Pith. sign in

REVIEW 2 major objections 4 minor 31 references

Which device you use matters little for continuous face authentication; how your face appears to the camera matters a lot, and everyday phone use with occluded or off-screen faces causes most false lockouts.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-05 00:19 UTC pith:EKCQGYQU

load-bearing objection Solid comparative study; the cross-platform null result holds up, but the abstract's visibility claim overreaches the descriptive evidence. the 2 major comments →

arxiv 2608.00763 v1 pith:EKCQGYQU submitted 2026-08-01 cs.HC

Continuous Face Authentication on Mobile and Desktop Platforms: A Comparative Study

Miriam Palmetshofer , Rainhard Dieter Findling , Marc Kurz This is my paper
classification cs.HC
keywords continuous authenticationface recognitionbiometric authenticationmobile authenticationdesktop authenticationInsightFacetemporal trust decayin-the-wild evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to establish that a face-based continuous authentication system behaves nearly identically on phones and laptops, and that the conditions of use — not the platform — determine how often it fails. Using an InsightFace-based verifier with a decaying trust score, the authors test 25 people in controlled sessions (normal use, repeated head turns, low light) on both an iPhone and a MacBook, and then in 121 recordings of unscripted everyday phone use. Controlled results show device differences are small and mostly vanish under multiple-testing correction, while scenario differences are large and significant, with low light as the hardest condition. Everyday use is harder still: at the standard threshold, 22.3% of genuine-user sessions end in a false lockout, and 21 of 27 such failures coincide with the face occluded, cut off, or outside the camera frame. If correct, the upshot is that deployment effort should go into robustness to real-world face visibility and user-configurable thresholds, not platform-specific tuning.

Core claim

Central claim: continuous face authentication is governed by how the face reaches the camera, not by the device. Across 25 participants, an InsightFace-based verifier with exponential trust decay produced statistically comparable trust trajectories on an iPhone 12 Pro and a MacBook Pro; the only significant device difference was small, traced to a posture mismatch between upright mobile enrollment and lowered typing. Usage conditions mattered far more: low light cut trust most and caused the only false lockouts; head turns recovered on return to frontal view. Everyday phone use was hardest: false rejections rose from 0% to 22.3% of recordings at threshold 0.5, with 21 of 27 failures involvin

What carries the argument

The pipeline combines InsightFace's compact buffalo_sc model — a face-detection and recognition pipeline pairing SCRFD-500MF detection with MobileFaceNet recognition producing 512-dimensional L2-normalized embeddings — with an enrollment set of 50 reference embeddings per participant (five head directions, two rotation directions, five frames each). Frame-level similarity is the 90th percentile of cosine similarities to enrollment, and a temporal trust rule turns per-frame scores into continuous decisions: detected faces update trust with weight e^(−Δt/45), undetected faces decay it with e^(−Δt/60), and trust below τ = 0.5 locks the device. The trust-decay equations are what carry the argume

Load-bearing premise

The paper's main causal story — that occlusion and off-screen faces cause false rejections — rests on the authors' own post-hoc labeling of 27 failing recordings, with no second coder and no formal protocol, and the paper itself concedes this analysis does not isolate causal effects.

