Pith. sign in

REVIEW 4 major objections 5 minor 36 references

Strength-Parity Ensembling with Parameter-Isolated Experts for Multi-Task Affect Recognition

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A new admission rule determines when adding a model to an ensemble actually helps, and a low-rank parameter-isolation method supplies the members that pass it.

desk verdict The real contribution is the 0.98-vs-0.91 correlation measurement showing LoRA parameter isolation preserves ensemble diversity; the headline gain is within noise and the admission rule is not validated out-of-sample. read the letter →

arxiv 2607.16290 v2 pith:YNQJINFF submitted 2026-07-12 cs.CV

Tung Hung Bui , Hong Hai Nguyen , Van Thong Huynh This is my paper
classification cs.CV
keywords multi-taskaffectrecognitionensemblingstrength-parityruleparameterisolationlow-rankadaptationABAWchallengeexpressionensemblediversity
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 tackles the practical question of which member is worth adding to an already strong ensemble for multi-task affect recognition. It proposes a strength-parity rule: a new model helps only if it is both decorrelated from the current ensemble and about as accurate as its members. The paper then shows that full fine-tuning of a shared backbone re-converges to nearly identical predictions, while confining each adaptation to a disjoint low-rank subspace yields decorrelated near-peer experts. Applying this rule and mechanism on the ABAW multi-task track raises the validation score from 1.6669 to 1.6949, with the strongest configuration reaching 1.7259.

What carries the argument

The strength-parity rule is an admission criterion derived from the classic accuracy–diversity decomposition of ensemble error: an added member helps only when it lowers the mean pairwise correlation without raising the error level. The mechanism that satisfies it is parameter-isolation: freezing a shared backbone and adapting it to different affect corpora through disjoint low-rank residual subspaces (rank 16, scaling 32), which keeps adaptations decorrelated while preserving individual strength.

What would settle it

Run the same parameter-isolation ensemble on a held-out test split (not used for member admission) and compare the gain from an admitted AffectNet-LoRA expert against a re-seeded head; if the expert's advantage disappears or reverses, the rule's predictive power on unseen data is not supported.

Watch

Extended reading notes

Core claim

The paper establishes that on a shared backbone, unconstrained fine-tuning curricula re-converge to a prediction correlation of 0.98, adding no ensemble value, whereas parameter-isolated low-rank experts (LoRA-style) on the same backbone stay decorrelated at 0.91 while remaining near-peers. The strongest such expert, adapted on AffectNet, reaches expression macro-F1 0.4762 and, when added to the expression ensemble, raises the overall multi-task score from 1.6669 to 1.6949. The paper frames this as supplying the mechanism required by its strength-parity rule: a way to produce decorrelated near-peers from a single backbone.

Load-bearing premise

The central claim rests on the assumption that the strength-parity rule, evaluated on validation metrics, reliably predicts which members improve the ensemble on held-out data, even though the rule's MSE-based derivation is not directly matched to macro-F1 or CCC and the validation split is used both for admission and reporting.

Editorial extensions

If this is right

  • If correct, the strength-parity rule provides a pre-ensemble admission test: measure a candidate's prediction correlation with the current ensemble and its own task metric; only admit members that are both decorrelated and near-peers.
  • Parameter-isolation becomes a practical tool for building diverse ensembles from a single pretrained model, avoiding the cost of multiple backbones or full fine-tuning runs.
  • The observed plateau after a few decorrelated near-peers suggests that further ensemble gains will require genuinely different representations rather than more averaging.
  • The finding that full fine-tuning re-converges to 0.98 correlation highlights a fundamental limitation of diversity-by-seed or diversity-by-curriculum on a shared backbone.
  • The reported gains (e.g., adding the AffectNet-LoRA expert) are positive-within-noise at 50 validation videos; the only member gain that clears significance is pooling the shared-latent V-A byproduct.

Reading between the lines

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

  • The strength-parity rule, though derived from MSE, is applied to macro-F1 and CCC; a reader might infer that the rule is being used heuristically rather than as an exact predictor, and a direct test on these metrics would clarify its scope.
  • The validation split is used both for member admission and for reporting gains, which risks optimistic selection bias; a cross-validated or held-out evaluation would test whether the rule transfers.
  • The paper's own paired bootstrap labels most gains as positive-within-noise, suggesting that the practical benefit of parameter-isolated experts may be real but smaller than the point estimates imply.
  • This mechanism could generalize beyond affect recognition: any task where a shared backbone is adapted to multiple domains may benefit from disjoint low-rank subspaces as a diversity source.
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 / 5 minor

Summary. The paper addresses the question of when adding one more member helps an already strong ensemble for the ABAW multi-task affect recognition benchmark. It proposes a 'strength-parity' admission rule, based on the Krogh–Vedelsby accuracy–diversity decomposition: a new member should be both decorrelated from the current ensemble and of comparable individual accuracy. The paper reports that full fine-tuning of a single backbone re-converges to a prediction correlation of 0.98, whereas parameter-isolated LoRA experts on the same backbone remain at 0.91 while staying near-peers. The strongest LoRA expert (AffectNet-adapted) raises the expression macro-F1 from 0.4698 to 0.4978 and the overall PMTL from 1.6669 to 1.6949; pooling the shared-latent members' valence–arousal byproduct further raises the score to 1.7259. All differences are accompanied by paired video-level bootstrap p-values, and source code is provided.

