Pith. sign in

REVIEW 3 major objections 7 minor 40 references

Discovering Geometric Biases in 3D Face Reconstruction: A Curvature-Aware Spectral Framework for Fairness Evaluation

T0 review · 3 major / 7 minor · reviewed 2026-07-09 · glm-5.2

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

Pith's one-line read Curvature metric exposes age bias in 3D face models

desk verdict Curvature-based metric for 3D face reconstruction fairness: solid core idea, but the perceptual validation doesn't cover where the strongest bias findings are. read the letter →

arxiv 2607.07486 v1 pith:2SYB2HMS submitted 2026-07-08 cs.CV

Veronika Shilova , Emmanuel Malherbe , Giovanni Palma , Panagiotis-Alexandros Bokaris , Laurent Risser , Jean-Michel Loubes This is my paper
classification cs.CV
keywords 3DfacereconstructioncurvatureerrorLaplace-Beltramioperatordemographicbiasmorphablemodelfairnessspectralclusteringmeshprocessing
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

Standard 3D face reconstruction evaluation relies on Euclidean point-to-point distances, which measure whether the overall shell of a reconstructed face is in the right place but miss whether the surface detail — ridges, folds, creases, local undulations — is faithfully captured. This paper proposes replacing that metric with a curvature-based error measure derived from the Laplace-Beltrami operator, a mathematical object that captures how a surface bends at every point. The authors call this the Curvature Reconstruction Error (CRE). In a user study with 19 annotators judging 1,111 pairs of nose reconstructions, CRE matched human perceptual judgments 73.63% of the time, while the standard Euclidean metric performed near chance (42–54%). Having established that CRE sees what humans see, the authors use it to probe two widely deployed 3D Morphable Model bases — BFM and FLAME — for demographic bias. They find that reconstruction error in the cheek region correlates with subject age at Pearson r = 0.77 (BFM) and r = 0.82 (FLAME), meaning older faces are systematically smoothed out because the linear model cannot represent the high-frequency geometric detail of aging skin. This bias persists through five downstream 2D-to-3D reconstruction algorithms, indicating the problem originates in the 3DMM basis itself, not in how it is fitted. Preliminary clustering analysis also finds that African and Asian faces concentrate in specific failure modes where the nose is unnaturally creased and rounded, though the authors flag this as preliminary given the small sample sizes.

What carries the argument

The Laplace-Beltrami operator estimated on triangle meshes via local quadratic surface fitting; its spectral decomposition into manifold harmonics (eigenfunctions of the generalized eigenvalue problem Lφ = λMφ); projection of per-vertex curvature error maps into this spectral basis; and K-means clustering of the resulting spectral coefficients to discover recurring failure modes linked to demographic attributes.

What would settle it

If CRE were computed on a larger, demographically balanced 3D face dataset and the Pearson correlations between age and cheek-region CRE dropped below 0.3 (or lost significance after multiple-comparison correction), the central bias-discovery claim would be substantially weakened.

Watch

Extended reading notes

Core claim

The central object is the Curvature Reconstruction Error (CRE), defined as the L1-normed mass-weighted difference in mean curvature H between a ground-truth mesh and its reconstruction, where H is estimated via a local quadratic regression of the Laplace-Beltrami operator on R-ring mesh neighborhoods. The paper's core claim is twofold: first, that CRE is a substantially better proxy for human perceptual judgment of reconstruction fidelity than Euclidean NMSE (73.63% vs. 42–54% pairwise accuracy), and second, that when CRE is applied across demographic groups, it reveals that 3DMM bases carry systematic age-related representational bias (Pearson r up to 0.82 between age and cheek-region CRE),

Load-bearing premise

The demographic bias findings rest on the REALY benchmark of 100 subjects with imbalanced subgroup sizes (very few non-Caucasian samples), and the 48 Pearson correlations in Table 2 are reported without confidence intervals or multiple-comparison correction, making the age-bias correlations suggestive but not yet statistically airtight.

