Pith. sign in

REVIEW 5 major objections 7 minor 1 cited by

Self-Supervised Multi-View Representation Learning using Vision-Language Model for 3D/4D Facial Expression Recognition

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read SMILE-VLM claims that 3D/4D facial expression recognition can match or beat supervised classifiers using only self-supervised multiview learning aligned to natural-language emotion descriptions.

desk verdict A sensible combination of known SSL and VLM components for 3D/4D FER, but the 'self-supervised, label-free' claim does not survive contact with the method: the class-specific text prompts leak the ground-truth label. read the letter →

arxiv 2506.01203 v1 pith:TB4IAAE7 submitted 2025-06-01 cs.CV

classification cs.CV
keywords facialexpressionrecognitionself-supervisedlearningvision-languagemodels3D/4Dpointcloudsmulti-viewcontrastiveredundancyreductionmicro-expression
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

SMILE-VLM claims that 3D/4D facial expression recognition can be solved with self-supervised, view-invariant embeddings aligned to natural-language emotion descriptions, without categorical labels. The framework combines Barlow-Twins-style multiview decorrelation, contrastive image-to-text alignment, and cross-modal redundancy reduction, producing a shared visual-textual space where a face is matched against text prompts at inference. On BU-3DFE, BU-4DFE, BP4D-Spontaneous, and Bosphorus it reports state-of-the-art accuracy for an unsupervised method, and in several settings it edges out supervised baselines. If the claim holds, annotation cost drops dramatically for affective computing, and zero-shot recognition of new emotion descriptions becomes possible.

What carries the argument

Three jointly optimized losses carry the argument. A multiview Barlow-Twins-style cross-correlation loss averages per-view cross-correlation matrices, pushes the diagonal toward 1 and the off-diagonals toward 0, yielding view-invariant and non-redundant visual features. A vision-language contrastive InfoNCE loss aligns both each individual view and the fused multiview embedding to the text prompt of the sample. A cross-modal redundancy-minimization loss decorrelates the visual-textual cross-correlation matrix. The final objective is $L_{\mathrm{SMILE-VLM}} = \alpha L_{\mathrm{mv-bt}} + \beta L_{\mathrm{vl-align}} + \gamma L_{\mathrm{red-min}}$, and the view embeddings are combined through learned softmax attention weights.

What would settle it

An ablation that removes the text prompts entirely or replaces them with randomly mismatched emotion descriptions while keeping the multiview and redundancy losses intact: if SMILE-VLM retains most of its reported 88% to 97% accuracy under mismatched prompts, the language alignment is not the source of the gains, whereas if accuracy collapses, the prompts are supplying emotion-category information and the label-free claim fails.

Watch

Extended reading notes

Core claim

The paper's central discovery is that multiview 3D/4D facial inputs can be mapped into the same embedding space as emotion-describing text using only self-supervised losses, and that this alignment alone is enough to rival and occasionally beat supervised classifiers. The reported numbers are 89.51% on BU-3DFE Subset I, 84.01% on Subset II, 79.97% on Bosphorus, 96.57% on BU-4DFE, 88.45% on BP4D-Spontaneous recognition, 80.66% on the BP4D cross-dataset setting, and 86.61% accuracy with 0.8023 F1 on 4DME micro-expression recognition. The method extends to 4D micro-expression recognition, where subtle cues benefit from multi-angle fusion. The paper concludes that multimodal self-supervision is a scalable, annotation-efficient path for expressive facial behavior understanding.

Load-bearing premise

The load-bearing premise is that text prompts such as "a smiling person" or "a surprised face" — which are generated from the true emotion category and paired with each training sample — are auxiliary semantic cues rather than supervised labels; if a reader treats category-derived prompts as labels, the central self-supervised claim collapses.

Editorial extensions

If this is right

  • On BU-3DFE Subset II, SMILE-VLM reaches 84.01% accuracy, beating the best supervised baseline by 2.68% without using emotion labels.
  • On BU-4DFE, the model reaches 96.57% accuracy, ahead of the prior unsupervised baseline by 0.81% and of key-frame supervised methods.
  • On spontaneous BP4D, the model reaches 88.45% recognition accuracy and 80.66% cross-dataset accuracy, trailing the strongest supervised method by only 0.11% and 1.04%.
  • In the 4D micro-expression extension, multi-view fusion with emotion-sensitive prompts yields 86.61% accuracy and 0.8023 F1 on 4DME, improving over the strongest prior multi-view baseline.
  • Ablation results show that removing any one of the three losses lowers accuracy across all benchmarks, so each component is load-bearing for the claimed gains.

