Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

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

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Cranio-ID claims that automatic landmark detection plus graph-based semantic correspondence can match skull X-rays to face photos and to sketches, outperforming existing cross-modal retrieval methods on the S2F and CUHK datasets.

desk verdict The landmark detection is real work, but the split is arithmetically impossible, the hyperparameters were tuned on the test set, and the OT/CA equations don't define local alignment — the retrieval results shouldn't be cited. read the letter →

arxiv 2511.14411 v1 pith:6IGYXKAD submitted 2025-11-18 cs.CV

Ravi Shankar Prasad , Nandani Sharma , Dinesh Singh This is my paper
classification cs.CV
keywords craniofacialidentificationlandmarklocalizationcross-modalmatchinggraphneuralnetworkoptimaltransportcross-attentionskull-to-faceretrievalforensic
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

The paper seeks to establish that craniofacial identification can be automated from ordinary 2D skull X-rays, replacing slow manual landmark placement. Its claim is that a pipeline which localizes anatomical landmarks with a pose-detection model, arranges them into graphs, and then aligns those graphs across modalities using cross-attention and optimal transport can retrieve the matching face photo with high recall. On the skull-to-face dataset the best configuration reaches 50.0% R@1 and 85.8% R@20; on the sketch-to-face dataset it reaches 88.4% R@1 and 90.6% R@20. If the claim holds, forensic examiners could search face databases from a skull image alone, and the same correspondence machinery would transfer to sketch-to-face identification.

What carries the argument

The load-bearing mechanism is the landmark graph: each of the automatically detected craniofacial landmarks supplies a node, whose features are the landmark coordinates concatenated with a patch embedding from a pretrained network, and edges connect the k nearest landmarks by Euclidean distance. This graph is embedded with a two-layer graph convolutional network, fused with a global ViT embedding of the whole image, and then aligned across modalities by bidirectional multi-head cross-attention followed by an entropic-regularized optimal transport that computes a soft correspondence between the two feature sequences. The transport cost and a combined global-plus-local triplet loss are optimiz

What would settle it

Enumerate the subject IDs in the test and training splits of the skull-to-face dataset and check whether any subject appears in both; a single repeated identity would invalidate the 50.0% R@1 and 85.8% R@20 results. Alternatively, retrain with a strict disjoint split of the 51 subjects (for example 10 test subjects rather than 10 pairs) and compare whether recall@k drops substantially.

Watch

Extended reading notes

Core claim

Cranio-ID's central claim is that a skull and a face photo of the same person can be matched by representing each image as a graph whose nodes are local patches around automatically detected craniofacial landmarks, with edges defined by k-nearest neighbors in landmark space, and then learning a semantic correspondence between the two graphs. The paper argues that this landmark-based graph representation, enriched by global vision-transformer features and refined through bidirectional cross-attention followed by entropic optimal transport, closes the large domain gap between X-ray and visible-light imagery. Under the reported evaluation setting, the method outperforms the baselines it compare

Load-bearing premise

The load-bearing premise is that the test set contains only identities never seen in training; the described split into 10 test, 21 validation, and 71 training pairs does not arithmetically match the dataset's 51 subjects, so identity leakage would inflate every reported recall number.

Editorial extensions

If this is right

  • If a skull X-ray can retrieve a face photo at 85.8% top-20 recall, forensic identification could begin with an automated database search instead of manual landmarking and superimposition.
  • The same landmark-graph plus cross-attention/optimal-transport pipeline transfers to sketch-to-face matching, reaching 88.4% top-1 recall, indicating the correspondence machinery is not specific to X-rays.
  • Automatic landmark annotation removes the need for expert manual landmark placement; the reported localization mAP around 99.5% for frontal and side face and skull suggests detection is accurate enough to feed the matching stage.
  • Ablations show that optimal transport and cross-attention each contribute gains over raw global similarity, and the best configuration uses patch size 128 and margin 0.3, giving a concrete recipe for practitioners.

