Cooperative Face Liveness Detection from Optical Flow

Anton Konushin; Artem Sokolov; Mikhail Nikitin

REVIEW 4 major objections 2 minor 25 references

Cooperative Face Liveness Detection from Optical Flow

T0 review · 4 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The authors claim that a cooperative protocol—having a user slowly move their face toward the camera—lets a neural optical-flow and RGB classifier read 3D facial volume cues, improving liveness discrimination against printed photos, screen

desk verdict The abstract describes a face-liveness method; the full text is an unrelated math paper, so the submission cannot be evaluated for the claimed CV work. read the letter →

arxiv 2508.10786 v1 pith:WHSNSHSA submitted 2025-08-14 cs.CV

Artem Sokolov , Mikhail Nikitin , Anton Konushin This is my paper · ORCID

classification cs.CV

keywords facelivenessdetectionpresentationattackopticalflowcooperativeprotocolfacialvolumevideoreplayneuralclassifierspatio-temporalfeatures

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper is trying to establish that face liveness can be checked by asking the user to move their face slowly toward the camera and analyzing the resulting optical flow. The claim is that this movement makes the 3D shape of a real face visible in the motion pattern, while flat photos, screens, rigid masks, and video replays produce flow fields that a neural classifier can separate from genuine faces. If true, this cooperative protocol would give video-based liveness systems a simple, hardware-free way to reduce presentation-attack success compared with passive methods. The central assumption is that a replay of an approaching face does not fool the classifier.

What carries the argument

The controlled approaching-face protocol is the load-bearing interaction: it converts liveness detection from observing whatever motion happens to watching a specific, repeatable motion whose optical flow should encode 3D facial volume. The neural optical-flow estimator extracts dense frame-to-frame motion, and the resulting flow maps, concatenated with RGB frames, are fed to a neural classifier that learns spatial-temporal patterns separating live faces from presentation attacks. The protocol is what makes 'volume from motion' available as a discriminative signal.

What would settle it

Record a high-frame-rate video of a real face slowly approaching a camera, display that video on a phone or tablet, and move the display toward a second camera at the same physical speed; run the trained classifier. If the replay clips are accepted at a rate close to genuine clips, the central discrimination claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that depth, or facial volume, can be extracted from motion alone under a controlled interaction: participants are instructed to keep their face frontal and move it steadily toward the camera. In that setting, the optical flow field is not just an overall expansion; because nearby parts of a 3D surface move differently from distant parts, the flow carries parallax information tied to facial geometry. The authors state that neural optical-flow estimation plus a classifier operating on both flow and RGB frames significantly improves discrimination between genuine faces and printed-photo, screen, mask, and video-replay attacks, and that this beats passive methods tha

Load-bearing premise

The load-bearing premise is that the optical flow generated by a live 3D face moving toward the camera is different enough from the flow of a flat photo, a screen, a rigid mask, and especially a video replay of an approaching face that a learned classifier can tell them apart; the paper supplies no argument or experiment isolating the replay case.

Editorial extensions

If this is right

A liveness check could be administered as a short, natural gesture—'slowly bring your face closer'—requiring no special sensor beyond an RGB camera.
If the flow-based volume cue is as discriminative as claimed, printed photos and rigid masks should be rejected at lower false-acceptance rates than in passive systems, because their motion fields lack the parallax of a genuine 3D face.
The hardest attack class becomes a video replay of someone moving toward the camera; the paper's claim that the classifier still separates such replays is the decisive point to test.
Combining flow and RGB features lets the model exploit both motion geometry and appearance, which should generalize better than appearance-only detectors to unseen attack types.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension is to treat the protocol as a challenge-response: the system could randomize the required speed or direction of approach, forcing an attacker to replay a matching motion; whether such randomization actually increases security is not tested in the paper.
Because the discriminator relies on learned optical flow, its behavior on very close-range faces, motion blur, or atypical skin and lighting is likely to depend on the flow estimator's training distribution; a hand-computed divergence or parallax statistic could offer an interpretable diagnostic or a cheaper baseline.
The same 'volume from controlled motion' cue might transfer to other verification tasks, such as object or document liveness, where a user moves an item toward the camera and depth is inferred from the resulting flow.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague, here is my read. The abstract of 2508.10786 promises a cooperative video-based face liveness detection system built on a slow approach-to-camera protocol, neural optical flow, and a classifier separating live faces from prints, screens, masks, and replays. The full text, however, is a mathematics paper by different authors about Ricci shrinkers. So the document as submitted does not contain the method it advertises. I cannot referee the CV claim because there is no architecture, no training detail, no dataset, no evaluation, and no analysis tying the abstract to the body. This is a submission-level failure, not a matter of interpretation.