Reading between the lines

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

  • The same three-loss recipe should transfer to other multiview 3D tasks such as identity verification or face anti-spoofing, since view decorrelation, language anchoring, and redundancy reduction are not expression-specific.
  • A harder test than closed-set zero-shot matching would be open-vocabulary inference, where text prompts name emotions or affective states not seen in training; the shared space is built for that, but the paper only reports fixed-category matching.
  • Because prompts are randomly sampled per training step, the model's sensitivity to prompt wording could be measured by swapping synonyms; if the alignment loss is doing real semantic work, accuracy should survive paraphrase, whereas if it is only copying category anchors, it should drop.
  • The paper's ablations suggest that withholding language entirely causes the largest degradation on some datasets, which makes the label-free claim the empirical crux: the prompts are the likely source of the label-like signal.
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

5 major / 7 minor

Summary. The paper proposes SMILE-VLM, a framework for 3D/4D facial expression recognition that combines a multi-view Barlow Twins-style decorrelation loss, a vision-language contrastive alignment loss with frozen CLIP text encoders, and a cross-modal redundancy minimization loss. The authors claim that SMILE-VLM is fully self-supervised, does not use emotion labels, outperforms prior unsupervised methods, and matches or exceeds supervised baselines on BU-3DFE, BU-4DFE, BP4D-Spontaneous, and Bosphorus. The method is also extended to 4D micro-expression recognition on the 4DME dataset. The core technical ingredients are reasonable: multi-view projections at three angles, learned view fusion, and three auxiliary losses. However, the central claim of label-free self-supervision is undermined by the use of class-conditioned text prompts, and the reported performance gains over existing methods are small and lack statistical support.

Significance. If the method were genuinely self-supervised and label-free, a vision-language model that matches supervised 3D/4D FER performance would be a useful contribution, especially given the high cost of emotion annotation for 3D/4D data. The paper also ships extensive experiments on four standard benchmarks, includes ablation studies, and extends the framework to micro-expression recognition. However, the significance is substantially reduced because the method uses ground-truth class information to select text prompts during training, making it weakly supervised rather than self-supervised. The only unsupervised baseline is the author's own prior work, and the accuracy margins over strong baselines are mostly below 1.5% with no error bars or significance tests. The conceptual contribution and the empirical evidence are therefore not sufficient to support the stated central claims.