Reading between the lines

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

  • Editorial inference: the paper evaluates on tens of identities, not on large real-world face galleries, so scaling from the reported 85.8% R@20 to databases of thousands or millions of faces is an open question.
  • Editorial inference: because the pipeline depends on accurate landmarks, a testable prediction is that degrading or perturbing keypoint coordinates should lower retrieval recall; this could be measured directly by injecting noise into the detected landmarks.
  • Editorial inference: since the framework transfers to sketches, it may extend to other modalities such as 3D skull reconstructions or composite drawings, and the same cross-attention/optimal-transport alignment objective could be tested across those domains.
  • Editorial inference: the landmark-graph design is more interpretable than end-to-end embeddings, since one can inspect which landmark pairs the optimal transport couples; this could aid forensic validation if the correspondence maps were visualized per case.
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 proposes Cranio-ID, a pipeline for craniofacial identification from 2D skull X-rays and face photographs. The framework first detects anatomical landmarks with YOLO-pose models, then builds per-landmark patch-based graphs, extracts node features with pretrained CNNs/ViTs, passes them through GCNs, and finally aligns the two modalities via cross-attention and entropic optimal transport, trained with a combined triplet/OT loss. The authors validate the framework on the S2F skull-to-face dataset and the CUHK sketch-to-face dataset, reporting Recall@K and mAP@K for several backbones, and claim that the proposed method outperforms all other methods. The central retrieval claim is supported only by ablations within the proposed pipeline.

Significance. If the central claim were supported, the paper would contribute a practical, largely automatic alternative to manual craniofacial identification from 2D X-rays, with a natural extension to sketch-to-face matching. The landmark localization results in Table 1 are strong, and the framework is sensibly motivated by the difficulty of manual annotation. However, the evaluation as presented does not establish that Cranio-ID generalizes to unseen subjects: the split is arithmetically inconsistent with a subject-disjoint partition, hyperparameters are selected on the test set, and no external baseline anchors the retrieval numbers. The significance is therefore currently unproven.

major comments (4)
  1. [Section 3 (dataset split)] The split is arithmetically inconsistent with a subject-wise partition. S2F contains 51 subjects and 102 skull–face pairs (one lateral and one frontal per subject); a subject-wise split must produce even counts in every subset. The text states 10 test, 21 validation, and 71 training pairs. Since 21 is odd, the validation set cannot be subject-wise. If pairs were instead split randomly, the same subject's skull and face can appear in both training and test, leaking identity information for a person-identification task. Because Table 2's R@K/mAP@K are the headline result, the authors must report subject IDs/overlap statistics and re-run the evaluation on a genuinely subject-disjoint split.
  2. [Section 5.3 / Tables 3–4 (test-set model selection)] The paper explicitly states that ablation studies are conducted on the testing set and chooses d=128 and m=0.3 from those tables. This means hyperparameters are selected on the test set, so the numbers in Table 2 are optimistic estimates of generalization. The authors should either use a validation set for model selection or report a nested/independent test protocol with the final hyperparameters frozen before any test-set evaluation.
  3. [Section 5.2 / Table 2 (no external baseline)] Table 2 only compares variants of the proposed pipeline with OT/CA on/off. No existing skull-to-face or sketch-to-face method is included. The conclusion that "our proposed method outperforms all other methods" is not supported by these comparisons. At minimum, the paper should compare against standard cross-domain matching baselines (e.g., pretrained embeddings with cosine similarity, or previously published methods on S2F/CUHK) under the same split.
  4. [Sections 4.2–4.4 / Eqs. (7)–(16) (method specification)] The notation for F_gs is inconsistent. Eq. (7) defines F_gs as a single concatenated vector in R^d, but Eq. (11) sums over T_s and Eq. (15) indexes F_enh_s and F_enh_f by i/j, implying sequence inputs; the OT cost matrix is also typed as R^{d×d} where it should be sequence-length × sequence-length. As written, the cross-attention and OT modules are not implementable from the equations. The authors should clarify whether 'd' denotes feature dimension or sequence length and provide exact tensor shapes for the attention and OT steps.
minor comments (4)
  1. [Abstract, §1, §3.1] The abstract and introduction say "total 19 landmarks," while §3.1 and Figure 1 describe 18 frontal and 13 lateral landmarks. The total is unclear (18+13=31); please state the exact landmark count and how the two views relate.
  2. [Eq. (1)] Eq. (1) is missing a comma in the subscript: it should be I[x_i−P:x_i+P, y_i−P:y_i+P]. Algorithm 1 uses the order y then x; make the notation consistent.
  3. [Section 3 / CUHK dataset] The split is described only for S2F. The CUHK sketch-to-face dataset is used for the second half of Table 2, but the train/validation/test protocol for CUHK is not stated. Please add this information and the number of subjects/identities in each split.
  4. [Table 3 caption] The caption says "hidden space dimensionality (d), corresponding to the patch size per keypoint," but d values are 32/64/128. The relationship between patch size and hidden dimension should be clarified: is d the patch side length or the node feature dimension?

Circularity Check

2 steps flagged · score 6.0 of 10

Test-set hyperparameter selection and an arithmetically impossible subject-disjoint split make the reported skull-to-face retrieval numbers partly fitted rather than predicted.

