Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Cross-Domain Identity Representation for Skull to Face Matching with Benchmark DataSet

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A Siamese network with triplet loss can match a skull X-ray to the right face, on a new 40-volunteer benchmark.

desk verdict The new X-ray/face dataset is the real contribution, but the triplet-level split leaks identities and the reported identification performance is close to chance. read the letter →

arxiv 2507.08329 v1 pith:M5J4O2QG submitted 2025-07-11 cs.CV

Ravi Shankar Prasad , Dinesh Singh This is my paper · ORCID
classification cs.CV
keywords Siamesenetworkscraniofacialrecognitionskull-to-facematchingcross-domainidentityrepresentationtripletlossX-rayskullimagesfaceretrievalbenchmarkdataset
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 tries to establish that a person can be identified from an X-ray image of their skull by learning a shared identity space for skulls and faces. The proposed solution is a Siamese network with a common backbone, trained with triplet loss so that the embedding of a skull query sits close to the embedding of the same person's face and far from other faces. Because paired skull and face images are scarce, the authors collected their own benchmark dataset from 40 volunteers, with front and side skull X-rays and optical face images. On that dataset the trained models retrieve the correct face among the top matches, and the paper argues this is a step toward automatic craniofacial identification.

What carries the argument

The load-bearing object is the Siamese network with shared backbone, trained on triplets consisting of a skull anchor, a same-identity face positive, and a different-identity face negative. The triplet loss $\max(0, d(a,p) - d(a,n) + \alpha)$ pushes the anchor embedding toward the positive face and away from the negative face. The paper's design choice is to freeze the face-side tower at a pretrained model while keeping the skull-side tower learnable, so the domain gap is absorbed mostly in the skull branch. The architecture is paired with a benchmark dataset of front and side X-ray and face images, from which 12,480 triplets are formed and split 70:30 into training and validation.

What would settle it

Take a set of real skull images (for example, CT volume renders or photographs of skulls) with known identities and run the trained retrieval system: if accuracy on real skulls is close to chance while accuracy on soft-tissue-removed X-rays is high, the central claim collapses. A simpler test is to compare shape landmarks of the X-ray-derived skull images against true skull images and check whether the supposed skull images preserve the same identity-discriminative geometry.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-domain identity representation can be learned directly from paired 2D skull-like X-ray images and optical face images. Using a Siamese architecture in which the face side of the network is a pretrained frozen backbone and the skull side is the same architecture with learnable weights, the model is trained with triplet loss on 12,480 triplets built from four anchor-positive pairs per volunteer. The result is a Euclidean embedding in which a query skull is matched to gallery faces by confidence score $e^{-\delta}$, where $\delta$ is the Euclidean distance between embeddings. Retrieval experiments show the best backbones reach about 75% validation accuracy and top-30 recall around 0.87 on the 40-person gallery, and the method still retrieves faces when the gallery is mixed with an external face dataset.

Load-bearing premise

The whole result rests on the assumption that a 2D X-ray image with soft tissue removed is a valid stand-in for a real forensic skull, preserving the identity-relevant shape cues; the paper states the remaining hard tissue resembles the skull but does not compare against actual skulls.

Editorial extensions

If this is right

  • If the central claim holds, skull-to-face identification becomes a nearest-neighbour retrieval problem in a learned embedding, replacing manual clay reconstruction and subjective superimposition.
  • The collected 40-volunteer paired X-ray and face dataset can serve as a training and evaluation benchmark for craniofacial recognition and reconstruction research.
  • Freezing the face tower and tuning only the skull tower is enough to align the two domains, suggesting pretrained face representations carry usable identity geometry.
  • Retrieval degrades gracefully but stays meaningful when the gallery is enlarged by mixing in an external face dataset, supporting use in realistic missing-person galleries.
  • The confidence score based on Euclidean distance gives a way to threshold matches, so a high score would support a tentative identification claim in practice.