Editorial extensions

If this is right

  • If CRE is adopted as a standard evaluation metric, 3D face reconstruction benchmarks that currently report only Euclidean errors would need to be re-evaluated, potentially changing the ranking of state-of-the-art methods.
  • The finding that age bias is inherent to the linear 3DMM subspace — not introduced by fitting algorithms — suggests that debiasing efforts must target the model basis (e.g., nonlinear or richer subspaces) rather than the fitting pipeline.
  • Regulatory frameworks like the EU AI Act require declared accuracy metrics for high-risk AI; if curvature-based metrics reveal biases that Euclidean metrics miss, regulators may need to specify which geometric fidelity metric is acceptable for compliance.
  • The spectral clustering of error maps into demographic-specific failure modes provides a template for bias auditing in other template-based 3D reconstruction domains beyond faces (e.g., body shape, medical implants).

Reading between the lines

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

  • The 73.63% perceptual accuracy, while far above chance, also means CRE disagrees with human judgment roughly 26% of the time — suggesting there are perceptually salient reconstruction qualities that even curvature does not fully capture (e.g., texture, specular highlights, or higher-order differential properties).
  • If the age bias stems from the linear PCA basis being unable to represent high-frequency surface detail, then nonlinear generative models (e.g., neural implicit surfaces, diffusion-based shape priors) should be directly testable with CRE to quantify whether they reduce this bias — a natural next experiment the paper does not run.
  • The spectral clustering approach (Algorithm 1) requires choosing K and the number of clusters without a stability analysis; an independent reader would want to know whether the ethnicity-linked failure modes survive perturbation of these parameters before drawing qualitative conclusions about systematic bias against specific populations.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 7 minor

Summary. This paper introduces a curvature-aware evaluation framework for 3D face reconstruction based on the Laplace-Beltrami operator. The Curvature Reconstruction Error (CRE) metric is validated through a user study showing 73.63% pairwise accuracy in matching human perceptual judgments, substantially outperforming Euclidean NMSE metrics. Using CRE, the authors uncover systematic age-related biases in 3DMM bases (BFM, FLAME) and reconstruction algorithms, with Pearson correlations up to 0.82 between age and cheek-region CRE. Preliminary evidence for ethnicity-linked failure modes is presented via spectral clustering of error maps. The framework is method-agnostic and includes code release.

Significance. The paper addresses an important and underexplored gap: demographic bias quantification in 3D face reconstruction. The CRE metric is derived from mesh geometry (not fitted to demographic labels), avoiding circularity. The user study provides external perceptual validation. Code and annotation data are publicly released, supporting reproducibility. The spectral clustering approach for discovering unknown failure modes is a novel and potentially impactful contribution. The finding that biases are inherent to the linear 3DMM subspace rather than specific fitting algorithms is well-motivated by the experimental design (projecting ground-truth 3D scans directly into the 3DMM basis).

major comments (3)
  1. §4.2, Table 1: The perceptual validation of CRE (73.63% pairwise accuracy) is conducted exclusively on nose-region reconstructions using BFM. However, the strongest bias findings in Table 2 are in the cheek region with FLAME (r=0.82). The paper does not provide evidence that CRE aligns with human perceptual judgment for cheek reconstructions or for FLAME. If curvature variations in cheeks are less perceptually salient than in the nose, the age-bias correlation may capture geometric discrepancies that humans do not perceive as reconstruction failures. This generalization gap should be addressed—either by extending the user study to at least one additional region/model combination, or by explicitly acknowledging this as a limitation and tempering the perceptual-claim language for the bias findings.
  2. §4.3, Table 2: The Pearson correlations between age and CRE are reported without confidence intervals, p-values, or multiple-comparison corrections. Table 2 reports 48 correlations (8 models × 4 regions × 1 'All' column, plus gender-stratified columns). Without correction for multiple testing, the strong correlations (e.g., r=0.82) may be partially inflated by selection across many region/demographic/model combinations. Please report 95% confidence intervals and apply a multiple-comparison correction (e.g., Bonferroni or Benjamini-Hochberg) to the set of tests in Table 2.
  3. §4.2, Table 1: The best hyperparameters (R=3, K=128) appear to have been selected on the same 1,111 nose/BFM pairs used to report the 73.63% accuracy figure, with 18 configurations tested. If the hyperparameters were selected to maximize pairwise accuracy on this set, the reported accuracy is optimistically biased. Please clarify whether any form of train/test split or cross-validation was used for hyperparameter selection, or acknowledge this as a limitation.
minor comments (7)
  1. §3.2, Eq. (9): The choice of L1 norm is stated but not motivated. The ablation (Table 1, lower portion) shows L1 outperforms L2 and Huber, but no explanation is given for why L1 is superior in this context.
  2. §3.3, Algorithm 1: The number of clusters for K-means is not specified. The text mentions '9 (resp. 7) unique failure modes' for BFM (resp. FLAME) but does not explain how K was chosen. A stability analysis or external criterion would strengthen this choice.
  3. §4.3, Fig. 6: The ethnicity analysis is based on very few non-Caucasian samples in REALY (Fig. 2). The paper acknowledges this limitation but still draws qualitative conclusions about 'systematic failure modes for African and Asian populations.' Please temper the language to reflect the preliminary nature of the finding.
  4. §3.2: The matrix tilde-Delta is described as nonsymmetric, while the cotangent-weight Laplacian L in §3.3 (Eq. 11) is symmetric. The relationship between these two operators and why different formulations are used for curvature estimation vs. spectral decomposition should be clarified.
  5. Table 1: The 'Pairwise-ROC-AUC' column is reported but not defined in the text. Please add a definition.
  6. §4.1: The optimization in Eq. (14) uses vertex-to-vertex mapping, while CRE in Eq. (8) uses point-to-surface mapping. The inconsistency between the fitting objective and the evaluation metric should be noted.
  7. Fig. 2: The demographic statistics are shown as a bar chart but exact sample counts per ethnicity category are not provided in tabular form. Given the small sample sizes, a table with exact counts would be helpful.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The three major comments are all valid concerns that we will address in the revised manuscript. Below we respond to each point.