  1. fitted input called prediction [Section 5.3 (Ablation studies), Tables 3-4; reported in Section 5.2/Table 2]
    "Ablation studies are conducted on S2F and CUHK dataset testing set ... We obtain best performance when m is set to 0.3."

    The headline Table 2 results (e.g., R@1 = 50.0, R@20 = 85.8 for ViT, d=128, m=0.3) are obtained after selecting patch size d and margin m by evaluating on the testing set. Thus the test set is reused for model selection: the reported score is the best over the tested configurations on the test set, not an unbiased out-of-sample prediction. The test-selected hyperparameters are then presented as the method's performance, so the central retrieval numbers are partly a fit to the test data rather than a prediction.

  2. fitted input called prediction [Section 3 (Additional Curation of S2F Dataset), with Table 2]
    "More precisely, for training YOLO pose models, this dataset contains 102 skull–face pairs, which consist of 51 lateral views and 51 frontal views. ... Specifically, 10 pairs were subject wise randomly selected for testing, 21 for validation, and 71 for training."

    The S2F dataset has 51 subjects each contributing two views, so a subject-wise partition must assign both views of a subject to the same subset, requiring even subset sizes. A validation set of 21 pairs is arithmetically impossible for a subject-wise split. If the actual split is pair-wise random, the same subject can appear in both training and test. Because the task is person identification, this identity leakage means the Table 2 R@K/mAP@K values are not a clean generalization measure; they are partially forced by the model having seen the test identities during training.

full rationale

Most of the technical derivation (graph construction in Eqs. 1-4, GCN embedding in Eqs. 5-6, cross-attention in Eqs. 9-14, OT in Eqs. 15-20, and the triplet loss in Eqs. 21-25) is self-contained; no equation reduces to its input by construction, and the framework is also tested on the external CUHK benchmark. Self-citations to the authors' S2F dataset [37,38] and their earlier graph works [33,41] are present but are not the mechanism that forces the result. However, the central empirical claim is partially a fitted quantity. Section 5.3 explicitly performs ablation/hyperparameter selection on the testing set and then reports those test-selected numbers as the method's performance. In addition, the Section 3 'subject-wise' split is arithmetically inconsistent with the 51-subject/102-pair dataset (21 is odd), so the actual split likely allows the same subject in both training and test, leaking identity information. These are evaluation-design circularities: the headline retrieval numbers are not a clean out-of-sample prediction. Because the core framework has independent content, the score is 6 (partial circularity) rather than higher.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities. It relies on several unstated or test-set-fitted hyperparameters (beta, lambda_OT, k, d, m) and on the domain assumption that skull X-ray landmarks and face/sketch landmarks are semantically aligned. The benchmark dataset itself originates in the authors' prior work, which increases the circularity burden.

free parameters (5)
  • Patch size / hidden dimension d = 128 (selected via test-set ablation, Table 3)
    Controls node feature dimension; best R@1 on S2F was chosen by evaluating directly on the testing set.
  • Triplet margin m = 0.3 (Table 4)
    Ablated on the testing set; the value 0.3 is used in the final results.
  • Similarity weighting beta = not reported
    Introduced in Eqs. 22-23 to balance global and OT similarities; no value or ablation is given.
  • OT loss weight lambda_OT = not reported
    Appears in Eq. 25 as the total training objective weight; no value is reported.
  • k-NN graph degree k = not reported
    Defines graph edges in Eq. 4 but the value of k is never specified in the paper.
assumptions (4)
  • domain assumption Landmarks on skull X-rays after soft-tissue elimination correspond semantically to landmarks on optical face and sketch images.
    Used throughout Section 4; the graph-matching problem assumes this correspondence, but no anatomical validation or quantitative assessment is provided.
  • domain assumption S2F skull images are correlated enough with CUHK sketch images, and S2F face images with CUHK face images, to draw conclusions across the two datasets.
    Section 5 states this correlation without quantitative evidence, yet it underlies the cross-dataset claims.
  • domain assumption Manual soft-tissue elimination on all S2F X-rays faithfully isolates skeletal structures without removing identity-bearing features.
    Section 3 describes manual outlining via Roboflow; the accuracy and consistency of this preprocessing are not evaluated.
  • standard math Entropic optimal transport with Sinkhorn iterations produces a meaningful transport plan for the given features.
    Standard OT theory, invoked in Eq. 16; however, the input dimensions in the written equations are mismatched.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cranio-ID: Graph-Based Craniofacial Identification via Automatic Landmark Annotation in 2D Multi-View X-rays." pith.science (2026). https://pith.science/paper/6IGYXKAD