Reading between the lines

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

  • If the X-ray-derived skull images genuinely preserve identity-relevant shape, the same framework could be applied to existing medical X-ray archives, giving far larger training sets than paired CT data.
  • The frozen-face-tower design suggests a natural extension: aligning skull embeddings into a large pretrained face embedding space would allow skull queries to search face galleries built for ordinary face recognition, far beyond the 40-person benchmark.
  • Because all volunteers are Indians aged 21 to 30, a decisive test of the claim requires evaluating on skulls and faces from other age groups and populations, where craniofacial relationships differ.
  • The paper's reliance on soft-tissue-removed 2D X-rays rather than actual skulls means the method's forensic validity ultimately depends on a comparison against real skull imagery; that comparison is the missing experiment.
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.

Referee Report

4 major / 4 minor

Summary. The paper introduces a new dataset (IITMandi_S2F) of paired skull X-ray images and face photographs from 40 volunteers, and proposes a Siamese network trained with triplet loss to retrieve a person's face from a query skull image. The authors experiment with nine CNN backbones and report training and validation accuracy, Recall@30, mAP, and MRR on their dataset, and additionally evaluate retrieval on a mixed gallery containing 485 face images. The central claim is that the learned cross-domain embedding can identify a person from a skull image with satisfactory accuracy.

Significance. If the proposed method and dataset were validated, they would constitute a useful benchmark and a meaningful step toward automated craniofacial identification, an area with limited public data. The dataset of paired X-ray and face images is a potentially valuable resource, and the paper explicitly frames it as a benchmark to encourage further research. The use of triplet loss with cross-domain pairs is a reasonable technical direction. However, the reported results do not currently establish the central identification claim because of flaws in the evaluation protocol, the proximity of the reported recall to random chance, and the unvalidated use of soft-tissue-removed X-rays as skull proxies. The paper also provides no baselines or statistical significance tests, making it difficult to judge whether the observed performance exceeds trivial methods.

major comments (4)
  1. [Section 3 (Dataset split) and Section 5 (Evaluation)] The training/validation split is performed at the triplet level rather than at the identity level. Section 3 states that 12,480 triplets are formed from 40 subjects and then 'split the triplets into training and validation sets with a 70:30 ratio.' Since each subject contributes multiple anchor-positive and negative roles, the same 40 identities appear in both training and validation. Consequently, the retrieval results in Table 1 and Figures 5-8 are evaluated on gallery faces whose identities were seen during training. This identity leakage means the reported recall does not demonstrate generalization to unseen people, which is the core requirement for the forensic identification claim. The decisive check is an identity-disjoint split, where no subject appears in both training and test sets.
  2. [Table 1, Section 5, Figures 5-8] The reported Recall@30 values (0.78-0.87) are evaluated against a 40-face gallery, where random chance is 30/40 = 0.75. The reported values are thus only marginally above chance, and the observed excess could plausibly be attributed to identity memorization from the triplet-level split. Even if the split were corrected, these numbers provide little evidence that the model learns identity-specific cross-domain representations. The paper does not report confidence intervals or a statistical comparison against a chance baseline, so the reader cannot assess whether the observed performance is significant. This near-chance performance undermines the central claim that a high similarity score indicates the skull belongs to the corresponding person.
  3. [Section 3, Dataset construction] The skull images are not actual skull images but 2D X-ray images from which 'soft tissue' is removed, leaving a hard-tissue image that 'resembles the skull.' The paper provides no validation that this processed X-ray image preserves the identity-relevant morphological cues of a real skull, such as its 3D shape or thickness. If the preprocessed X-ray is not a faithful proxy for a forensic skull, the method's results have no bearing on real craniofacial identification. The authors should compare their processed images against actual skull images or CT-derived skull surfaces, or at minimum justify why the hard-tissue X-ray is sufficient for identity matching.
  4. [Section 5, Evaluation protocol and reporting] The evaluation protocol is incompletely specified. Section 5 says 'We use the 70% of the image pairs for training' but the earlier description in Section 3 refers to a 70:30 split of triplets; the relationship between the two is unclear. The composition of the gallery used for Table 1 is not explicitly stated (presumably the 40 face images), and the paper does not clarify whether the query skulls in the validation set are from the same subjects as the gallery faces or whether the mixed-gallery experiments in Figures 7-8 include identity overlap between the training set and the added IISCIFD faces. These details are essential for interpreting the reported numbers and should be clarified.
minor comments (4)
  1. [Throughout] The paper contains numerous typos and grammatical errors, including 'an benchmark', 'the the', and 'IITMandi_S2S' in the caption of Figure 2 versus 'IITMandi_S2F' elsewhere. These should be corrected.
  2. [Equation (1), Section 4] The notation f(x_a^i)* is undefined. It should be clarified that the asterisk denotes learnable parameters for the skull branch, and the embedding notation should be made consistent.
  3. [Section 5, Evaluation] No baseline methods are compared, such as random retrieval, a simple nearest-neighbor baseline on raw image features, or an existing craniofacial superimposition method. Without such baselines, the practical value of the proposed Siamese framework is not established.
  4. [Section 3, Data collection] The dataset is collected from 40 volunteers aged 21-30, mostly from North and East India, with an imbalanced gender ratio (18 female, 22 male). The paper does not discuss the ethical approval for collecting and releasing human X-ray and face data, nor does it provide a clear data-availability statement beyond 'on request.' For a benchmark dataset, these details are important.