What is actually new, and worth saying, is the abstract's idea. The cooperative protocol—ask the user to move their face toward the camera, then use optical flow to infer facial volume—is not a standard liveness cue in the visibility of the literature to me, and combining it with a learned neural classifier to cover print, screen, mask, and replay attacks is a plausible design. If a real paper were written around that idea, it would be worth a serious referee.

But on this submission, the soft spots are structural. There is nothing to check. The abstract's claim of 'significantly improving discrimination' is completely unsupported. The only specific concern I would flag for the eventual real paper is the replay attack: a video of a face moving toward the camera produces nearly the same optical expansion as a live face. The abstract does not acknowledge this, let alone provide an argument or experiment showing the learned classifier can separate them. That would be the load-bearing test.

The Ricci shrinker paper itself may be perfectly good mathematics, but it is not this abstract's paper, and the authors are not even the same. My recommendation for the editor: desk-reject or return to authors to resubmit with the correct manuscript. The CV idea has some promise, but this version is not reviewable. I would not cite this document for anything, and I would not bring it to our reading group as-is.

Referee Report

4 major / 2 minor

Summary. The abstract announces a cooperative video-based face liveness detection method built around a 'controlled approaching face protocol': the user is asked to move their frontal-oriented face toward the camera, neural optical flow is estimated, and a classifier separates live faces from printed photos, screen displays, masks, and video replays. The substantive text supplied under 'Full Text', however, is a mathematics paper on the geometric structure of ends of Ricci shrinkers by Bertellotti and Buzano (arXiv:2508.10790), with no mention of face liveness, optical flow, presentation attacks, datasets, or experiments. The central claim—that the proposed method significantly improves discrimination between genuine and attacked faces—is therefore completely unsupported by the manuscript as submitted.

Significance. If substantiated, the proposed cooperative protocol could be a genuinely useful contribution to face anti-spoofing: exploiting the expansion pattern of a live face approaching the camera is a plausible source of 3D/volume cues and could complement passive methods. However, the submitted manuscript provides no substantiation whatsoever. There is no model architecture, no training procedure, no protocol specification, no dataset, no baseline comparison, and no evaluation metric. The only rigorous content is an unrelated differential-geometry article. The manuscript therefore has no current scientific value for the computer-vision audience and cannot be assessed for correctness, novelty, or reproducibility.

major comments (4)

[Full text, §1 (Eq. 1.1)] The body of the manuscript is not the described liveness-detection paper. The abstract promises a neural optical-flow estimator and a liveness classifier; the full text is 'Geometric Structure of Ends of Ricci Shrinkers' and its principal equation is Ric_g + Hess f = 1/2 g. None of the claimed components—optical flow, classifier, attack types, interaction protocol—appear anywhere in the full text. This is a load-bearing failure: the central claim cannot be checked because the method itself is absent.
[Abstract (claim of improved discrimination)] The abstract asserts 'significantly improving discrimination between genuine faces and various presentation attacks' and states the method is more reliable than passive methods. The manuscript contains no dataset, no experimental setup, no baselines, and no results (no APCER, BPCER, EER, ROC, or any other metric). For a claim of empirical superiority, evaluation is mandatory. As submitted, the assertion is entirely unsupported.
[Abstract ('controlled approaching face protocol')] The core innovation is described only in the abstract as a two-sentence interaction scenario. The full text contains no definition of the protocol: no instructions for the user motion, no camera configuration, no video duration, no frame-rate or temporal-window choices, and no description of how flow fields and RGB frames are combined. The method is therefore neither reproducible nor testable.
[Full text, throughout (unspecified models)] No neural network architecture, loss function, training procedure, or inference details are given for either the optical flow estimator or the final classifier. Without these, it is impossible to verify whether the flow features provide independent information or simply memorize training-set idiosyncrasies. The absence of any equations or pseudocode describing the liveness pipeline makes the manuscript impossible to evaluate even as a theoretical proposal.