@misc{pith2026251114411,
  author       = {Pith},
  title        = {Pith review of: Cranio-ID: Graph-Based Craniofacial Identification via Automatic Landmark Annotation in 2D Multi-View X-rays},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6IGYXKAD}},
  note         = {Machine review of arXiv:2511.14411}
}
read the original abstract

In forensic craniofacial identification and in many biomedical applications, craniometric landmarks are important. Traditional methods for locating landmarks are time-consuming and require specialized knowledge and expertise. Current methods utilize superimposition and deep learning-based methods that employ automatic annotation of landmarks. However, these methods are not reliable due to insufficient large-scale validation studies. In this paper, we proposed a novel framework Cranio-ID: First, an automatic annotation of landmarks on 2D skulls (which are X-ray scans of faces) with their respective optical images using our trained YOLO-pose models. Second, cross-modal matching by formulating these landmarks into graph representations and then finding semantic correspondence between graphs of these two modalities using cross-attention and optimal transport framework. Our proposed framework is validated on the S2F and CUHK datasets (CUHK dataset resembles with S2F dataset). Extensive experiments have been conducted to evaluate the performance of our proposed framework, which demonstrates significant improvements in both reliability and accuracy, as well as its effectiveness in cross-domain skull-to-face and sketch-to-face matching in forensic science.

Figures

Figures reproduced from arXiv: 2511.14411 by the authors.

Figure 1
Figure 1. Sample image showing landmark localization on the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Sample image of X-ray dataset used, where soft tis [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The proposed framework consists of five stages. In the first stage, the face and skull regions are detected, and keypoints [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Predictions on test dataset samples for landmark [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Top-10 retrieval results for given query in S2F [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: (a) 2D visualisation of embeddings before training shows a large domain gap between two modalities (i.e., skull and [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in 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-Diff: Diffusion-based Cross-domain Craniofacial Reconstruction with 2D X-ray Skull Guidance and Structural Identity Constraints

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    Cranio-Diff applies a conditioned diffusion model with ControlNet and text prompts to reconstruct faces from skull X-rays, evaluated on a synthesized dataset of 4320 samples from 120 subjects across age and BMI variations.

Reference graph

Works this paper leans on

59 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Ter- ca-wgnn: trimodel emotion recognition using cumulative attribute-weighted graph neural network.Applied Sci- ences, 14(6):2252, 2024

    Hussein Farooq Tayeb Al-Saadawi and Resul Das. Ter- ca-wgnn: trimodel emotion recognition using cumulative attribute-weighted graph neural network.Applied Sci- ences, 14(6):2252, 2024. 2

    2024

  2. [2]

    Automatic landmark 8 annotation in 3d surface scans of skulls: Methodological proposal and reliability study.Computer Methods and Programs in Biomedicine, 210:106380, 2021

    Enrique Bermejo, Kei Taniguchi, Yoshinori Ogawa, Rub´en Martos, Andrea Valsecchi, Pablo Mesejo, Oscar Ib´a˜nez, and Kazuhiko Imaizumi. Automatic landmark 8 annotation in 3d surface scans of skulls: Methodological proposal and reliability study.Computer Methods and Programs in Biomedicine, 210:106380, 2021. 1

    2021

  3. [3]

    Bookstein.Morphometric Tools for Landmark Data: Geometry and Biology

    Fred L. Bookstein.Morphometric Tools for Landmark Data: Geometry and Biology. Cambridge University Press, Cambridge, UK, 1997. 2

    1997

  4. [4]

    Dual subspace manifold learning based on gcn for intensity-invariant fa- cial expression recognition.Pattern Recognition, 148: 110157, 2024

    Jingying Chen, Jinxin Shi, and Ruyi Xu. Dual subspace manifold learning based on gcn for intensity-invariant fa- cial expression recognition.Pattern Recognition, 148: 110157, 2024. 2

    2024

  5. [5]

    Bayesian estimation of optimal craniofacial reconstruc- tions.Forensic science international, 201(1-3):146–152,

    Peter Claes, Dirk Vandermeulen, Sven De Greef, Guy Willems, John Gerald Clement, and Paul Suetens. Bayesian estimation of optimal craniofacial reconstruc- tions.Forensic science international, 201(1-3):146–152,

  6. [6]

    Com- puterized craniofacial reconstruction: conceptual frame- work and review.Forensic science international, 201(1- 3):138–145, 2010

    Peter Claes, Dirk Vandermeulen, Sven De Greef, Guy Willems, John Gerald Clement, and Paul Suetens. Com- puterized craniofacial reconstruction: conceptual frame- work and review.Forensic science international, 201(1- 3):138–145, 2010. 2

    2010

  7. [7]