major comments (5)
  1. [Section II.A, Table 1, Eq. (6)] The claim that SMILE-VLM is self-supervised and does not use emotion labels is contradicted by the training procedure. In Section II.A, the paper states that prompts are generated from expression categories and that 'these prompts are randomly sampled at training time,' while Table 1 shows prompts grouped under emotion labels such as Happy, Sad, Surprise, Angry, Disgust, and Fear. Since the CLIP text encoder is frozen, the only channel by which class identity can enter the loss is the per-sample selection of a prompt from the true class's prompt set. Equation (6) then aligns each view and the fused embedding to that class-specific text via InfoNCE, which is mathematically equivalent to using the true category through class-name text embeddings. The statement 'The model is not provided with categorical emotion labels' is therefore not accurate; the method is weakly supervised or label-supervised. This invalidates the central 'self-supervised' claim and makes the comparison against the fully unsupervised MiFaR baseline not like-for-like.
  2. [Section IV.A-C, Tables 2-4] The state-of-the-art claim is not supported by the magnitude of the reported gains. On BU-3DFE Subset I the improvement over the best supervised method is 0.20%; on Bosphorus SMILE-VLM is 0.03% below the best supervised method; on BU-4DFE the gain over MiFaR is 0.81%; and on BP4D the gain over Yao et al. is 1.86%. The paper does not report error bars, confidence intervals, or significance tests for the 10-fold subject-independent protocol. Given that such margins are often within cross-validation noise, the reported numbers are insufficient to establish that SMILE-VLM 'surpasses existing supervised methods.'
  3. [Section II.G and Section III.D] The hyperparameters that control the loss balance are not reported. Equation (10) defines α, β, and γ as weighting coefficients, and Section II.G says 'These weights are critical in balancing the learning dynamics of the model.' However, Section III.D reports only 'an initial learning rate of 1e-4 with a weight decay factor' and does not provide values for α, β, γ, λ_mv-bt, λ_red-min, τ, batch size, number of epochs, or the number of views. Without these settings, the reported accuracies cannot be reproduced, and the sensitivity of the results to the loss weighting is unknown.
  4. [Section IV, Tables 2-4] The only unsupervised baseline used in the evaluation is MiFaR [48], which is the author's own prior method. No comparison is made with standard self-supervised learning baselines (SimCLR, MoCo, BYOL) or with other vision-language self-supervised methods applied to the same 3D/4D projection data. As a result, the claim that SMILE-VLM 'surpasses existing unsupervised methods' rests on a single baseline and does not provide a meaningful picture of the state of the art in self-supervised 3D/4D representation learning.
  5. [Section IV.E, Table 5] The extension to 4D micro-expression recognition also relies on class information. The paper states that the model is 'fine-tuned with emotion-sensitive textual prompts' where '[CLS] is replaced by the emotion category,' and Table 5 reports per-class results for Positive, Negative, Surprise, Repression, and Others. This fine-tuning uses the ground-truth micro-expression categories to construct prompts, again contradicting the label-free claim. Additionally, the only comparison is with the original dataset paper [56], so the evidence for generalizability of the proposed framework to MER is limited.
minor comments (7)
  1. [Throughout] The terms 'self-supervised,' 'unsupervised,' and 'label-free' are used interchangeably even though the method uses class-conditioned prompts; the wording should be corrected to 'weakly supervised' or 'label-efficient' if the class-conditioned prompt selection is retained.
  2. [Table 2] The notation '(↑↓)' and the parenthetical numbers (e.g., '5.01 ↑') are not explained; the reader cannot tell whether the arrow indicates that the baseline is above or below the proposed method, or whether the number is an absolute difference.
  3. [Eqs. (3)-(4)] The symbol λ appears in Eq. (3) without a subscript, while Eq. (4) uses λ_mv-bt; the notation should be made consistent.
  4. [Fig. 2 and Fig. 3] The figure captions are too vague. 'Ablation study of SMILE-VLM on multiple datasets' does not specify which configurations are compared, and 'Accuracy improvements across multiple benchmark datasets' does not define what the blue gradient encodes or what the baseline configuration is.
  5. [References] Reference [29] is cited as MoCo but the actual citation is Caron et al., SwAV; either the citation or the method name is incorrect. Also, the paper refers to 'Y ao et al.' and 'Mul timodality' with spacing typos.
  6. [Section IV.C] Danelakis et al. is described as 'fully supervised,' but retrieval-based facial expression retrieval may not be trained in a classification setting; this description should be clarified or the comparison should be justified.
  7. [Section I.B] The claim of 'zero-shot expression recognition' is misleading because the prompts used at inference are the same as those used during training; the evaluation is closer to a nearest-class-mean classification in the CLIP embedding space rather than a zero-shot transfer to unseen categories.

Circularity Check

3 steps flagged · score 6.0 of 10

Class-specific prompt pairing in Eq. (6) leaks ground-truth expression labels into training, invalidating the central self-supervised, label-free claim; the only unsupervised comparison is the author's own prior work.

  1. self definitional [Section II-A (Problem Formulation), Section II-D Eq. (6), Table 1]
    "These prompts are generated using the GPT language model to map expression categories to semantically rich natural language descriptions. ... Each expression category is associated with a set of semantically rich prompts ... These descriptions are randomly sampled at training time ... The model is not provided with categorical emotion labels. Instead, it receives semantic cues in the form of descriptive text templates that do not require manual annotation."

    The prompt set is indexed by the six expression categories (Table 1), so selecting a prompt t for a given training sample requires knowing that sample's ground-truth category. In Eq. (6), the InfoNCE loss uses zt_b, the text embedding of that sample's associated prompt, as the positive anchor; the frozen CLIP text encoder converts the category into a fixed class prototype. Optimizing this loss moves each visual embedding toward the prototype of its true class, which is mathematically a supervised contrastive classifier over the six emotion categories. The claim that the model is 'fully self-supervised' and label-free is therefore true only by renaming categorical labels as 'semantic cues'; the class label is exactly the variable that selects the training text.