minor comments (2)

[Full text, references] The reference list consists entirely of mathematics papers on Ricci flow and Ricci shrinkers. There is no citation of prior work on face anti-spoofing, optical-flow-based motion analysis, or presentation-attack detection, which would be expected even in a short conference-style submission.
[Full text, formatting] The displayed equation in §1 is garbled in the provided text (e.g., 'Ricg +∇2f = 1 2 g'), and several labels contain placeholder characters. This is secondary to the substantive mismatch, but it further obstructs reading.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the claimed CV method is not present in the full text, so there is no derivation chain whose outputs reduce to its inputs.

full rationale

The abstract at the top describes a cooperative face-liveness-detection pipeline: users move their face toward the camera, neural optical flow is estimated, and a classifier separates live faces from print/screen/mask/replay attacks. The full text supplied is an unrelated mathematics article, 'Geometric Structure of Ends of Ricci Shrinkers' by Bertellotti and Buzano (arXiv:2508.10790v1), containing only Ricci-shrinker geometry equations such as Ric_g + ∇²f = ½g. There is no optical-flow estimator, classifier architecture, interaction protocol, dataset, training procedure, evaluation metric, or held-out experiment anywhere in the full text. Circularity requires an exhibited reduction—e.g., a fitted parameter renamed as a prediction, a self-citation used to forbid alternatives, or an output that equals its input by definition. None of these structures can be located because the claimed derivation chain is absent. The mismatch is a severe omission/verifiability problem and makes the abstract's claim unsupported, but unsupported is not the same as circular. Therefore the circularity score is 0.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The abstract introduces a cooperative protocol that assumes live 3D faces and planar attacks produce distinguishable optical flow. This domain assumption is the main unproven premise; learned parameters are undisclosed, and no independent evidence is provided.

free parameters (1)

unspecified neural network weights (optical flow estimator and liveness classifier)
The method is a learned system, so its performance depends on weights fitted to training data; the abstract provides no architectures or trained values, so these are unaccounted free parameters.

assumptions (2)

domain assumption The cooperative 'slow approach' protocol reliably induces distinct optical flow for live faces versus planar attacks
The abstract asserts this as the core innovation; it assumes users follow the protocol and that planar attacks cannot reproduce the 3D motion cue.
domain assumption Optical flow estimated from RGB frames preserves facial volume information
The method depends on neural optical flow estimation retaining the 3D structure cue despite potential estimation errors and low-light or low-texture regions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cooperative Face Liveness Detection from Optical Flow." pith.science (2026). https://pith.science/paper/WHSNSHSA

@misc{pith2026250810786,
  author       = {Pith},
  title        = {Pith review of: Cooperative Face Liveness Detection from Optical Flow},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WHSNSHSA}},
  note         = {Machine review of arXiv:2508.10786}
}

read the original abstract

In this work, we proposed a novel cooperative video-based face liveness detection method based on a new user interaction scenario where participants are instructed to slowly move their frontal-oriented face closer to the camera. This controlled approaching face protocol, combined with optical flow analysis, represents the core innovation of our approach. By designing a system where users follow this specific movement pattern, we enable robust extraction of facial volume information through neural optical flow estimation, significantly improving discrimination between genuine faces and various presentation attacks (including printed photos, screen displays, masks, and video replays). Our method processes both the predicted optical flows and RGB frames through a neural classifier, effectively leveraging spatial-temporal features for more reliable liveness detection compared to passive methods.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 22 canonical work pages

[1]

Introduction The main goal of this article is to initiate the study of the geometric structure of the ends of a gradient shrinking Ricci soliton ( M, g, f) �� global curvature assumptions. In order to do so, we study pointed limits of such solitons along sequences of points going to infinity along an integral curve of ∇f , or equivalently blow-up lim...

arXiv 2025
[9]

P. Y. Chan, Z. Ma, and Y. Zhang. �� . Calc. Var. �� :179 (2024)

work page 2024
[10]

Chow, et al

B. Chow, et al. �� . Mathematical Surveys and Monographs. Vol. 163, AMS (2010)

work page 2010
[11]

Enders, R

J. Enders, R. M¨ uller, and P. Topping.�� . Comm. Anal. Geom. �� :5 (2011), 905–922

work page 2011
[12]