What would settle it

Score every frame of the 121 in-the-wild recordings with an automated visibility measure (face-detector confidence, face bounding-box size, distance from frame edges, and detected occlusion), then test whether visibility predicts the 27 false rejections while controlling for motion blur, head angle, and phone model. If visibility does not predict lockouts, the occlusion explanation fails; if it does, the count of 21 in 27 is confirmed quantitatively. A second annotator independently coding the same 27 recordings for visibility would directly test the reliability of that count.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Device choice is a minor factor: the same approach can be expected to behave similarly on a phone and a laptop under matched conditions, so platform-specific tuning is unlikely to pay off as much as condition-specific robustness.
  • Benchmarking in controlled conditions overstates real-world performance: at the standard threshold, one in five everyday mobile sessions would end in a false lockout, mostly when the face is not fully visible.
  • The lock threshold is a genuine usability/security dial: lowering it from 0.5 to 0.4 cuts false rejections from 22.3% to 8.3% while keeping the imposter lockout rate at 0%, but slows unauthorized-user lockout from 9s to 20s.
  • Low light is the hardest controlled condition — the only one that produced false lockouts — while head turns are survivable because users periodically return to a frontal view, letting trust recover.
  • Unauthorized users were locked out in every composed scenario (FAR 0%), meaning the approach's practical weakness is interrupting legitimate users, not admitting imposters.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 21-of-27 visibility finding is descriptive, not causal; an automated per-frame visibility score (detector confidence, face size, distance to frame edge, occlusion masks) could convert it into a quantitative predictor of lockouts and test whether visibility beats motion blur and phone model as an explanation.
  • The enrollment posture mismatch suggests a cheap, testable fix: enrolling while holding the phone in the natural typing position (rather than upright in front of the face) may close most of the remaining mobile-desktop trust gap without changing the model.
  • Because FAR stays at 0% while FRR drops sharply at τ = 0.4, an adaptive threshold — lower in privacy-sensitive apps, higher in contexts with frequent occlusion like walking or eating — could capture most of the usability gain without a global security tradeoff.
  • The study's in-the-wild leg used only phones; extending the naturalistic protocol to laptops would test whether the device-type conclusion holds outside controlled conditions, where desktop lighting and posture are different.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper implements a continuous face-authentication pipeline combining InsightFace embeddings with an exponential temporal trust-decay mechanism and evaluates it in two studies. A controlled study with 25 participants compares mobile (iPhone 12 Pro) and desktop (MacBook Pro) performance under baseline, repeated-head-turn, and low-light conditions using composed genuine/impostor videos. An in-the-wild study additionally records 121 naturalistic mobile sessions on participants' own smartphones. The paper's main claims are that device type alone has little effect, that usage conditions substantially affect performance, and that everyday mobile use is more challenging, with reduced face visibility being a main contributor to false rejections. The implementation is publicly released.

Significance. The controlled cross-platform comparison is a genuine and non-obvious contribution: face-based continuous authentication is usually evaluated on a single platform, and the paired three-scenario design with composed impostor segments and multiple-testing correction provides useful evidence that lighting and head posture can matter more than platform choice. The release of the implementation is a further strength. However, the headline inference about reduced face visibility as the main driver of false rejections is not supported by the quantitative analysis as reported, and the in-the-wild comparison conflates usage condition with device hardware. If the authors address these evidential gaps, the paper would be a solid empirical addition to the continuous-authentication literature.

major comments (2)
  1. [§5.2 / Abstract] The claim that reduced face visibility is a main contributor to false rejections is not established. Section 5.2 reports reduced visibility in 21 of the 27 in-the-wild recordings that false-rejected at τ=0.5, but it gives no corresponding rate for the 94 recordings that did not false-reject. Without this base rate, 21/27 is not evidence: if reduced visibility is equally common in accepted recordings, the count is diagnostically weak. The passage itself states that the analysis 'does not isolate causal effects' yet the abstract and conclusion elevate it to 'main contributor.' The visibility coding also appears to be single-rater and post-hoc, with no protocol or inter-rater reliability, and it may be confounded with motion blur, head pose, and the different trust-update branch when no face is detected (Eq. 3). Please add a base-rate comparison for accepted vs false-rejected recordings, a
  2. [§4.3 / Table 3] The in-the-wild versus controlled-mobile comparison is confounded by device. The controlled mobile condition used a single iPhone 12 Pro, whereas the in-the-wild recordings used participants' own smartphones (18 iPhones, 6 Android devices). The lower GT, higher FRR, and shorter ULT in the in-the-wild condition could therefore reflect differences in camera hardware, image processing, or recording conditions specific to those phones rather than naturalistic usage per se. The abstract's statement that 'everyday mobile device use is in general more challenging' is accordingly not cleanly supported. Please report results separately by device type, or at least on the iPhone-only subset, and state the distribution of recordings across devices.
minor comments (4)
  1. [Typos/formatting] There are small typos and formatting artifacts: 'T able 1' and 'T able 3' in the text, and 'F AR' in the Results section heading. These should be fixed in the final version.
  2. [§3.2] The functions f_wei and f_dec are used in Eq. (3) but are only defined in the surrounding text. Please define them directly with the equations for readability.
  3. [§4.4] The in-the-wild evaluation pairs each genuine-user recording with 18 randomly selected recordings from other participants using a fixed seed. It would be helpful to state whether the random selection is stratified by device type or scenario and to report the resulting number of pairs per device, since this affects the FAR and ULT estimates.
  4. [Figures 4 and 5] The trust-score traces for hundreds or thousands of composed videos are overlaid as single colored lines, which makes the figures difficult to read beyond the mean and P90 markers. Consider using quantile bands or a heatmap-style density display.