  2. fitted input called prediction [Section III-D (Implementation Details), Section IV (Results)]
    "The standard classification is done in a zero-shot setting, where the fused embedding of a query sample is matched directly against the encoded textual prompts. Finally, the class with the highest similarity score is selected as the predicted expression."

    The inference-time prompts are the same per-class prompt sets used during training. Because Eq. (6) explicitly aligned each class's visual embeddings to those prompts, the 'zero-shot' evaluation is nearest-prototype matching to the very classes whose labels selected the training prompts. The predicted expression is, by construction, the class that was used to construct the positive text for that sample during training. This is the fitted-input-called-prediction pattern: the class identity is an input to the prompt-generation step, and the same class identity is then reported as an output prediction.

1 more flagged steps
  1. self citation load bearing [Section IV, first paragraph]
    "To the best of our knowledge, only one prior method has explored 3D/4D facial expression recognition in a fully unsupervised setting [48]. We include this method in our evaluation to establish a direct baseline for self-supervised learning in this domain."

    Reference [48] is the author's own prior work (Behzad and Zhao, FG 2021), and the paper uses it as the sole 'unsupervised' baseline. The claim of surpassing existing unsupervised methods is established only relative to this self-citation, and the premise that no other fully unsupervised method exists is also asserted via the same citation. This makes the unsupervised-SOTA comparison self-referential, although the more severe circularity is the label leakage through class-specific prompt selection in Eq. (6).

full rationale

The core derivation chain of SMILE-VLM is not algebraically circular, but the paper's headline claim — that the model is fully self-supervised and trained without emotion labels — is circular by definition. Section III.C states 'Each expression category is associated with a set of semantically rich prompts,' and Section II.A says the prompts 'map expression categories' to text. Therefore choosing the positive text t_b in the InfoNCE loss of Eq. (6) requires knowing the ground-truth category of sample b. The frozen CLIP text encoder converts that category into a fixed class prototype, so the training objective is supervised contrastive alignment to the true class. The subsequent 'zero-shot' inference matches against the same class-specific prompts, so the predicted class is the class whose prompt supervised the embedding; this is the fitted-input-called-prediction pattern. The unsupervised state-of-the-art claim is also fragile because the only unsupervised baseline [48] is the author's own prior work, and the premise that it is the only such method is a self-citation. The benchmark accuracies themselves are externally measured numbers, and the visual encoder still must learn to project facial geometry, so the circularity is concentrated in the self-supervised/label-free interpretation rather than in every experimental result. This warrants a score of 6: the central claim partially reduces to label supervision by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on tuning choices and assumptions, the most consequential being that class-conditioned prompt selection is not supervision. None of the losses or modules are new; they are combinations of Barlow Twins, InfoNCE, and redundancy reduction. No new entities are introduced.

free parameters (6)
  • lambda_mvbt
    Weight for the multiview Barlow Twins loss in Eq. (4); no value is reported.
  • lambda_redmin
    Weight for the cross-modal redundancy loss in Eq. (9); no value is reported.
  • alpha_beta_gamma
    Weights in the joint loss Eq. (10); their values are discussed qualitatively but never specified.
  • tau
    Temperature in the InfoNCE loss Eq. (6); no value is reported.
  • learning_rate_weight_decay
    Optimizer settings are mentioned in Section III.D but exact values are omitted.
  • number_of_views_and_angles = 3 views at 0, -30, +30 degrees
    The choice of three angles is a modeling decision; no experiments justify this selection.
assumptions (5)
  • domain assumption CLIP's pretrained text and image encoders provide a semantic space that transfers to 3D/4D facial data projected to 2D.
    The entire alignment relies on CLIP embeddings; invoked throughout Section II as the baseline model.
  • domain assumption GPT-generated prompts for the six basic emotions are semantically appropriate for supervising expression recognition.
    Used in Section II.C to build the prompt set T, assuming these prompts map cleanly to visual expressions.
  • ad hoc to paper Pairing a training sample with the prompt from its true expression category is not label supervision.
    Section II.A asserts this to justify the 'self-supervised' label; it is contradicted by the method, which selects prompts using ground-truth class information.
  • domain assumption 2D projections at frontal and plus or minus 30 degrees preserve expression-discriminative information of 3D/4D scans.
    Section III.B preprocesses all data this way without validating that these views capture the relevant facial deformations.
  • domain assumption Rank pooling compresses 4D sequences without losing temporal expression dynamics.
    Section III.B applies rank pooling to 4D data, assuming it retains the subtle motion cues needed for expression recognition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Supervised Multi-View Representation Learning using Vision-Language Model for 3D/4D Facial Expression Recognition." pith.science (2026). https://pith.science/paper/TB4IAAE7

