facial Comparison, FISWG. Available online: https://www.fiswg.org/fiswg_guide_for_role-based_training_in_facial_comparison_v1.0_20200717.pdf.

  • It may state two separate claims as one.

    Premisenone given

    Compression employed in CCTV systems is lossy and, once performed during recording, cannot be removed or reversed.

    “Compression employed in CCTV systems is lossy and, once performed during recording, cannot be removed or reversed.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim.

    Study design and conductcitation

    The conditions in the studies were limited to the major questions broadly investigated by each study.

    “In doing so, although realistic, the conditions were limited to the major questions broadly investigated by each study [21,48,49].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [21] Bacci, 2021, Validation of forensic facial comparison by morphological analysis in photographic and CCTV samples, Int. J. Legal Med. source
    • cites [48] Bacci, 2021, Performance of forensic facial comparison by morphological analysis across optimal and suboptimal CCTV settings, Sci. Justice source
    • cites [49] Bacci, 2021, Assessing the effect of facial disguises on forensic facial comparison by morphological analysis, J. Forensic Sci. source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    Both the FISWG and the ENFSI recommend MA as the best practice for forensic facial identification.

    “Both the FISWG and the ENFSI recommend MA as the best practice for forensic facial identification [27,125].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [27] Facial Identification Scientific Working Group (2021, November 25). Facial Comparison Overview and Methodology Guidelines. Available online: https://fiswg.org/fiswg_facial_comparison_overview_and_methodology_guidelines_V1.0_20191025.pdf.
    • cites [125] Viték, S., Klíma, M., and Krasula, L. (2014, January 13–16). Video compression technique impact on efficiency of person identification in CCTV systems. Proceedings of the Proceedings—International Carnahan Conference on Security Technology, Rome, Italy. source

  • Probably not a substantial claim. The paper’s conclusions may not depend on it.

    Premisecitation

    The training course outline proposed by members of the research group inadvertently follows most of the proposed FISWG guidelines.

    “Recently, members of our research group proposed an outline for a training course with a three-tiered approach offered to the police force [17]; it inadvertently follows most of the proposed guidelines from FISWG.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [17] Steyn, 2018, Forensic facial comparison in South Africa: State of the science, Forensic Sci. Int. source

  • It may state two separate claims as one.

    Premisecitation

    CCTV systems in general do not accurately capture color information from a scene and have been deemed mostly unreliable in a forensic context.

    “CCTV systems in general do not accurately capture color information from a scene [119] and have been deemed mostly unreliable in a forensic context [91,120].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [119] Hunt, R.W.G. (1995). The Reproduction of Colour, Fountain Press. [5th ed.]. source
    • cites [91] Smith, 2008, Colour analysis and verification of CCTV images under different lighting conditions, Image Qual. Syst. Perform. V source
    • cites [120] MacDonald, 2007, Fatal Flaws: Uncertainty in the Interpretation of Colour in CCTV Images, Ann. Br. Mach. Vis. Assoc.
42 of 42

  • The paper relies on this without support.

    The FISWG feature list is generic enough to be applicable across sex and population groups.

    “In principle, the FISWG feature list should be generic enough to make it applicable across sex and population groups, but facial variations may potentially lead to variations in accuracies and reliability based on the biases and abilities of the observers.”

  • The paper relies on this without support.

    Little data exist describing the types and quantity of data loss incurred in CCTV systems globally and how this impacts surveillance and criminal investigations.

    “Little data exist describing the types and quantity of data loss incurred in CCTV systems globally and how this impacts surveillance and criminal investigations.”

  • The paper relies on this without support.

    There is a significant lack of research on IR CCTV recording in the context of facial identification and testing.

    “There is a significant lack of research on IR CCTV recording in the context of facial identification and testing, and further validation of MA on a sample of IR surveillance data of comparable quality and conditions to the standard CCTV camera should be conducted.”

  • The paper relies on this without support.

    The main justification behind typical surveillance camera mounting heights is that it lowers the risk of cameras being vandalized, stolen, or obstructed.

    “The main justification behind these mounting heights is that it lowers the risk of cameras being vandalized, stolen, or obstructed.”