Circularity Check

0 steps flagged

No circular derivation: the evaluation is a forward measurement over fixed configuration parameters, the GT–ULT relation is an explicit definitional consequence rather than a fitted prediction, and the visibility limitation is a causal-inference concern, not circularity.

full rationale

The paper's central claims are empirical measurements over recorded videos processed with a fixed InsightFace pipeline and a fixed temporal-decay rule (Eqs. 1–4). The parameters k_weight=45s, k_decay=60s, τ=0.5, p=0.90 are stated configuration values (Table 1); they are not fitted to the evaluation outcomes. FRR, FAR, GT, and ULT are reported consequences of running that fixed configuration, not predictions derived from fitted parameters. The temporal-decay logic is adopted from prior work [2] as a design choice, not as an external uniqueness theorem or as a source of the empirical claims. The GT–ULT inverse relation described in §5.1 is a definitional property of the trust-update rule: a lower starting trust leaves less distance to the threshold, so the unauthorized-user crossing time is shorter. The paper presents this as an explanation of the observed table values, not as an independent empirical discovery, so it is not a self-definitional prediction. The in-the-wild visibility analysis (§5.2) is a descriptive, post-hoc coding of 27 false-rejected recordings; the paper itself cautions that it 'does not isolate causal effects,' and the abstract's 'main contributor' wording overstates the support. That is a missing base-rate / causal-inference limitation — and the paper flags it — but it is not circular reasoning: the visibility coding is not an input to the model that later reappears as the output. No self-citation is load-bearing: the authors' prior work (e.g., [10,11,12,15]) appears only as contextual related work, and the adapted temporal-decay reference [2] is not by the present authors. Under the required standard — exhibiting a specific reduction of a claimed result to its own inputs — no circular step can be identified.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central claims rest on the pretrained InsightFace model, the hand-chosen temporal decay parameters, the enrollment assumptions, and the simulation design. No new physical or algorithmic entities are postulated beyond the standard trust-score construct inherited from prior work.

free parameters (6)
  • kweight = 45 s
    Time constant for combining new similarity evidence; empirically determined to balance responsiveness and tolerance (Section 3.3).
  • kdecay = 60 s
    Time constant for trust decay when no face is detected; empirically determined (Section 3.3).
  • tau = 0.5 (and 0.4 in threshold comparison)
    Lock threshold; chosen by authors; the paper also evaluates tau=0.4 to illustrate usability tradeoff (Table 3).
  • p = 0.90
    Percentile used to aggregate enrollment similarities; chosen by authors (Section 3.1, Table 1).
  • X0 = 1.0
    Initial trust after successful non-continuous authentication; assumed value (Section 3.3).
  • authentication rate = 1 Hz
    Frame processing rate used in evaluation; affects delta-t and trust dynamics (Section 3.3).
axioms (5)
  • domain assumption InsightFace (SCRFD-500MF + MobileFaceNet) embeddings are reliable for the face verification task across poses and lighting
    The paper relies on a pretrained model rather than training or validating it on this dataset (Section 3.1).
  • domain assumption The largest detected face in a frame is the primary device user
    Stated in Section 3.1; if a secondary face were larger, the approach would verify the wrong identity.
  • domain assumption Enrollment under good lighting with rotating head covers the pose space well enough to be reused across all scenarios
    Enrollment is performed once and reused for baseline, head-turn, low-light, and in-the-wild evaluations (Section 4.1).
  • domain assumption Composed videos with black-frame transitions and non-adversarial imposter segments simulate realistic device takeover
    The unauthorized-user simulation uses other participants who do not avoid detection and a 3s black transition (Section 4.4); the paper itself lists exclusion of adversarial behavior as a limitation.
  • domain assumption The hand-coded categorization of reduced facial visibility is accurate and not confounded by motion blur or angle
    The descriptive analysis in Section 5.2 labels 21 of 27 false-rejected recordings as having reduced visibility; the paper states this analysis does not isolate causal effects.