Circularity Check

1 steps flagged · score 6.0 of 10

Triplet-level data split lets the same 40 identities occur in both training and validation, so the reported skull-to-face retrieval is evaluated on fitted identities rather than unseen people.

  1. fitted input called prediction [Section 3 (triplet construction and train/validation split); Section 5/Table 1 evaluation]
    "For each volunteer, 4 (= 2 × 2) anchor-positive pairs are created, while each anchor-positive pair can have 78 (= 39 × 2) negatives. Thus, for 𝑛 = 40 volunteers, a total of 12, 480 (= 4 × 78 × 𝑛) triplets are formed. Finally, we split the triplets into training and validation sets with a 70 ∶ 30 ratio, respectively."

    The split is at the triplet level, not the subject level, so every one of the 40 identities appears in both training and validation. The network weights are fitted using skull and face images of all subjects, and the validation then reports retrieval of faces for those same subjects. Table 1 and Figures 5-8 therefore measure how well the model recognizes identities whose images were part of the fitted input, not whether a skull can be matched to a face of an identity never seen during training. With a 40-face gallery, Recall@30 values near 0.69-0.87 are close to the 30/40 = 0.75 chance level, and identity-specific memorization can explain the excess.

full rationale

The proposed Siamese/triplet-loss framework is a standard learning approach and is not circular in itself; there is no load-bearing self-citation, ansatz-smuggling, or renaming of known results. The circularity enters at evaluation: Section 3 splits 12,480 triplets from 40 volunteers by triplet, so the same 40 identities appear in both training and validation. Consequently the central empirical claim, that a skull query retrieves the correct face, is tested against gallery identities whose images were used to fit the model. This makes the validation a retrieval-among-fitted-identities exercise rather than an identity-disjoint generalization test, which is the property required for forensic identification. The separate concern that X-ray hard-tissue images may not faithfully represent real skulls is an external-validity assumption, not a circularity, and it is noted but not scored here.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on two domain assumptions: that X-ray-derived hard-tissue images stand in for real skulls, and that skull geometry carries identity information transferable to face appearance. The hyperparameters and split ratio are free parameters that shape the reported performance.

free parameters (3)
  • Triplet loss margin alpha
    The margin hyperparameter in Eq. (1) controls the separation between positive and negative pairs; the value is chosen by hand, is not reported, and affects all reported accuracies.
  • Soft-tissue removal threshold
    X-ray images are processed to remove soft tissue, a thresholding/segmentation step whose parameters are not specified and which determines the 'skull' appearance.
  • Train/validation split ratio = 70:30
    The 70:30 triplet split is a design choice; because it is not identity-disjoint, it determines the optimistic evaluation.
assumptions (3)
  • domain assumption The hard-tissue portion of a 2D X-ray image resembles a skull well enough to carry identity information.
    Section 3 states the soft tissue is removed so the image 'resembles the skull', but no validation against actual skulls is provided.
  • domain assumption Skull shape determines face appearance sufficiently for automated identity matching.
    This premise underlies the whole skull-to-face task and is assumed rather than demonstrated.
  • domain assumption A shared-embedding triplet-loss Siamese network can align X-ray and optical face images of the same identity.
    Section 4 assumes that minimizing triplet loss across domains yields a meaningful common representation; no theoretical justification is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Domain Identity Representation for Skull to Face Matching with Benchmark DataSet." pith.science (2026). https://pith.science/paper/M5J4O2QG