@misc{pith2026250601203,
  author       = {Pith},
  title        = {Pith review of: Self-Supervised Multi-View Representation Learning using Vision-Language Model for 3D/4D Facial Expression Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TB4IAAE7}},
  note         = {Machine review of arXiv:2506.01203}
}
read the original abstract

Facial expression recognition (FER) is a fundamental task in affective computing with applications in human-computer interaction, mental health analysis, and behavioral understanding. In this paper, we propose SMILE-VLM, a self-supervised vision-language model for 3D/4D FER that unifies multiview visual representation learning with natural language supervision. SMILE-VLM learns robust, semantically aligned, and view-invariant embeddings by proposing three core components: multiview decorrelation via a Barlow Twins-style loss, vision-language contrastive alignment, and cross-modal redundancy minimization. Our framework achieves the state-of-the-art performance on multiple benchmarks. We further extend SMILE-VLM to the task of 4D micro-expression recognition (MER) to recognize the subtle affective cues. The extensive results demonstrate that SMILE-VLM not only surpasses existing unsupervised methods but also matches or exceeds supervised baselines, offering a scalable and annotation-efficient solution for expressive facial behavior understanding.

Figures

Figures reproduced from arXiv: 2506.01203 by the authors.

Figure 1
Figure 1. FIGURE 1 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. FIGURE 2 [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. FIGURE 3 [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

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. Facial Emotion Learning with Text-Guided Multiview Fusion via Vision-Language Model for 3D/4D Facial Expression Recognition

    cs.CV 2025-07 conditional novelty 4.0 of 10

    FACET-VLM combines CLIP text prompts with multiview fusion to claim state-of-the-art 3D/4D facial expression recognition, but lacks released artifacts and a vision-only baseline.

Reference graph

Works this paper leans on

54 extracted references · 54 canonical work pages · cited by 1 Pith paper

  1. [48]

    Behzad and G

    M. Behzad and G. Zhao, ‘‘Self-supervised learning via multi-view facial rendezvous for 3d/4d affect recognition,’’ in 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pp. 1– 5, IEEE, 2021

  2. [56]

    X. Li, S. Cheng, Y . Li, M. Behzad, J. Shen, S. Zafeiriou, M. Pantic, and G. Zhao, ‘‘4dme: A spontaneous 4d micro-expression dataset with multimodalities,’’ IEEE Transactions on Affective Computing , vol. 14, no. 4, pp. 3031–3047, 2022. 10 VOLUME 11, 2023

  3. [1]

    Bordes, R

    F. Bordes, R. Y . Pang, A. Ajay, A. C. Li, A. Bardes, S. Petryk, O. Mañas, Z. Lin, A. Mahmoud, B. Jayaraman, M. Ibrahim, M. Hall, Y . Xiong, J. Lebensold, C. Ross, S. Jayakumar, C. Guo, D. Bouchacourt, H. Al-Tahan, K. Padthe, V . Sharma, H. Xu, X. E. Tan, M. Richards, S. Lavoie, P . Astolfi, R. A. Hemmat, J. Chen, K. Tirumala, R. Assouel, M. Moayeri, A. T...