    Forensic identification by computer- aided craniofacial superimposition: a survey.ACM Com- puting Surveys (CSUR), 43(4):1–27, 2011

    Sergio Damas, Oscar Cordon, Oscar Ibanez, Jose San- tamaria, Inmaculada Alem ´an, Miguel Botella, and Fer- nando Navarro. Forensic identification by computer- aided craniofacial superimposition: a survey.ACM Com- puting Surveys (CSUR), 43(4):1–27, 2011. 1, 2

    2011

  8. [8]

    Springer Nature, 2020

    Sergio Damas, Oscar Cord ´on, and Oscar Ib ´a˜nez.Hand- book on craniofacial superimposition: The MEPROCS project. Springer Nature, 2020. 1, 2

    2020

Show all 59 references
  1. [9]

    3d semi-landmarks based statistical face reconstruction.Journal of computing and Information technology, 14(1):31–43, 2006

    Michel Desvignes, Gerard Bailly, Yohan Payan, and Maxime Berar. 3d semi-landmarks based statistical face reconstruction.Journal of computing and Information technology, 14(1):31–43, 2006. 2

    2006

  2. [10]

    Attentional visual graph neural network based facial expression recognition method.Signal, Im- age and Video Processing, 18(12):8693–8705, 2024

    Wenmin Dong, Xiangwei Zheng, Lifeng Zhang, and Yuang Zhang. Attentional visual graph neural network based facial expression recognition method.Signal, Im- age and Video Processing, 18(12):8693–8705, 2024. 2

    2024

  3. [11]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Trans- formers for image recognition ...

    2021

  4. [12]

    Variations in facial relationships: their significance in treatment and prognosis.American journal of orthodontics, 34(10):812–840, 1948

    William B Downs. Variations in facial relationships: their significance in treatment and prognosis.American journal of orthodontics, 34(10):812–840, 1948. 1

    1948

  5. [13]

    Craniofacial recon- struction based on multi-linear subspace analysis.Multi- media Tools and Applications, 73(2):809–823, 2014

    Fuqing Duan, Sen Yang, Donghua Huang, Yongli Hu, Zhongke Wu, and Mingquan Zhou. Craniofacial recon- struction based on multi-linear subspace analysis.Multi- media Tools and Applications, 73(2):809–823, 2014. 2

    2014

  6. [14]

    3d facial landmarks: Inter-operator variability of manual annotation.BMC medical imaging, 14(1):35, 2014

    Jens Fagertun, Stine Harder, Anders Rosengren, Chris- tian Moeller, Thomas Werge, Rasmus R Paulsen, and Thomas F Hansen. 3d facial landmarks: Inter-operator variability of manual annotation.BMC medical imaging, 14(1):35, 2014. 1

    2014

  7. [15]

    Inter- national validation study of ai-guided craniofacial super- imposition in a contemporary population sample.Foren- sic Science International, page 112628, 2025

    Rosario Guerra, Rub ´en Martos, ´Oscar Ib ´a˜nez, An- drea Valsecchi, Enrique Bermejo, Stefano De Luca, Mar´ıa Alejandra Guativonza, Guillermo R-Garc ´ıa, Ver´onica Mart´ınez-Garc´ıa, Daniel Casallas, et al. Inter- national validation study of ai-guided craniofacial super- imp...

    2025

  8. [16]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. 4

    2016

  9. [17]

    Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam

    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision appli- cations.arXiv preprint arXiv:1704.04861, 2017. 4

    2017 arXiv

  10. [18]

    A hier- archical dense deformable model for 3d face reconstruc- tion from skull.Multimedia tools and applications, 64 (2):345–364, 2013

    Yongli Hu, Fuqing Duan, Baocai Yin, Mingquan Zhou, Yanfeng Sun, Zhongke Wu, and Guohua Geng. A hier- archical dense deformable model for 3d face reconstruc- tion from skull.Multimedia tools and applications, 64 (2):345–364, 2013. 2

    2013

  11. [19]

    Modeling fine-grained relations in dynamic space-time graphs for video-based facial expression recognition.IEEE Transactions on Affective Computing,

    Changqin Huang, Fan Jiang, Zhongmei Han, Xiaodi Huang, Shijin Wang, Yanlai Zhu, Yunliang Jiang, and Bin Hu. Modeling fine-grained relations in dynamic space-time graphs for video-based facial expression recognition.IEEE Transactions on Affective Computing,

  12. [20]

    The weighted landmark-based algorithm for skull identification

    Jingbo Huang, Mingquan Zhou, Fuqing Duan, Qingqong Deng, Zhongke Wu, and Yun Tian. The weighted landmark-based algorithm for skull identification. InIn- ternational Conference on Computer Analysis of Images and Patterns, pages 42–48. Springer, 2011. 2