@misc{pith2026250708329,
  author       = {Pith},
  title        = {Pith review of: Cross-Domain Identity Representation for Skull to Face Matching with Benchmark DataSet},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M5J4O2QG}},
  note         = {Machine review of arXiv:2507.08329}
}
read the original abstract

Craniofacial reconstruction in forensic science is crucial for the identification of the victims of crimes and disasters. The objective is to map a given skull to its corresponding face in a corpus of faces with known identities using recent advancements in computer vision, such as deep learning. In this paper, we presented a framework for the identification of a person given the X-ray image of a skull using convolutional Siamese networks for cross-domain identity representation. Siamese networks are twin networks that share the same architecture and can be trained to discover a feature space where nearby observations that are similar are grouped and dissimilar observations are moved apart. To do this, the network is exposed to two sets of comparable and different data. The Euclidean distance is then minimized between similar pairs and maximized between dissimilar ones. Since getting pairs of skull and face images are difficult, we prepared our own dataset of 40 volunteers whose front and side skull X-ray images and optical face images were collected. Experiments were conducted on the collected cross-domain dataset to train and validate the Siamese networks. The experimental results provide satisfactory results on the identification of a person from the given skull.

Figures

Figures reproduced from arXiv: 2507.08329 by the authors.

Figure 1
Figure 1. Randomly picked positive and negative pairs for evaluation dataset. a moldable material by hand onto an unidentified skull using anatomical data and reference material. According to Wilkinson's detailed examination, this is a very subjective process that requires a great deal of creative interpretation and sometimes produces random results [21]. Later on, mainly there are two main methods for this process. The first… view at source ↗
Figure 2
Figure 2. Paired face and X-ray samples from the prepared dataset IITMandi_S2S for skull to face matching been taken into account in recognition. But, again, this method fails to get the optimal landmark points on the face as landmarks on the 2D and 3D faces are done manually. All these methods face challenges with data insufficiency due to the non-availability of enough pairwise data of the skull and its respective face imag… view at source ↗
Figure 3
Figure 3. Proposed Overall Framework for Skull-to-Face Image Retrieval and Architecture Using Triplet Loss. In this framework, Convolutional Neural Networks (CNNs) serve as the backbone of our proposed Siamese networks, which share a common backbone. For the face images, we utilize pretrained deep models with frozen layers, while for the skull images, we train the same pretrained model with adjustable weights. On the right si… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: 3D t-SNE on the left side visualizes query skull embedding in red star, ground truth face embedding in green triangle, and IITMandi_S2F gallery face embeddings in blue circle. While 3D t-SNE on the right side visualizes the same, but with a mixed gallery of face embedd…
Figure 5
Figure 5. Figure 5: Recall results for retrieving the top k matched faces from the face gallery based on the confidence score based on the Euclidean distance between the given query skull and the gallery face using different models. color, illumination, position and scale of the skull and…
Figure 6
Figure 6. Figure 6: Mean average precision results for retrieving the top k-matched faces from the face gallery based on the confidence score based on the Euclidean distance between the given query skull and the gallery face using different models. 25 50 75 100 125 150 175 200 Top-k 0.0 0…
Figure 9
Figure 9. Figure 9: Qualitative results of top 10 retrievals using (a) ResNet50 and (b) DenseNet121 model. The face with a green border is the correct matched face of the given skull. 𝑆𝐼𝐷01 to 𝑆𝐼𝐷16 are query skulls. ViT_B_16 ResNet101 VGG16 EfficientNet_B0 ResNet18 ConvNeXt_Tiny MobileNe…
Figure 7
Figure 7. Figure 7: Comparison of various models on recall results for retrieving the top 𝑘 matched faces given query skull from the mixed gallery face using different models. 25 50 75 100 125 150 175 200 Top-k 0.25 0.30 0.35 0.40 0.45 0.50 mAP mAP (ResNet18) mAP (ResNet50) mAP (ResNet101…
Figure 10
Figure 10. Figure 10: Comparison of different models for validation accu￾racy, out of which MobileNet_V2, ResNet50, DenseNet201 are showing better validation accuracy. Here, the superscript a, p, and n indicate anchor, positive, and negative image, respectively. However, in the event of a …