  4. [2]

    Minaee, T

    S. Minaee, T. Mikolov, N. Nikzad, M. Chenaghlu, R. Socher, X. Amatriain, and J. Gao, ‘‘Large language models: A survey,’’ 2024

  5. [3]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P . Mishkin, J. Clark, G. Krueger, and I. Sutskever, ‘‘Learning transferable visual models from natural language supervision,’’ 2021

  6. [4]

    Y . Zhai, H. Bai, Z. Lin, J. Pan, S. Tong, Y . Zhou, A. Suhr, S. Xie, Y . Le- Cun, Y . Ma, and S. Levine, ‘‘Fine-tuning large vision-language models as decision-making agents via reinforcement learning,’’ 2024

  7. [5]

    M. K. Chowdary, T. N. Nguyen, and D. J. Hemanth, ‘‘Deep learning-based facial emotion recognition for human–computer interaction applications,’’ Neural Computing and Applications , vol. 35, no. 32, pp. 23311–23328, 2023

  8. [6]

    N. M. Foteinopoulou and I. Patras, ‘‘Learning from label relationships in human affect,’’ in Proceedings of the 30th ACM International Conference on Multimedia, vol. 33 of MM ’22, ACM, 2022

Show all 54 references
  1. [7]

    Y adegaridehkordi, N

    E. Y adegaridehkordi, N. F. B. M. Noor, M. N. B. Ayub, H. B. Affal, and N. B. Hussin, ‘‘Affective computing in education: A systematic review and future research,’’ Computers & Education, vol. 142, 2019

  2. [8]

    C. A. Corneanu, M. O. Simón, J. F. Cohn, and S. E. Guerrero, ‘‘Survey on rgb, 3d, thermal, and multimodal approaches for facial expression recognition: History, trends, and affect-related applications,’’IEEE TPAMI, vol. 38, no. 8, pp. 1548–1568, 2016

  3. [9]

    Ekman and W

    P . Ekman and W. V . Friesen, ‘‘Constants across cultures in the face and emotion.,’’ Journal of personality and social psychology , vol. 17, no. 2, p. 124, 1971

  4. [10]

    Y .-J. Liu, B. Wang, L. Gao, J. Zhao, R. Yi, M. Y u, Z. Pan, and X. Gu, ‘‘4d facial analysis: A survey of datasets, algorithms and applications,’’ Computers & Graphics, vol. 115, pp. 423–445, 2023

  5. [11]

    Li et al., ‘‘3d facial expression recognition via multiple kernel learning of multi-scale local normal patterns,’’ in ICPR, 2012

    H. Li et al., ‘‘3d facial expression recognition via multiple kernel learning of multi-scale local normal patterns,’’ in ICPR, 2012

  6. [12]

    Li, Tao Jia, and H

    X. Li, Tao Jia, and H. Zhang, ‘‘Expression-insensitive 3d face recognition using sparse representation,’’ in CVPR, pp. 2575–2582, 2009

  7. [13]

    H. Li, H. Ding, D. Huang, Y . Wang, X. Zhao, J.-M. Morvan, and L. Chen, ‘‘An efficient multimodal 2d+ 3d feature-based approach to automatic facial expression recognition,’’ CVIU, pp. 83–92, 2015

  8. [14]

    Mpiperis, S

    I. Mpiperis, S. Malassiotis, and M. G. Strintzis, ‘‘Bilinear models for 3-d VOLUME 11, 2023 9 Author et al.: Preparation of Papers for IEEE TRANSACTIONS and JOURNALS face and facial expression recognition,’’IEEE Transactions on Information F orensics and Security, vol. 3, no. ...

  9. [15]

    X. Zhao, D. Huang, E. Dellandréa, and L. Chen, ‘‘Automatic 3d facial expression recognition based on a bayesian belief net and a statistical facial feature model,’’ in ICPR, pp. 3724–3727, 2010

  10. [16]

    Samir et al., ‘‘An intrinsic framework for analysis of facial surfaces,’’ IJCV, 2009

    C. Samir et al., ‘‘An intrinsic framework for analysis of facial surfaces,’’ IJCV, 2009

  11. [17]

    Maalej, B

    A. Maalej, B. B. Amor, M. Daoudi, A. Srivastava, and S. Berretti, ‘‘Shape analysis of local facial patches for 3d facial expression recognition,’’ Pattern Recognition, vol. 44, no. 8, pp. 1581–1589, 2011

  12. [18]

    H. Li, J. Sun, Z. Xu, and L. Chen, ‘‘Multimodal 2d+3d facial expression recognition with deep fusion convolutional neural network,’’ IEEE Trans- actions on Multimedia, vol. 19, 2017