pith-pipeline@v1.3.0-alltime-deepseek · 3656 in / 5343 out tokens · 153610 ms · 2026-08-05T00:19:15.388473+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Continuous Face Authentication on Mobile and Desktop Platforms: A Comparative Study." pith.science (2026). https://pith.science/paper/EKCQGYQU

@misc{pith2026260800763,
  author       = {Pith},
  title        = {Pith review of: Continuous Face Authentication on Mobile and Desktop Platforms: A Comparative Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EKCQGYQU}},
  note         = {Machine review of arXiv:2608.00763}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Personal devices hold sensitive data and provide access to sensitive services. Conventional personal device authentication verifies users' identity only at the moment access is granted. An unlocked device may be accessed by an unauthorized person if the user stops using the device without locking it, or if another person takes over. Continuous authentication addresses this gap. This paper investigates how device type and usage conditions influence continuous mobile face authentication with an InsightFace-based approach with temporal trust decay. We evaluate the approach with mobile and desktop recordings with different head directions and lighting conditions. We also evaluate recordings from everyday mobile device use without predefined tasks. The results show that device type alone has little impact, while different usage conditions do have impact on the authentication performance. Results also show that everyday mobile device use is in general more challenging for continuous face authentication, where reduced face visibility, including occlusions and faces outside the camera viewport, is a main contributor to false rejections.

Figures

Figures reproduced from arXiv: 2608.00763 by Marc Kurz, Miriam Palmetshofer, Rainhard Dieter Findling.

Figure 1
Figure 1. Figure 1: Overview of the continuous face authentication pipeline, from image capture and InsightFace-based verification to temporal trust updating and the authentication decision. in the buffalo model family at 16 MB. This variant uses SCRFD-500MF [14] for face detection and MobileFaceNet [5] for recognition. The recognition model generates 512-dimensional L2-normalized embedding vectors, and when multiple faces ar… view at source ↗
Figure 2
Figure 2. Figure 2: Controlled study examples on desktop and mobile. Rows show the baseline, repeated head turns, and low-light scenarios. explained its features. Participants were asked to complete at least four 5-minute recordings over several days. To keep participants in the app during recording, the app provided links to games, digital newspapers, and other websites that opened in an in-app browser. Participants freely s… view at source ↗
Figure 3
Figure 3. Figure 3: In-the-wild study examples with varying lighting, head positions, and occlu￾sions. 3. An unauthorized-user segment showing another study participant using the device without attempting to avoid detection. The approach continues to use the genuine user’s enrollment data, and the segment has the same duration as the genuine-user segment. For the controlled study, each genuine-user recording is paired with re… view at source ↗
Figure 4
Figure 4. Figure 4: Controlled-study trust scores by scenario and device. Colored lines show com￾posed videos, black lines show means, and the dashed line marks τ = 0.5. Diamonds mark mean and P90 ULT [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: In-the-wild trust scores. Colored lines show composed videos, the black line shows the mean, and dashed lines mark τ = 0.5 and τ = 0.4. Diamonds mark mean and P90 ULT. 6 Conclusion This paper examined how continuous face authentication is affected by desktop and mobile platform choice and by different usage conditions. The controlled study shows only small differences between mobile and desktop devices. Ov… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 30 canonical work pages

  1. [1]

    IEEE IoT J

    Abuhamad, M., Abuhmed, T., Mohaisen, D., Nyang, D.: AUToSen: Deep-learning- based implicit continuous authentication using smartphone sensors. IEEE IoT J. 7(6), 5008–5020 (2020)

    work page 2020

  2. [2]