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Cranio-ID: Graph-Based Craniofacial Identification via Automatic Landmark Annotation in 2D Multi-View X-rays

    cs.CV 2025-11 reject novelty 4.0 of 10

    Cranio-ID retrieves faces from skull X-rays and sketches by combining YOLO-pose landmark detection with graph-based cross-attention and optimal-transport matching.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Learningrepresen- tationsbymaximizingmutualinformationacrossviews

    Bachman,P.,Hjelm,R.D.,Buchwalter,W.,2019. Learningrepresen- tationsbymaximizingmutualinformationacrossviews. Advancesin neural information processing systems 32

    work page 2019

  2. [2]

    Cran- iofacial reconstruction as a prediction problem using a latent root regression model

    Berar, M., Tilotta, F.M., Glaunès, J.A., Rozenholc, Y., 2011. Cran- iofacial reconstruction as a prediction problem using a latent root regression model. Forensic science international 210, 228–236

    work page 2011

  3. [3]

    Bayesian estimation of optimal craniofacial reconstructions

    Claes, P., Vandermeulen, D., De Greef, S., Willems, G., Clement, J.G., Suetens, P., 2010a. Bayesian estimation of optimal craniofacial reconstructions. Forensic science international 201, 146–152

    work page

  4. [4]

    Computerized craniofacial reconstruction: conceptualframeworkandreview.Forensicscienceinternational201, 138–145

    Claes, P., Vandermeulen, D., De Greef, S., Willems, G., Clement, J.G., Suetens, P., 2010b. Computerized craniofacial reconstruction: conceptualframeworkandreview.Forensicscienceinternational201, 138–145

    work page

  5. [5]

    Damas,S.,Cordón,O.,Ibáñez,O.,Damas,S.,Cordón,O.,Ibáñez,O.,

    work page

  6. [6]

    Forensic identification by computer-aided craniofacial superimposition: a survey

    Damas,S.,Cordon,O.,Ibanez,O.,Santamaria,J.,Alemán,I.,Botella, M., Navarro, F., 2011. Forensic identification by computer-aided craniofacial superimposition: a survey. ACM Computing Surveys (CSUR) 43, 1–27

    work page 2011

  7. [7]

    Craniofacial reconstruction based on multi-linear subspace analysis

    Duan, F., Yang, S., Huang, D., Hu, Y., Wu, Z., Zhou, M., 2014. Craniofacial reconstruction based on multi-linear subspace analysis. Multimedia Tools and Applications 73, 809–823

    work page 2014

  8. [8]

    Dimensionality reduction by learning an invariant mapping, in: 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06), IEEE

    Hadsell, R., Chopra, S., LeCun, Y., 2006. Dimensionality reduction by learning an invariant mapping, in: 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR’06), IEEE. pp. 1735–1742

    work page 2006

Show all 23 references
  1. [9]

    A hierarchical dense deformable model for 3d face reconstruction from skull

    Hu,Y.,Duan,F.,Yin,B.,Zhou,M.,Sun,Y.,Wu,Z.,Geng,G.,2013. A hierarchical dense deformable model for 3d face reconstruction from skull. Multimedia tools and applications 64, 345–364

    2013

  2. [10]

    Huang,J.,Zhou,M.,Duan,F.,Deng,Q.,Wu,Z.,Tian,Y.,2011. The weightedlandmark-basedalgorithmforskullidentification,in:Com- puterAnalysisofImagesandPatterns:14thInternationalConference, CAIP 2011, Seville, Spain, August 29-31, 2011, Proceedings, Part II 14, Springer. pp. 42–48

    2011

  3. [11]

    Robust3dfacemodeling and reconstruction from frontal and side images

    Jin,H.,Wang,X.,Zhong,Z.,Hua,J.,2017. Robust3dfacemodeling and reconstruction from frontal and side images. Computer Aided Geometric Design 50, 1–13

    2017

  4. [12]

    Are you from north or south india? a hard face-classificationtaskrevealssystematicrepresentationaldifferences between humans and machines

    Katti, H., Arun, S., 2019. Are you from north or south india? a hard face-classificationtaskrevealssystematicrepresentationaldifferences between humans and machines. Journal of Vision 19, 1–1

    2019

  5. [13]

    Siamese network features forimagematching,in:201623rdinternationalconferenceonpattern recognition (ICPR), IEEE