  13. [19]

    O. K. Oyedotun, G. Demisse, A. E. R. Shabayek, D. Aouada, and B. Ot- tersten, ‘‘Facial expression recognition via joint deep learning of rgb-depth map latent representations,’’ in ICCVW, 2017

  14. [20]

    Zhang, L

    X. Zhang, L. Yin, J. F. Cohn, S. Canavan, M. Reale, A. Horowitz, and P . Liu, ‘‘A high-resolution spontaneous 3d dynamic facial expression database,’’ in FG, 2013

  15. [21]

    Y . Sun, X. Chen, M. Rosato, and L. Yin, ‘‘Tracking vertex flow and model adaptation for three-dimensional spatiotemporal face analysis,’’ IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, vol. 40, no. 3, pp. 461–474, 2010

  16. [22]

    Sandbach, S

    G. Sandbach, S. Zafeiriou, M. Pantic, and D. Rueckert, ‘‘Recognition of 3d facial expression dynamics,’’ Image and Vision Computing, 2012

  17. [23]

    B. B. Amor, H. Drira, S. Berretti, M. Daoudi, and A. Srivastava, ‘‘4-d facial expression recognition by learning geometric deformations,’’ IEEE transactions on cybernetics, vol. 44, 2014

  18. [24]

    T. Fang, X. Zhao, O. Ocegueda, S. K. Shah, and I. A. Kakadiaris, ‘‘3d/4d facial expression analysis: An advanced annotated face model approach,’’ Image and vision Computing , vol. 30, no. 10, 2012

  19. [25]

    T. Fang, X. Zhao, S. K. Shah, and I. A. Kakadiaris, ‘‘4d facial expression recognition,’’ in ICCVW, 2011

  20. [26]

    Reale, X

    M. Reale, X. Zhang, and L. Yin, ‘‘Nebula feature: A space-time feature for posed and spontaneous 4d facial behavior analysis,’’ in FG, 2013

  21. [27]

    W. Li, D. Huang, H. Li, and Y . Wang, ‘‘Automatic 4d facial expression recognition using dynamic geometrical image network,’’ in FG, 2018

  22. [28]

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, ‘‘A simple framework for contrastive learning of visual representations,’’ in ICML, 2020

  23. [29]

    Caron, I

    M. Caron, I. Misra, J. Mairal, P . Goyal, P . Bojanowski, and A. Joulin, ‘‘Un- supervised learning of visual features by contrasting cluster assignments,’’ in NeurIPS, 2020

  24. [30]

    Grill, F

    J.-B. Grill, F. Strub, F. Altché, C. Tallec, P . Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Gheshlaghi Azar, B. Piot, k. kavukcuoglu, R. Munos, and M. V alko, ‘‘Bootstrap your own latent - a new approach to self-supervised learning,’’ in NeurIPS, 2020

  25. [31]

    Zbontar, L

    J. Zbontar, L. Jing, I. Misra, Y . LeCun, and S. Deny, ‘‘Barlow twins: Self- supervised learning via redundancy reduction,’’ in ICML, 2021

  26. [32]

    R. Gao, F. Y ang, W. Y ang, and Q. Liao, ‘‘Margin loss: Making faces more separable,’’ IEEE Signal Processing Letters , vol. 25, no. 2, pp. 308–312, 2018

  27. [33]

    Y . Tian, J. Cheng, Y . Li, and S. Wang, ‘‘Secondary information aware facial expression recognition,’’ IEEE Signal Processing Letters , vol. 26, no. 12, pp. 1753–1757, 2019

  28. [34]

    Jiang, B

    P . Jiang, B. Wan, Q. Wang, and J. Wu, ‘‘Fast and efficient facial expression recognition using a gabor convolutional network,’’IEEE Signal Processing Letters, vol. 27, pp. 1954–1958, 2020

  29. [35]

    M. Hu, Q. Chu, X. Wang, L. He, and F. Ren, ‘‘A two-stage spatiotemporal attention convolution network for continuous dimensional emotion recog- nition from facial video,’’IEEE Signal Processing Letters, vol. 28, pp. 698– 702, 2021

  30. [36]