read point-by-point responses

  1. Referee: §4.2, Table 1: Perceptual validation of CRE conducted only on nose/BFM, while strongest bias findings are in cheek/FLAME (r=0.82). Generalization gap for perceptual alignment across regions and models.

    Authors: The referee raises a valid concern. Our user study was conducted exclusively on nose-region reconstructions using BFM, yet the strongest age-bias correlations are found in the cheek region with FLAME. We acknowledge this generalization gap. In the revised manuscript, we will take two steps. First, we will explicitly add a limitation paragraph in Section 5 noting that the perceptual validation covers only the nose/BFM combination and that perceptual alignment for cheek reconstructions or FLAME has not been directly verified. Second, we will temper the perceptual-claim language in Sections 4.3 and 5: rather than stating that CRE aligns with human perception across all regions and models, we will scope the claim to the validated setting (nose/BFM) and frame the cheek/FLAME bias findings as geometrically motivated but not yet perceptually validated. We note that the CRE metric is derived from differential geometry (mean curvature via the Laplace-Beltrami operator) and is not fitted to any demographic or perceptual labels, so the risk of circularity is low; however, the referee is correct that curvature variations in the cheek may differ in perceptual salience from those in the nose, and we will state this explicitly. We will also add extending the user study to at least one additional region/model combination as a priority for future work. revision: yes

  2. Referee: §4.3, Table 2: Pearson correlations reported without confidence intervals, p-values, or multiple-comparison corrections across 48 tests.

    Authors: This is a correct and important observation. In the revised manuscript, we will report 95% confidence intervals for all Pearson correlations in Table 2 and apply a Benjamini-Hochberg correction (FDR at 0.05) across the full family of tests in the table. We will also report the corresponding adjusted p-values. We expect the strongest correlations (e.g., r=0.82 for FLAME cheek, r=0.88 for BFM cheek/male) to remain significant given their magnitude and the sample size of 100 subjects, but we will transparently flag any correlations that do not survive correction. We chose Benjamini-Hochberg over Bonferroni because the tests are not fully independent (they share the same subjects across regions and models), and FDR control is more appropriate under such dependence. We will add a footnote explaining the correction procedure and its assumptions. revision: yes

  3. Referee: §4.2, Table 1: Hyperparameters (R=3, K=128) appear selected on the same 1,111 pairs used to report 73.63% accuracy, with 18 configurations tested. Potential optimistic bias without train/test split.