    2011

  13. [21]

    Facial soft tissue thickness database for cran- iofacial reconstruction in korean adults.Journal of foren- sic sciences, 57(6):1442–1447, 2012

    Hyeon-Shik Hwang, Myoung-Kyu Park, Won-Joon Lee, Jin-Hyoung Cho, Byung-Kuk Kim, and Caroline M Wilkinson. Facial soft tissue thickness database for cran- iofacial reconstruction in korean adults.Journal of foren- sic sciences, 57(6):1442–1447, 2012. 1

    2012

  14. [22]

    Conceptual transitions in methods of skull-photo superimposition that impact the reliability of identification: a review.Forensic science international, 246:110–121, 2015

    Paul T Jayaprakash. Conceptual transitions in methods of skull-photo superimposition that impact the reliability of identification: a review.Forensic science international, 246:110–121, 2015. 1

    2015

  15. [23]

    Transformer embedded spectral-based graph network for facial expression recognition.International Journal of Machine Learning and Cybernetics, 15(6):2063–2077,

    Xing Jin, Xulin Song, Xiyin Wu, and Wenzhu Yan. Transformer embedded spectral-based graph network for facial expression recognition.International Journal of Machine Learning and Cybernetics, 15(6):2063–2077,

  16. [24]

    Gcf: Graph convolutional networks for facial expression recognition

    Hozaifa Kassab, Mohamed Bahaa, and Ali Hamdi. Gcf: Graph convolutional networks for facial expression recognition. In2024 Intelligent Methods, Systems, and Applications (IMSA), pages 166–171. IEEE, 2024. 2

    2024

  17. [25]

    A three- dimensional analysis of soft and hard tissue changes af- ter a mandibular setback surgery.Computer methods and programs in biomedicine, 83(3):178–187, 2006

    Nam-Kug Kim, Cheol Lee, Suk-Ho Kang, Jae-Woo Park, Myung-Jin Kim, and Young-Il Chang. A three- dimensional analysis of soft and hard tissue changes af- ter a mandibular setback surgery.Computer methods and programs in biomedicine, 83(3):178–187, 2006. 1

    2006

  18. [26]

    Craniofacial an- thropometry: practical measurement of the head and face for clinical, surgical, and research use.(No Title), 1997

    John C Kolar and Elizabeth M Salter. Craniofacial an- thropometry: practical measurement of the head and face for clinical, surgical, and research use.(No Title), 1997. 1 9

    1997

  19. [27]

    Cross-domain image matching with deep feature maps.International Journal of Computer Vision, 127(11):1738–1750, 2019

    Bailey Kong, James Supan ˘cic˘ III, Deva Ramanan, and Charless C Fowlkes. Cross-domain image matching with deep feature maps.International Journal of Computer Vision, 127(11):1738–1750, 2019. 3

    2019

  20. [28]

    Radio- graphic evaluation of orthodontic treatment by means of four different cephalometric superimposition methods

    Marcos Augusto Lenza, Adilson Alves de Carvalho, Ed- uardo Beaton Lenza, Mauricio Guilherme Lenza, Hianne Miranda de Torres, and Jo ˜ao Batista de Souza. Radio- graphic evaluation of orthodontic treatment by means of four different cephalometric superimposition methods. Dental...

    2015

  21. [29]

    Shuai Liu, Shichen Huang, Weina Fu, and Jerry Chun- Wei Lin. A descriptive human visual cognitive strategy using graph neural network for facial expression recog- nition.International Journal of Machine Learning and Cybernetics, 15(1):19–35, 2024. 2

    2024

  22. [30]

    Semantic correspondence as an optimal transport prob- lem

    Yanbin Liu, Linchao Zhu, Makoto Yamada, and Yi Yang. Semantic correspondence as an optimal transport prob- lem. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4463– 4472, 2020. 2

    2020

  23. [31]

    Facial action units as a joint dataset training bridge for facial expression recognition.IEEE Transac- tions on Multimedia, 2025

    Shuyi Mao, Xinpeng Li, Fan Zhang, Xiaojiang Peng, and Yang Yang. Facial action units as a joint dataset training bridge for facial expression recognition.IEEE Transac- tions on Multimedia, 2025. 2

    2025

  24. [32]

    Forensic facial reconstruction of skele- tonized and highly decomposed human remains

    Stephen Missal. Forensic facial reconstruction of skele- tonized and highly decomposed human remains. In Forensic genetic approaches for identification of human skeletal remains, pages 549–569. Elsevier, 2023. 2