    Li, J.-M

    H. Li, J.-M. Morvan, and L. Chen, ‘‘3d facial expression recognition based on histograms of surface differential quantities,’’ in International Conference on Advanced Concepts for Intelligent Vision Systems, pp. 483– 494, Springer, 2011

  31. [38]

    Y ang, D

    X. Y ang, D. Huang, Y . Wang, and L. Chen, ‘‘Automatic 3d facial expression recognition using geometric scattering representation,’’ in IEEE FG, 2015

  32. [39]

    van den Oord, Y

    A. van den Oord, Y . Li, and O. Vinyals, ‘‘Representation learning with contrastive predictive coding,’’ 2019

  33. [40]

    Savran, N

    A. Savran, N. Alyüz, H. Dibeklioğlu, O. Çeliktutan, B. Gökberk, B. Sankur, and L. Akarun, ‘‘Bosphorus database for 3d face analysis,’’ in European workshop on biometrics and identity management , 2008

  34. [41]

    L. Yin, X. Wei, Y . Sun, J. Wang, and M. J. Rosato, ‘‘A 3d facial expression database for facial behavior research,’’ in FG, 2006

  35. [42]

    Zhang, L

    X. Zhang, L. Yin, J. F. Cohn, S. Canavan, M. Reale, A. Horowitz, P . Liu, and J. M. Girard, ‘‘Bp4d-spontaneous: a high-resolution spontaneous 3d dynamic facial expression database,’’Image and Vision Computing, vol. 32, no. 10, pp. 692 – 706, 2014

  36. [44]

    Behzad, N

    M. Behzad, N. V o, X. Li, and G. Zhao, ‘‘Landmarks-assisted collaborative deep framework for automatic 4d facial expression recognition,’’ in FG, 2020

  37. [45]

    Behzad, N

    M. Behzad, N. V o, X. Li, and G. Zhao, ‘‘Towards reading beyond faces for sparsity-aware 3d/4d affect recognition,’’ Neurocomputing, 2021

  38. [46]

    Behzad, X

    M. Behzad, X. Li, and G. Zhao, ‘‘Disentangling 3d/4d facial affect recogni- tion with faster multi-view transformer,’’ IEEE Signal Processing Letters , vol. 28, pp. 1913–1917, 2021

  39. [47]

    Bilen, B

    H. Bilen, B. Fernando, E. Gavves, and A. V edaldi, ‘‘Action recognition with dynamic image networks,’’ IEEE TPAMI, 2017

  40. [49]

    Xue et al., ‘‘Automatic 4d facial expression recognition using dct features,’’ in WACV, 2015

    M. Xue et al., ‘‘Automatic 4d facial expression recognition using dct features,’’ in WACV, 2015

  41. [50]

    Q. Zhen, D. Huang, Y . Wang, and L. Chen, ‘‘Muscular movement model- based automatic 3d/4d facial expression recognition,’’ IEEE Transactions on Multimedia, vol. 18, no. 7, pp. 1438–1450, 2016

  42. [51]

    Y . Y ao, D. Huang, X. Y ang, Y . Wang, and L. Chen, ‘‘Texture and geome- try scattering representation-based facial expression recognition in 2d+3d videos,’’ ACM Trans. Mult. Comput. Commun. Appl. , 2018

  43. [52]

    Bejaoui, H

    H. Bejaoui, H. Ghazouani, and W. Barhoumi, ‘‘Sparse coding-based rep- resentation of lbp difference for 3d/4d facial expression recognition,’’ Multimedia Tools and Applications, 2019

  44. [53]

    Behzad, N

    M. Behzad, N. V o, X. Li, and G. Zhao, ‘‘Automatic 4d facial expression recognition via collaborative cross-domain dynamic image network,’’ in BMVC, British Machine Vision Association Press, 2019

  45. [54]

    Danelakis, T

    A. Danelakis, T. Theoharis, I. Pratikakis, and P . Perakis, ‘‘An effective methodology for dynamic 3d facial expression retrieval,’’ Pattern Recog- nition, vol. 52, 2016

  46. [55]

    Q. Zhen, D. Huang, H. Drira, B. B. Amor, Y . Wang, and M. Daoudi, ‘‘Magnifying subtle facial motions for effective 4d expression recogni- tion,’’ IEEE Transactions on Affective Computing , 2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.