    Authors: The referee is correct to flag this. We did not use a train/test split or cross-validation for hyperparameter selection; the 18 configurations in the ablation table were all evaluated on the same 1,111 labeled pairs, and the best-performing configuration (R=3, K=128) was selected based on this set. We will acknowledge this as a limitation in the revised manuscript. That said, we note two mitigating factors that we will also include in the discussion. First, the ablation results show that CRE is robust across hyperparameters: the accuracy ranges from 66.52% to 73.63%, and several configurations achieve accuracy within 1-2 percentage points of the best (e.g., R=4, K=128 at 73.00%; R=2, K=128 at 72.46%; R=1, K=128 at 71.38%). This suggests the reported accuracy is not an artifact of a fragile hyperparameter setting. Second, all CRE configurations substantially outperform the NMSE baselines (which range from 42% to 54%), so the qualitative conclusion—that curvature-based metrics align better with human perception than Euclidean metrics—holds regardless of hyperparameter choice. Nevertheless, we agree that the specific figure of 73.63% is optimistically biased, and we will state this clearly, noting it should be interpreted as an upper bound on perceptual accuracy. In future work, we plan to conduct a follow-up user study with a held-out test set to obtain an unbiased estimate. revision: yes

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity found; CRE is defined from mesh geometry, validated against external human judgments, and bias findings are correlations with external demographic labels.

full rationale

The paper's derivation chain is largely self-contained and does not reduce to its inputs by construction. (1) The CRE metric (Eqs. 8–9) is defined purely from differential geometry: it computes the difference in mean curvature H between ground truth and reconstructed meshes, where H is estimated via the Laplace-Beltrami operator (Eq. 7). This definition is independent of demographic labels or perceptual judgments. (2) The 3DMM fitting procedure (Eq. 14) optimizes vertex-to-vertex NMSE, not CRE, so the fitting objective and the evaluation metric are distinct—no self-definitional circularity. (3) The perceptual validation (Section 4.2) uses an external user study with 19 annotators providing pairwise comparisons, which is independent ground truth, not a quantity the metric was designed to fit. (4) The bias findings (Section 4.3) are Pearson correlations between CRE and demographic metadata (age, gender, ethnicity) from the REALY benchmark—external labels not used in CRE's definition. (5) The spectral clustering (Algorithm 1) clusters error maps without reference to demographic labels; the demographic distributions of clusters are computed post hoc. The one minor concern is that hyperparameters R=3, K=128 were selected on the same 1,111 nose/BFM pairs used to report the 73.63% accuracy, but the ablation (Table 1) shows all CRE configurations outperform NMSE (worst CRE: 61.57% vs. best NMSE: 53.92%), so the qualitative finding is not forced by the selection. Reference [34] is a self-citation by overlapping authors but is cited only for general ML bias context in the introduction, not as a load-bearing premise. The generalization gap between nose/BFM validation and cheek/FLAME bias findings is a correctness risk, not a circularity issue.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities, particles, or forces. The CRE metric and spectral clustering pipeline are constructed from established mathematical operators (Laplace-Beltrami, mean curvature, K-means). The free parameters (R, K, norm, cluster count) are standard hyperparameters selected by ablation, not theoretical constants. The axioms are domain assumptions about the validity of curvature as a perceptual proxy and the representativeness of the benchmark, both of which are empirically tested but not theoretically derived.

free parameters (4)
  • R (ring neighborhood radius) = 3 (best), tested 1-4
    Controls the size of the local neighborhood for Laplace-Beltrami estimation. Selected via ablation on user-study data (Table 1). R=3 gives best pairwise accuracy.
  • K (number of eigenfunctions) = 128 (best), tested 128-1024
    Dimension of spectral space for error map projection and clustering. Selected via ablation (Table 1). K=128 gives best pairwise accuracy.
  • Number of spectral clusters = 9 (BFM nose), 7 (FLAME nose)
    Number of K-means clusters for bias discovery. Not justified by stability analysis or external criterion; appears chosen manually.
  • Norm in CRE scalar metric = L1
    Selected over L2 and Huber via ablation (Table 1). L1 gives best pairwise accuracy.
assumptions (4)
  • domain assumption Mean curvature H is a meaningful proxy for human perception of facial reconstruction fidelity.
    The entire framework rests on curvature capturing perceptually salient features. Validated empirically via user study (73.63% accuracy) but not derived from perceptual theory. Invoked in Section 3.2.
  • domain assumption The REALY benchmark (100 subjects) is sufficiently representative to detect demographic biases.
    All bias findings depend on this sample. The paper acknowledges limited ethnicity sample sizes but still draws qualitative conclusions. Invoked in Section 4.1.
  • domain assumption Rigid ICP alignment is sufficient pre-processing for curvature comparison between meshes.
    Curvature is rotation/translation invariant in theory, but discrete estimation on meshes depends on vertex sampling and triangulation quality. The paper assumes ICP alignment plus vertex correspondence is adequate. Invoked in Section 3.1.
  • domain assumption Linear PCA basis of 3DMMs is the primary bottleneck for demographic fairness, not the fitting algorithm.
    The paper claims biases are 'inherent to their linear prediction subspaces' based on observing similar bias patterns across 3DMM bases and 2D reconstruction algorithms. This is supported by Table 2 but not formally proven. Invoked in Section 5.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Discovering Geometric Biases in 3D Face Reconstruction: A Curvature-Aware Spectral Framework for Fairness Evaluation." pith.science (2026). https://pith.science/paper/2SYB2HMS