    Master’s thesis, University of Applied Sciences Upper Austria (Dec 2017)

    Altenhofer, C.: Continuous Mobile Face Authentication. Master’s thesis, University of Applied Sciences Upper Austria (Dec 2017)

    work page 2017

  3. [3]

    In: Proc

    Ashwinkumar, J.S., Kumaran, H.S., Sivakarthikeyan, U., Rajesh, K.P.B.V., La- vanya, R.: Deep learning based approach for facilitating online proctoring using transfer learning. In: Proc. IEEE ICCCSP 2021. pp. 306–312 (2021)

    work page 2021

  4. [4]

    In: CVPRW (2019)

    Bazarevsky, V., Kartynnik, Y., Vakunov, A., Raveendran, K., Grundmann, M.: BlazeFace: Sub-millisecond neural face detection on mobile GPUs. In: CVPRW (2019)

    work page 2019

  5. [5]

    In: CCBR

    Chen, S., Liu, Y., Gao, X., Han, Z.: MobileFaceNets: Efficient CNNs for accurate real-time face verification on mobile devices. In: CCBR. vol. 10996 (2018)

    work page 2018

  6. [6]

    In: Proc

    Crouse, D., Han, H., Chandra, D., Barbello, B., Jain, A.K.: Continuous authen- tication of mobile user: Fusion of face image and inertial measurement unit data. In: Proc. IEEE ICB 2015. pp. 135–142 (2015)

    work page 2015

  7. [7]

    In: Proc

    Damer,N.,Maul,F.,Busch,C.:Multi-biometriccontinuousauthentication:Atrust model for an asynchronous system. In: Proc. FUSION 2016. pp. 2192–2199 (2016)

    work page 2016

  8. [8]

    IEEE Trans

    Deng, J., Guo, J., Yang, J., Xue, N., Kotsia, I., Zafeiriou, S.: ArcFace: Additive angular margin loss for deep face recognition. IEEE Trans. Pattern Anal. Mach. Intell.44(10), 5962–5979 (2022)

    work page 2022

  9. [9]

    Pattern Recognit

    Fenu, G., Marras, M., Boratto, L.: A multi-biometric system for continuous student authentication in e-learning platforms. Pattern Recognit. Lett.113, 83–92 (2018)

    work page 2018

  10. [10]

    IEEE TMC17(11), 2578–2590 (2018)

    Findling, R.D., Hölzl, M., Mayrhofer, R.: Mobile match-on-card authentication using offline-simplified models with gait and face biometrics. IEEE TMC17(11), 2578–2590 (2018)

    work page 2018

  11. [11]

    In: Proc

    Findling, R.D., Wenny, F., Holzmann, C., Mayrhofer, R.: Range face segmentation: Face detection and segmentation for authentication in mobile device range images. In: Proc. MoMM 2013. pp. 260–269 (2013)

    work page 2013

  12. [12]

    Findling, R.D., Mayrhofer, R.: Towards pan shot face unlock: Using biometric face information from different perspectives to unlock mobile devices. Int. J. Pervasive Comput. Commun.9(3), 190–208 (2013)

    work page 2013

  13. [13]

    In: Proc

    Ganidisastra, A.H.S., Bandung, Y.: An incremental training on deep learning face recognition for m-learning online exam proctoring. In: Proc. IEEE APWiMob 2021. pp. 213–219 (2021) Continuous Face Authentication on Mobile and Desktop Platforms 15

    work page 2021

  14. [14]

    In: ICLR (2022)

    Guo, J., Deng, J., Lattas, A., Zafeiriou, S.: Sample and computation redistribution for efficient face detection. In: ICLR (2022)

    work page 2022

  15. [15]

    Hintze, D., Füller, M., Scholz, S., Findling, R.D., Muaaz, M., Kapfer, P., Koch, E., Mayrhofer, R.: CORMORANT: Ubiquitous risk-aware multi-modal biometric authentication across mobile devices. Proc. ACM IMWUT3(3), 85:1–85:23 (2019)

    work page 2019

  16. [16]