    Melekhov, I., Kannala, J., Rahtu, E., 2016. Siamese network features forimagematching,in:201623rdinternationalconferenceonpattern recognition (ICPR), IEEE. pp. 378–383

    2016

  6. [14]

    The 3d reconstructionof facial features usingvolumedistortion,in:Proc.14thEurographicsUKConference

    Michael, S.,Chen, M., 1996. The 3d reconstructionof facial features usingvolumedistortion,in:Proc.14thEurographicsUKConference

    1996

  7. [15]

    Forensic facial reconstruction of skeletonized and highlydecomposedhumanremains,in:ForensicGeneticApproaches forIdentificationofHumanSkeletalRemains.Elsevier,pp.549–569

    Missal, S., 2023. Forensic facial reconstruction of skeletonized and highlydecomposedhumanremains,in:ForensicGeneticApproaches forIdentificationofHumanSkeletalRemains.Elsevier,pp.549–569

    2023

  8. [16]

    A template-deformation method for facial reproduction, in: ISPA 2005

    Muller, J., Mang, A., Buzug, T., 2005. A template-deformation method for facial reproduction, in: ISPA 2005. Proceedings of the 4th International Symposium on Image and Signal Processing and Analysis, 2005., IEEE. pp. 359–364

    2005

  9. [17]

    Face reconstruction from skull shapes and physical attributes, in: Pattern Recognition: 31st DAGM Symposium, Jena, Germany, September 9-11, 2009

    Paysan,P.,Lüthi,M.,Albrecht,T.,Lerch,A.,Amberg,B.,Santini,F., Vetter, T., 2009. Face reconstruction from skull shapes and physical attributes, in: Pattern Recognition: 31st DAGM Symposium, Jena, Germany, September 9-11, 2009. Proceedings 31, Springer. pp. 232– 241

    2009

  10. [18]

    Stir: Siamese transformerforimageretrievalpostprocessing,in:Proceedingsofthe 33rd ACM International Conference on Information and Knowledge Management, pp

    Shabanov, A., Tarasov, A., Nikolenko, S., 2024. Stir: Siamese transformerforimageretrievalpostprocessing,in:Proceedingsofthe 33rd ACM International Conference on Information and Knowledge Management, pp. 4852–4857

    2024

  11. [19]

    Automatic face recognition from skeletal remains, in: 2007IEEEConferenceonComputerVisionandPatternRecognition, IEEE

    Tu, P., Book, R., Liu, X., Krahnstoever, N., Adrian, C., Williams, P., 2007. Automatic face recognition from skeletal remains, in: 2007IEEEConferenceonComputerVisionandPatternRecognition, IEEE. pp. 1–7. Prasad and Singh: Preprint submitted to Elsevier Page 6 of 7 Cross-domain ...

    2007

  12. [20]

    Computerized craniofacial reconstruction using ct-derived implicit surface representations

    Vandermeulen, D., Claes, P., Loeckx, D., De Greef, S., Willems, G., Suetens, P., 2006. Computerized craniofacial reconstruction using ct-derived implicit surface representations. Forensic science international 159, S164–S174

    2006

  13. [21]

    Facial reconstruction–anatomical art or artistic anatomy? Journal of anatomy 216, 235–250

    Wilkinson, C., 2010. Facial reconstruction–anatomical art or artistic anatomy? Journal of anatomy 216, 235–250

    2010

  14. [22]

    mm3dface:Nonintrusive3dfacialreconstructionleveragingmmwave signals, in: Proceedings of the 21st Annual International Conference on Mobile Systems, Applications and Services, pp

    Xie, J., Kong, H., Yu, J., Chen, Y., Kong, L., Zhu, Y., Tang, F., 2023. mm3dface:Nonintrusive3dfacialreconstructionleveragingmmwave signals, in: Proceedings of the 21st Annual International Conference on Mobile Systems, Applications and Services, pp. 462–474. Ravi Shankar Pras...

    2023

  15. [2020]

    Handbookoncraniofacialsuperimposition: The MEPROCS project , 11–50

    Relationships between the skull and the face for forensic cran- iofacialsuperimposition. Handbookoncraniofacialsuperimposition: The MEPROCS project , 11–50

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.