@misc{pith2026260707486,
  author       = {Pith},
  title        = {Pith review of: Discovering Geometric Biases in 3D Face Reconstruction: A Curvature-Aware Spectral Framework for Fairness Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2SYB2HMS}},
  note         = {Machine review of arXiv:2607.07486}
}
read the original abstract

3D Morphable Models (3DMMs) remain the standard parametric shape priors for many state-of-the-art 3D face reconstruction algorithms. However, as these models are derived from a finite number of 3D face samples, they inherit the morphological biases of their training data, potentially limiting their generalizability across diverse global populations. In this paper, we propose a novel framework to analyze 3DMM reconstructions through the lens of surface curvature, with the objective to discover, quantify and visualize biases. While standard evaluation metrics often rely on Euclidean distances, our reconstruction error captures subtle surface nuances such as local topology or undulations. To do so, we leverage the Laplace-Beltrami Operator (LBO) to generate high-resolution curvature error maps, providing a localized and geometrically meaningful visualization of discrepancies between ground truth faces and reconstructed meshes. We derive from it an error metric that we validated through a user study, observing a significantly higher correlation to human perception compared to traditional methods. Furthermore, we conduct extensive experiments across several 3DMM bases and fitting algorithms, uncovering systematic age-related biases and providing preliminary evidence of biases associated with gender and ethnicity. Our findings highlight the necessity of adopting curvature-aware evaluation protocols to ensure demographic fairness and geometric precision in future 3D face reconstruction research.

Figures

Figures reproduced from arXiv: 2607.07486 by the authors.

Figure 1
Figure 1. Overview of our curvature-aware evaluation framework for bias discovery. Given a ground truth scan and its corresponding 3D reconstruction, we perform an anatomical segmentation of the reconstructed mesh. We compute Curvature-Based Error Maps using the Laplace-Beltrami Operator to isolate localized geometric discrepancies that Euclidean metrics often overlook. By performing spectral clustering on these error maps, w… view at source ↗
Figure 2
Figure 2. Overview of REALY benchmark: (a) REALY region masks, (b) REALY De￾mographic statistics To do so, given a 3DMM, we consider its template µ and aim at assigning a region label to each of its vertex vi , that can be later used for any face expressed in this base. Given a ground truth scan M = (V, F) in the REALY dataset, we first align it with the mesh of µ via the ICP algorithm [2, 3]. Then, for every vertex vi in the… view at source ↗
Figure 3
Figure 3. Example of three pairs with their metrics and the user choice. on which we computed the ratio of agreement. We retained as labeled pairs the 1, 111 pairs for which this ratio exceeds 75%. Given a metric E, we computed its accuracy as the probability that its choice matches the human majority consensus \text {Pairwise-Accuracy}(\mathcal {E}) = \frac {1}{N_{\text {pairs}}}\sum _{j=1}^{N_{\text {pairs}}}\mathds {1}_{\b… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Example of age bias on a senior person from REALY benchmark. These results reveal a significant representation bias: as age increases, the 3DMM basis and reconstruction models fails to capture high-frequency geometric details, effectively "smoothing out" the structural…
Figure 5
Figure 5. Figure 5: Error cluster centroids found for BFM (top row) and FLAME (bottom row) [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Distribution of failure modes @nose for each ethnicity in REALY benchmark [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Example of African and Asian subjects affected by a failure mode 1 of BFM [29] [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Centroids for error clusters identified for BFM [29] and FLAME [23]. In (·) we put the P C with maximal mean absolute value for the error cluster and its sign. cipal component – the dimension that primarily governs the cluster’s morphology – defined as the component wi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages

  1. [1]

    Regulation (eu) 2024/1689 of the european parliament and of the council of 13 june 2024 laying down harmonised rules on artificial intelligence and amending regulations (ec) no 300/2008, (eu) no 167/2013, (eu) no 168/2013, (eu) 2018/858, (eu) 2018/1139 and (eu) 2019/2144 and directives 2014/90/eu, (eu) 2016/797 and (eu) 2020/1828 (artificial intelligence ...

    work page 2024

  2. [2]

    In: 2007 IEEE conference on computer vision and pattern recognition

    Amberg, B., Romdhani, S., Vetter, T.: Optimal step nonrigid icp algorithms for surface registration. In: 2007 IEEE conference on computer vision and pattern recognition. pp. 1–8. IEEE (2007)

    work page 2007

  3. [3]

    In: Sensor fusion IV: control paradigms and data structures

    Besl, P.J., McKay, N.D.: Method for registration of 3-d shapes. In: Sensor fusion IV: control paradigms and data structures. vol. 1611, pp. 586–606. Spie (1992)

    work page 1992

  4. [4]

    Efficient error models for fault-tolerant architectures and the Pauli twirling approximation

    Besse, P., del Barrio, E., Gordaliza, P., Loubes, J.M., Risser, L.: A survey of bias in machine learning through the prism of statistical parity. The American Statistician 76(2), 188–198 (2022).https://doi.org/10.1080/00031305.2021.1952897

    work page Pith review arXiv doi:10.1080/00031305.2021.1952897 2022

  5. [5]

    In: Pro- ceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques

    Blanz, V., Vetter, T.: A morphable model for the synthesis of 3d faces. In: Pro- ceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques. p. 187–194. SIGGRAPH ’99, ACM Press/Addison-Wesley Publishing Co., USA (1999)

    work page 1999

  6. [6]

    IEEE Transactions on pattern analysis and machine intelligence25(9), 1063–1074 (2003)

    Blanz, V., Vetter, T.: Face recognition based on fitting a 3d morphable model. IEEE Transactions on pattern analysis and machine intelligence25(9), 1063–1074 (2003)

    work page 2003

  7. [7]

    CRC press (2010)

    Botsch, M., Kobbelt, L., Pauly, M., Alliez, P., Lévy, B.: Polygon mesh processing. CRC press (2010)

    work page 2010

  8. [8]

    In: European conference on computer vision

    Chai, Z., Zhang, H., Ren, J., Kang, D., Xu, Z., Zhe, X., Yuan, C., Bao, L.: Realy: Rethinking the evaluation of 3d face reconstruction. In: European conference on computer vision. pp. 74–92. Springer (2022)

    work page 2022

Show all 40 references
  1. [9]

    IEEE Transactions on Biometrics, Behavior, and Identity Science1(1), 32–41 (2019)

    Cook, C.M., Howard, J.J., Sirotin, Y.B., Tipton, J.L., Vemury, A.R.: Demographic effects in facial recognition and their dependence on image acquisition: An evalu- ation of eleven commercial systems. IEEE Transactions on Biometrics, Behavior, and Identity Science1(1), 32–41 (2019)

    2019

  2. [10]

    In: 2024 International Joint Conference on Neural Networks (IJCNN)

    Cui, Z., Yu, J.: Boosting fairness for 3d face reconstruction. In: 2024 International Joint Conference on Neural Networks (IJCNN). pp. 1–9. IEEE (2024)

    2024

  3. [11]

    International Journal of Computer Vision128(2), 547–571 (2020)

    Dai, H., Pears, N., Smith, W., Duncan, C.: Statistical modeling of craniofacial shape and texture. International Journal of Computer Vision128(2), 547–571 (2020)

    2020

  4. [12]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition workshops

    Deng, Y., Yang, J., Xu, S., Chen, D., Jia, Y., Tong, X.: Accurate 3d face recon- struction with weakly-supervised learning: From single image to image set. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition workshops. pp. 0–0 (2019)