    Hussain Al-Naji, F., Zagrouba, R.: CAB-IoT: Continuous authentication architec- ture based on blockchain for internet of things. J. King Saud Univ. Comput. Inf. Sci.34, 2497–2514 (2022)

    work page 2022

  17. [17]

    IEEE TMC22(3), 1551–1565 (2023)

    Keykhaie, S., Pierre, S.: Lightweight and secure face-based active authentication for mobile users. IEEE TMC22(3), 1551–1565 (2023)

    work page 2023

  18. [18]

    Image Vis

    Khoury, E., El Shafey, L., McCool, C., Günther, M., Marcel, S.: Bi-modal biometric authentication on mobile phones in challenging conditions. Image Vis. Comput. 32(12), 1147–1160 (2014)

    work page 2014

  19. [19]

    IEEE IoT J.11(9), 15957– 15968 (2024)

    Li, Y., Sun, X., Yang, Z., Huang, H.: SNNAuth: Sensor-based continuous authen- tication on smartphones using spiking neural networks. IEEE IoT J.11(9), 15957– 15968 (2024)

    work page 2024

  20. [20]

    In: Proc

    Mahbub, U., Sarkar, S., Patel, V.M., Chellappa, R.: Active user authentication for smartphones: A challenge data set and benchmark results. In: Proc. IEEE BTAS

    work page

  21. [21]

    IEEE TIFS5(4), 771–780 (2010)

    Niinuma, K., Park, U., Jain, A.K.: Soft biometric traits for continuous user au- thentication. IEEE TIFS5(4), 771–780 (2010)

    work page 2010

  22. [22]

    In: Proc

    Oza, P., Patel, V.M.: Active authentication using an autoencoder regularized CNN- based one-class classifier. In: Proc. IEEE FG 2019. pp. 1–8 (2019)

    work page 2019

  23. [23]

    IEEE SPM 33(4), 49–61 (2016)

    Patel, V.M., Chellappa, R., Chandra, D., Barbello, B.: Continuous user authenti- cation on mobile devices: Recent progress and remaining challenges. IEEE SPM 33(4), 49–61 (2016)

    work page 2016

  24. [24]

    IEEE TIFS14(5), 1240–1250 (2019)

    Perera, P., Patel, V.M.: Face-based multiple user active authentication on mobile devices. IEEE TIFS14(5), 1240–1250 (2019)

    work page 2019

  25. [25]

    Image Vis

    Samangouei, P., Patel, V.M., Chellappa, R.: Facial attributes for active authenti- cation on mobile devices. Image Vis. Comput.58, 181–192 (2017)

    work page 2017

  26. [26]

    In: Proc

    Schroff, F., Kalenichenko, D., Philbin, J.: FaceNet: A unified embedding for face recognition and clustering. In: Proc. IEEE CVPR 2015. pp. 815–823 (2015)

    work page 2015

  27. [27]

    In: CVPR Workshops

    Segundo, M.P., Sarkar, S., Goldgof, D., Silva, L., Bellon, O.: Continuous 3D face authentication using RGB-D cameras. In: CVPR Workshops. pp. 64–69 (2013)

    work page 2013

  28. [28]

    Multimed Tools Appl83(28), 71451–71493 (2024)

    Shende, S.W., Tembhurne, J.V., Ansari, N.A.: Deep learning based authentica- tion schemes for smart devices in different modalities: Progress, challenges, perfor- mance, datasets and future directions. Multimed Tools Appl83(28), 71451–71493 (2024)

    work page 2024

  29. [29]

    In: Proc

    Smith-Creasey, M., Albalooshi, F.A., Rajarajan, M.: Context awareness for improved continuous face authentication on mobile devices. In: Proc. IEEE DASC/PiCom/DataCom/CyberSciTech 2018. pp. 644–652 (2018)

    work page 2018

  30. [30]

    In: Proc

    Viola, P., Jones, M.: Rapid object detection using a boosted cascade of simple features. In: Proc. IEEE CVPR 2001. vol. 1 (2001)

    work page 2001

  31. [31]

    IEEE Signal Process

    Zhang, K., Zhang, Z., Li, Z., Qiao, Y.: Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Process. Lett.23(10), 1499–1503 (2016)

    2016