    2023

  25. [33]

    Para-X: Graph- based Facial Paralysis Detection using Structural De- formations of Facial Expression

    Dinesh Singh Nandani Sharma, Kajal. Para-X: Graph- based Facial Paralysis Detection using Structural De- formations of Facial Expression. InProceedings of the International Joint Conference on Neural Networks (IJCNN), 2025. Presented June 2025. 2

    2025

  26. [34]

    Facial landmark-based emotion recognition via directed graph neural network.Electronics, 9(5):764,

    Quang Tran Ngoc, Seunghyun Lee, and Byung Cheol Song. Facial landmark-based emotion recognition via directed graph neural network.Electronics, 9(5):764,

  27. [35]

    Face reconstruction from skull shapes and phys- ical attributes

    Pascal Paysan, Marcel L ¨uthi, Thomas Albrecht, Anita Lerch, Brian Amberg, Francesco Santini, and Thomas Vetter. Face reconstruction from skull shapes and phys- ical attributes. InJoint Pattern Recognition Symposium, pages 232–241. Springer, 2009. 2

    2009

  28. [36]

    Sparse-to-dense feature matching: Intra and inter domain cross-modal learning in domain adaptation for 3d semantic segmentation

    Duo Peng, Yinjie Lei, Wen Li, Pingping Zhang, and Yu- lan Guo. Sparse-to-dense feature matching: Intra and inter domain cross-modal learning in domain adaptation for 3d semantic segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 71...

    2021

  29. [37]

    Cross-domain identity representation for skull to face matching with benchmark dataset.arXiv preprint arXiv:2507.08329,

    Ravi Shankar Prasad and Dinesh Singh. Cross-domain identity representation for skull to face matching with benchmark dataset.arXiv preprint arXiv:2507.08329,

    arXiv

  30. [38]

    Fcr: Investigat- ing generative ai models for forensic craniofacial recon- struction, 2025

    Ravi Shankar Prasad and Dinesh Singh. Fcr: Investigat- ing generative ai models for forensic craniofacial recon- struction, 2025. 3, 6

    2025

  31. [39]

    Design and research of facial ex- pression recognition system based on key point extrac- tion.KSII Transactions on Internet & Information Sys- tems, 19(1), 2025

    Yan Qu and Yan Liu. Design and research of facial ex- pression recognition system based on key point extrac- tion.KSII Transactions on Internet & Information Sys- tems, 19(1), 2025. 2

    2025

  32. [40]

    new mor- phometry

    Ann H Ross, Ashley H McKeown, and Lyle W Konigs- berg. Allocation of crania to groups via the “new mor- phometry”.Journal of forensic sciences, 44(3):584–587,

  33. [41]

    Exp-graph: How connections learn facial attributes in graph-based ex- pression recognition.arXiv preprint arXiv:2507.14608,

    Nandani Sharma and Dinesh Singh. Exp-graph: How connections learn facial attributes in graph-based ex- pression recognition.arXiv preprint arXiv:2507.14608,

    arXiv

  34. [42]

    Optimal feature transport for cross-view image geo- localization

    Yujiao Shi, Xin Yu, Liu Liu, Tong Zhang, and Hongdong Li. Optimal feature transport for cross-view image geo- localization. InProceedings of the AAAI Conference on Artificial Intelligence, pages 11990–11997, 2020. 2

    2020

  35. [43]

    Self-supervised spatial correspondence across modalities

    Ayush Shrivastava and Andrew Owens. Self-supervised spatial correspondence across modalities. InProceedings of the Computer Vision and Pattern Recognition Confer- ence, pages 6383–6393, 2025. 2

    2025

  36. [44]

    Mingxing Tan and Quoc V . Le. Efficientnet: Rethink- ing model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Ma- chine Learning (ICML), pages 6105–6114, 2019. 4

    2019

  37. [45]

    Facial expression recognition using graph-based features and artificial neural networks

    Chaiyasit Tanchotsrinon, Suphakant Phimoltares, and Saranya Maneeroj. Facial expression recognition using graph-based features and artificial neural networks. In 2011 IEEE International Conference on Imaging Sys- tems and Techniques, pages 331–334. IEEE, 2011. 2

    2011

  38. [46]

    Roboflow.https://roboflow

    Roboflow Team. Roboflow.https://roboflow. com/, 2025. Accessed 13 November 2025. 3

    2025

  39. [47]

    Yolo (v11) – real-time object detec- tion and pose estimation framework.https://pypi

    Ultralytics Team. Yolo (v11) – real-time object detec- tion and pose estimation framework.https://pypi. org/project/ultralytics- v11/, 2024. Ver- sion 11, accessed 13 Nov. 2025. 3