    2019

  5. [13]

    ACM Transactions on Graphics (ToG)39(5), 1–38 (2020)

    Egger,B.,Smith,W.A.,Tewari,A.,Wuhrer,S.,Zollhoefer,M.,Beeler,T.,Bernard, F., Bolkart, T., Kortylewski, A., Romdhani, S., et al.: 3d morphable face mod- els—past, present, and future. ACM Transactions on Graphics (ToG)39(5), 1–38 (2020)

    2020

  6. [14]

    ACM Transactions on Graphics (ToG)40(4), 1–13 (2021)

    Feng, Y., Feng, H., Black, M.J., Bolkart, T.: Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG)40(4), 1–13 (2021)

    2021

  7. [15]

    In: 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018)

    Feng, Z.H., Huber, P., Kittler, J., Hancock, P., Wu, X.J., Zhao, Q., Koppen, P., Rätsch, M.: Evaluation of dense 3d reconstruction from 2d face images in the wild. In: 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018). pp. 780–786. IEEE ...

    2018

  8. [16]

    Wash- ington University, St

    Gatzke, T., Grimm, C.: Assessing curvature metrics on triangular meshes. Wash- ington University, St. Louis Missouri, Tech. Rep (2003)

    2003

  9. [17]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Gerogiannis, D., Papantoniou, F.P., Potamias, R.A., Lattas, A., Zafeiriou, S.: Arc2avatar: Generating expressive 3d avatars from a single image via id guidance. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 10770–10782 (2025)

    2025

  10. [18]

    National Institute of Standards and Technology Gaithersburg, MD (2019)

    Grother, P., Ngan, M., Hanaoka, K.: Face recognition vendor test (fvrt): Part 3, demographic effects. National Institute of Standards and Technology Gaithersburg, MD (2019)

    2019

  11. [19]

    In: Geometric modelling, pp

    Hamann, B.: Curvature approximation for triangulated surfaces. In: Geometric modelling, pp. 139–153. Springer (1992)

    1992

  12. [20]

    ACM Transactions on Graphics (ToG)36(6), 1–14 (2017)

    Hu, L., Saito, S., Wei, L., Nagano, K., Seo, J., Fursund, J., Sadeghi, I., Sun, C., Chen, Y.C., Li, H.: Avatar digitization from a single image for real-time rendering. ACM Transactions on Graphics (ToG)36(6), 1–14 (2017)

    2017

  13. [21]

    Lex Localis: Journal of Local Self-Government23(2025)

    Leelavathi, M., Kannan, D.: Enhanced 3d face anti-spoofing for secure biometric authentication using spatio-temporal deep learning networks (stdl-net) with robust feature representation and temporal consistency modeling. Lex Localis: Journal of Local Self-Government23(2025)

    2025

  14. [22]

    SIAM (1998)

    Lehoucq, R.B., Sorensen, D.C., Yang, C.: ARPACK users’ guide: solution of large- scale eigenvalue problems with implicitly restarted Arnoldi methods. SIAM (1998)

    1998

  15. [23]

    ACM Trans

    Li, T., Bolkart, T., Black, M.J., Li, H., Romero, J.: Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph.36(6), 194–1 (2017)

    2017

  16. [24]

    ACM Transactions on Graphics (ToG)37(4), 1–13 (2018)

    Lombardi, S., Saragih, J., Simon, T., Sheikh, Y.: Deep appearance models for face rendering. ACM Transactions on Graphics (ToG)37(4), 1–13 (2018)

    2018

  17. [25]

    In: Proceedings of Berkeley Symposium on Mathematical Statistics and Probability/University of California Press (1967)

    MacQueen, J.: Some methods for classification and analysis of multivariate obser- vations. In: Proceedings of Berkeley Symposium on Mathematical Statistics and Probability/University of California Press (1967)

    1967

  18. [26]

    Computer aided geometric design 17(6), 521–543 (2000)

    Meek, D.S., Walton, D.J.: On surface normal and gaussian curvature approxima- tions given data sampled from a smooth surface. Computer aided geometric design 17(6), 521–543 (2000)

    2000

  19. [27]

    British journal of oral and maxillofacial surgery49(8), e67–e71 (2011)

    Mueller, A., Paysan, P., Schumacher, R., Zeilhofer, H.F., Berg-Boerner, B.I., Mau- rer, J., Vetter, T., Schkommodau, E., Juergens, P., Schwenzer-Zimmerer, K.: Miss- ing facial parts computed by a morphable model and transferred directly to a polyamide laser-sintered prosthesis...