Gigli, A

N. Gigli, A. Mondino and G. Savar´ e. �� . Proc. London Math. Soc. �:11 (2015), 1071–1129

work page 2015
[13]

Hallgren

M. Hallgren. �� . Adv. Math. �� :108940 (2023)

work page 2023
[14]

R. S. Hamilton. � �� . Amer. J. Math. �� :3 (1995), 545–572

work page 1995
[15]

Haslhofer and R

R. Haslhofer and R. M¨ uller. � �� . Geom. Funct. Anal. �� (2011), 1091–1116

work page 2011

Show all 25 references

[16]

Haslhofer and R

R. Haslhofer and R. M¨ uller. � �� . Proc. Amer. Math. Soc. �� :10 (2015), 4433–4437

2015
[17]

N. V. Krylov. �� . Graduate Studies in Mathematics. Vol. 12, AMS (1996)

1996
[18]

H. Li, Y. Li, and B. Wang. �� . J. Funct. Anal. �� :108955 (2021)

2021
[19]

Li and B

Y. Li and B. Wang. �� . Calc. Var. �� :194 (2020)

2020
[20]

Li and B

Y. Li and B. Wang. �� . Acta Math. Sci. �� (2024), 1639–1695

2024
[21]

Li and B

Y. Li and B. Wang. �� . Preprint (2023), ArXiv:2301.09784v2

2023 arXiv
[22]

Mantegazza and R

C. Mantegazza and R. M¨ uller. �� . J. Reine Angew. Math. �� (2015), 173–199

2015
[23]

Munteanu, F

O. Munteanu, F. Schulze, and J. Wang. �� . J. Reine Angew. Math. �� (2021),185–217

2021
[24]

Munteanu and J

O. Munteanu and J. Wang. �� . Commun. Anal. Geom. �� :3 (2011), 451–486

2011
[25]

Munteanu and J

O. Munteanu and J. Wang. �� . Adv. Math. �� (2014), 269–305

2014
[26]

Munteanu and J

O. Munteanu and J. Wang. �� . Compos. Math. �� (2015), 2273–2300

2015
[27]

Munteanu and J

O. Munteanu and J. Wang. �� . Ann. Sci. ´Ec. Norm. Sup´ er. �� (2019), 891–925

2019
[28]

Munteanu and J

O. Munteanu and J. Wang. �� . J. Geom. Anal. �� :303 (2022)

2022
[29]

A. Naber. �� . J. Reine Angew. Math. �� (2010), 125–153

2010
[30]

Perelman

G. Perelman. �� Preprint (2002), ArXiv:math/0211159

2002 arXiv
[31]

Pigola, M

S. Pigola, M. Rimoldi, and A. Setti. �� . Math. Z. �� (2011), 777–790

2011
[32]

Z. H. Zhang. �� . Proc. Amer. Math. Soc. �� (2009), 2755–2759. �� ...

2009

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Introduction The main goal of this article is to initiate the study of the geometric structure of the ends of a gradient shrinking Ricci soliton ( M, g, f) �� global curvature assumptions. In order to do so, we study pointed limits of such solitons along sequences of points going to infinity along an integral curve of ∇f , or equivalently blow-up lim...

arXiv 2025

[2] [9]

P. Y. Chan, Z. Ma, and Y. Zhang. �� . Calc. Var. �� :179 (2024)

work page 2024

[3] [10]

Chow, et al

B. Chow, et al. �� . Mathematical Surveys and Monographs. Vol. 163, AMS (2010)

work page 2010

[4] [11]

Enders, R

J. Enders, R. M¨ uller, and P. Topping.�� . Comm. Anal. Geom. �� :5 (2011), 905–922

work page 2011

[5] [12]

Gigli, A

N. Gigli, A. Mondino and G. Savar´ e. �� . Proc. London Math. Soc. �:11 (2015), 1071–1129

work page 2015

[6] [13]

Hallgren

M. Hallgren. �� . Adv. Math. �� :108940 (2023)

work page 2023

[7] [14]

R. S. Hamilton. � �� . Amer. J. Math. �� :3 (1995), 545–572

work page 1995

[8] [15]

Haslhofer and R

R. Haslhofer and R. M¨ uller. � �� . Geom. Funct. Anal. �� (2011), 1091–1116

work page 2011