    2024

  40. [48]

    Automatic face recognition from skeletal remains

    Peter Tu, Rebecca Book, Xiaoming Liu, Nils Krahnsto- ever, Carl Adrian, and Phil Williams. Automatic face recognition from skeletal remains. In2007 IEEE Confer- ence on Computer Vision and Pattern Recognition, pages 1–7. IEEE, 2007. 2

    2007

  41. [49]

    Yolov8-pose: Ultralytics yolo pose esti- mation.https://github.com/ultralytics/ ultralytics, 2023

    Ultralytics. Yolov8-pose: Ultralytics yolo pose esti- mation.https://github.com/ultralytics/ ultralytics, 2023. Accessed: 2025-11-16. 3

    2023

  42. [50]

    Computer- ized craniofacial reconstruction using ct-derived implicit surface representations.Forensic science international, 159:S164–S174, 2006

    Dirk Vandermeulen, Peter Claes, Dirk Loeckx, Sven De Greef, Guy Willems, and Paul Suetens. Computer- ized craniofacial reconstruction using ct-derived implicit surface representations.Forensic science international, 159:S164–S174, 2006. 2

    2006

  43. [51]

    Multi-modality cross attention network for image and sentence matching

    Xi Wei, Tianzhu Zhang, Yan Li, Yongdong Zhang, and Feng Wu. Multi-modality cross attention network for image and sentence matching. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10941–10950, 2020. 2

    2020

  44. [52]

    Facial reconstruction–anatomical art or artistic anatomy?Journal of anatomy, 216(2): 235–250, 2010

    Caroline Wilkinson. Facial reconstruction–anatomical art or artistic anatomy?Journal of anatomy, 216(2): 235–250, 2010. 1

    2010

  45. [53]

    A joint hierarchical cross- attention graph convolutional network for multi-modal facial expression recognition.Computational Intelli- gence, 40(1):e12607, 2024

    Chujie Xu, Yong Du, Jingzi Wang, Wenjie Zheng, Tiejun Li, and Zhansheng Yuan. A joint hierarchical cross- attention graph convolutional network for multi-modal facial expression recognition.Computational Intelli- gence, 40(1):e12607, 2024. 2 10

    2024

  46. [54]

    Facial expression recognition based on graph neural network

    Xu Xu, Zhou Ruan, and Lei Yang. Facial expression recognition based on graph neural network. In2020 IEEE 5th International Conference on Image, Vision and Computing (ICIVC), pages 211–214. IEEE, 2020. 2

    2020

  47. [55]

    A semi-supervised learning approach for automated 3d cephalometric land- mark identification using computed tomography.PLoS One, 17(9):e0275114, 2022

    Hye Sun Yun, Chang Min Hyun, Seong Hyeon Baek, Sang-Hwy Lee, and Jin Keun Seo. A semi-supervised learning approach for automated 3d cephalometric land- mark identification using computed tomography.PLoS One, 17(9):e0275114, 2022. 1

    2022

  48. [56]

    Coupled information-theoretic encoding for face photo-sketch recognition

    Wei Zhang, Xiaogang Wang, and Xiaoou Tang. Coupled information-theoretic encoding for face photo-sketch recognition. InComputer Vision and Pattern Recogni- tion (CVPR), 2011 IEEE Conference on, pages 513–520. IEEE, 2011. 2, 6, 7

    2011

  49. [57]

    Facial landmark detection by deep multi- task learning

    Zhanpeng Zhang, Ping Luo, Chen Change Loy, and Xi- aoou Tang. Facial landmark detection by deep multi- task learning. InComputer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12,Proceedings, Part VI 13, pages 94–108. Springer,

    2014

  50. [58]

    Geometry-aware facial expression recognition via attentive graph convolutional networks

    Rui Zhao, Tianshan Liu, Zixun Huang, Daniel PK Lun, and Kin-Man Lam. Geometry-aware facial expression recognition via attentive graph convolutional networks. IEEE Transactions on Affective Computing, 14(2):1159– 1174, 2021

    2021

  51. [59]

    Spatial-temporal graphs plus trans- formers for geometry-guided facial expression recogni- tion.IEEE Transactions on Affective Computing, 14(4): 2751–2767, 2022

    Rui Zhao, Tianshan Liu, Zixun Huang, Daniel PK Lun, and Kin-Man Lam. Spatial-temporal graphs plus trans- formers for geometry-guided facial expression recogni- tion.IEEE Transactions on Affective Computing, 14(4): 2751–2767, 2022. 2 11

    2022

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.