    2011

  20. [28]

    Routledge (2013)

    Myers, J.L., Well, A.D., Lorch Jr, R.F.: Research design and statistical analysis. Routledge (2013)

    2013

  21. [29]

    In: 2009 sixth IEEE interna- tional conference on advanced video and signal based surveillance

    Paysan, P., Knothe, R., Amberg, B., Romdhani, S., Vetter, T.: A 3d face model for pose and illumination invariant face recognition. In: 2009 sixth IEEE interna- tional conference on advanced video and signal based surveillance. pp. 296–301. Ieee (2009)

    2009

  22. [30]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Sanyal, S., Bolkart, T., Feng, H., Black, M.J.: Learning to regress 3d face shape and expression from an image without 3d supervision. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7763–7772 (2019)

    2019

  23. [31]

    In: 2025 IEEE 19th International Conference on Automatic Face and Gesture Recognition (FG)

    Sariyanidi, E., Ferrari, C., Nocentini, F., Berretti, S., Cavallaro, A., Tunc, B.: 3d face reconstruction error decomposed: A modular benchmark for fair and fast method evaluation. In: 2025 IEEE 19th International Conference on Automatic Face and Gesture Recognition (FG). pp. ...

    2025

  24. [32]

    In: European Conference on Computer Vision

    Shang, J., Shen, T., Li, S., Zhou, L., Zhen, M., Fang, T., Quan, L.: Self-supervised monocular 3d face reconstruction by occlusion-aware multi-view geometry consis- tency. In: European Conference on Computer Vision. pp. 53–70. Springer (2020)

    2020

  25. [33]

    In: Computer Graphics Forum

    Sharp, N., Crane, K.: A laplacian for nonmanifold triangle meshes. In: Computer Graphics Forum. vol. 39, pp. 69–80. Wiley Online Library (2020)

    2020

  26. [34]

    arXiv preprint arXiv:2507.11247 (2025)

    Shilova, V., Malherbe, E., Palma, G., Risser, L., Loubes, J.M.: Fairness-aware grouping for continuous sensitive variables: Application for debiasing face analysis with respect to skin tone. arXiv preprint arXiv:2507.11247 (2025)

    2025 arXiv

  27. [35]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Thies, J., Zollhofer, M., Stamminger, M., Theobalt, C., Nießner, M.: Face2face: Real-time face capture and reenactment of rgb videos. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2387–2395 (2016)

    2016

  28. [36]

    arXiv preprint arXiv:1610.03151 (2016)

    Thies, J., Zollhöfer, M., Stamminger, M., Theobalt, C., Nießner, M.: Facevr: Real- time facial reenactment and eye gaze control in virtual reality. arXiv preprint arXiv:1610.03151 (2016)

    2016 arXiv

  29. [37]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wang, Z., Zhu, X., Zhang, T., Wang, B., Lei, Z.: 3d face reconstruction with the geometric guidance of facial part segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1672–1682 (2024)

    2024

  30. [38]

    ACM Transactions on Graphics (TOG)37(4), 1–14 (2018)

    Yamaguchi, S., Saito, S., Nagano, K., Zhao, Y., Chen, W., Olszewski, K., Mor- ishima, S., Li, H.: High-fidelity facial reflectance and geometry inference from an unconstrained image. ACM Transactions on Graphics (TOG)37(4), 1–14 (2018)

    2018

  31. [39]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yang, X., Taketomi, T., Endo, Y., Kanamori, Y.: Makeup prior models for 3d facial makeup estimation and applications. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2165–2176 (2024)

    2024

  32. [40]

    In: European conference on computer vision

    Zielonka, W., Bolkart, T., Thies, J.: Towards metrical reconstruction of human faces. In: European conference on computer vision. pp. 250–269. Springer (2022)

    2022

Pith tools

Reviewed July 9, 2026 · model on record in the stance chip above.