  • The paper relies on this without support.

    Trained police members have found success in their roles as facial examiners.

    “the trained police members have found success in their roles as facial examiners.”

  • The paper relies on this without support.

    Multiple factors influence the quality of an image for facial identification.

    “there are multiple factors that influence the quality of an image for facial identification.”

  • The paper relies on this without support.

    Image quality relies on more than just equipment resolution capacity; it involves lighting conditions, angle of incidence, SCD, distortions, color, visibility of features, and more.

    “Image quality relies on more than just equipment resolution capacity; it involves lighting conditions, angle of incidence, SCD, distortions, color, visibility of features, and more.”

  • The paper relies on this without support.

    MA is the most universally applicable and reliable method for facial comparison.

    “continued research in forensic facial comparison by MA is crucial as the most universally applicable and reliable method.”
8 of 8

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Self-check results appear here once an independent check has confirmed them.

Cited works that were retracted

Works this paper cites that were later retracted, from the retraction notices publishers deposit. Major when the notice came out before this paper was published.

  • Cites a retracted work: 10.6028/nist.sp.800-88r1

    Info

    The cited work was retracted on 2025-09-26 (notice 10.6028/nist.sp.800-88r1), not before this paper was published (2021-12-03).

    Cited work 10.6028/nist.sp.800-88r1 · retracted 2025-09-26 · notice 10.6028/nist.sp.800-88r1 · this paper published 2021-12-03

1 of 1

Citations

Cites

1 retracted

Flagged first, then by citations

Research works printed but not linked: 8

  1. Press Association (The Guardian, 2018). Welsh Police Wrongly Identify Thousands as Potential Criminals, The Guardian.
  2. Valentine, T., and Davis, J.P. (2015). Forensic Facial Identification, John Wiley & Sons. · DOI
  3. Speckeis, 2011, Can ACE-V be validated?, J. Forensic Identif., 61, 201
  4. Bacci, N., Davimes, J., Steyn, M., and Briers, N. (2021, October 30). Wits Face Database. Available online: https://hdl.handle.net/10539/29924.
  5. Ngan, M., Grother, P., and Hanaoka, K. (2020). Ongoing Face Recognition Vendor Test (FRVT) Part 6B: Face Recognition Accuracy with Face Masks Using Post-COVID-19 Algorithms. · DOI · Paper page
  6. Lee, 2009, Matching unfamiliar faces from poor quality closed-circuit television (CCTV) footage: An evaluation of the effect of training on facial identification ability, Axis Online J. CAHId, 1, 19
  7. Damjanovski, V. (2014). CCTV from Light to Pixels, Elsevier. [3rd ed.].
  8. MacDonald, 2007, Fatal Flaws: Uncertainty in the Interpretation of Colour in CCTV Images, Ann. Br. Mach. Vis. Assoc., 7, 1

Cited by

Sorted by citations

Related

Metadata

Authors
Bacci, N., Davimes, J. G., Steyn, M., & Briers, N.
Affiliations
University of the Witwatersrand (ZA)
Corresponding author
Bacci, Nicholas
ORCID
Venue
Biology
Publisher
MDPI AG
Published
2021-12-03
Volume
10
Issue
12
Pages
1269
ISSN
2079-7737 (ISSN-L)
DOI
10.3390/biology10121269
PMC
PMC8698381
OpenAlex
W4200507084
Semantic Scholar
d686e79b86b858b7a22369e0bfaacf0e0a870eb7
Cited by
26
References
108
Citations by year
2026: 4 · 2025: 5 · 2024: 11 · 2023: 2 · 2022: 4
FWCI
0.25
Type
journal article
Language
en
Open access
Licence
Locations
  • Biology (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Research Foundation (11858, CSUR160425163022)
  • JJJ Smieszek Fellowship, School of Anatomical Sciences (2017)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Face recognition and analysis
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Face Recognition and Perception
  • Social Sciences › Arts and Humanities › Archeology (arts and humanities) › Forensic Anthropology and Bioarchaeology Studies
Keywords
face identification, CCTV footage, hit rate, image quality, face recognition

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-09, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-09 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review