Significance. If the central claim holds, the paper offers a practical, low-cost mechanism for creating diverse ensemble members from a single backbone—parameter-isolated low-rank experts—and a simple admission test for deciding whether an expert is worth adding. The paper is commendably transparent: it reports noise-level gains as 'positive-within-noise,' separates dev-fit calibrated numbers from uncalibrated ones, and releases code. The correlation analysis (0.98 vs. 0.91 vs. 0.75) is an interesting empirical observation. However, the load-bearing empirical result—that the LoRA expert reliably improves the expression ensemble—is statistically indistinguishable from zero by the authors' own bootstrap, and the admission rule is applied and evaluated on the same validation split. The VA-pooling gain is significant, but it is a byproduct member rather than the paper's core mechanism. Because of these gaps, the contribution is promising but not yet fully secured.

major comments (4)
  1. [§II.B, Eq. (2)] The accuracy–diversity decomposition in Eq. (2) is derived for the mean squared error of averaged real-valued predictions under an equal-variance assumption. The paper applies this rule as a universal admission test for macro-F1 (expression, AU) and CCC (valence–arousal) ensembles without stating the transfer. Macro-F1 of averaged posteriors is not an MSE objective, and CCC is a concordance measure, so the conditions 'decorrelated and near-peer' being necessary/sufficient is not justified for these metrics. Please state the assumption explicitly and provide either a metric-appropriate derivation or an empirical validation that the rule predicts gains for each task metric.
  2. [§III.A–B, Table I] The primary result of the paper—that the AffectNet-LoRA expert improves the expression ensemble from 0.4698 to 0.4978 macro-F1 and the total to 1.6949—is described by the authors themselves as 'positive-within-noise' at the 50-video validation split. Yet the conclusion calls this 'our robust primary result.' A gain that is statistically indistinguishable from zero under the paper's own bootstrap cannot carry the weight of the central claim. The only significant member-level gain is the VA-pooling gain (p=0.005), which is a byproduct of the shared-latent heads, not of the parameter-isolation mechanism. Please rephrase the claims to match the evidence, or provide a larger validation sample/cross-validation to demonstrate that the LoRA-expert gain is reproducible.
  3. [§II.E and §III.A–B] Member admission and gain reporting are performed on the same 50-video validation split. The AffectNet-LoRA expert was admitted because it appeared strong and decorrelated on validation, and the reported improvement is measured on the same videos. This creates selection optimism: the gain may reflect the expert that happened to look best on this small split rather than a genuine effect of parameter isolation. The strength-parity rule is therefore not validated out-of-sample. Please add a nested or repeated split-cross-validation design, or use a held-out partition for at least the final evaluation of the admission rule.
  4. [§III.C, Table III and Fig. 2] The correlation values that support the mechanism—0.91 for LoRA experts vs. 0.98 for full fine-tuning on the same backbone, and 0.75 for AffectNet-LoRA vs. a different backbone—are reported as point estimates without confidence intervals. These correlations are also computed on the same validation set used for member selection. Since the strength-parity rule hinges on the difference between these correlations, please report bootstrap intervals or an explicit test of the difference, and clarify whether the correlations are stable across the 50 videos.
minor comments (5)
  1. [Abstract] Grammar: 'Source code are available' should be 'Source code is available.'
  2. [§II.B, Eq. (2)] The notation ε̄² and ρ̄ is introduced without a precise definition of the average over members and over the data distribution. A one-sentence definition would improve readability.
  3. [§II.B and throughout] The term 'near-peer' is never quantified. Please specify the tolerance used when admitting a member (e.g., within a certain F1 margin of the current ensemble members).
  4. [Table I] The row 'pool SAL members’ VA into VA ensemble' is the only gain reported as significant (p=0.005). Consider adding the p-value to the table for consistency with the text.
  5. [§III.B] The phrase 'robust primary result' for the 1.6949 figure is contradicted by the 'positive-within-noise' characterization on the same page. Please align the wording with the bootstrap outcome.

Circularity Check

0 steps flagged · score 1.0 of 10

No definitional circularity; strength-parity rule is an external decomposition and LoRA diversity is imported; main caveats are validation reuse and noise, which the paper discloses.

full rationale

The derivation chain is not circular in the definitional sense. The strength-parity rule is an application of the classic accuracy–diversity decomposition of Eq. (2), cited to Krogh and Vedelsby [15], and the two admission conditions (decorrelation and near-peer strength) are exactly the two terms of that external decomposition rather than being defined from the paper's own outputs. The parameter-isolation mechanism is likewise imported from external LoRA and LoRA-ensemble literature [18], [19]. The base system from companion work [1] is transparently adopted as an input, with the paper explicitly stating that its design and justification are not claimed here; while this is a self-citation by overlapping authors, the paper's incremental claims are measured against that adopted system rather than being derived from it. The main limitations are statistical rather than circular: member admission, hyperparameter selection, and per-AU threshold calibration all use the same 50-video validation split on which the gains are reported, and the paper itself discloses that the primary expression gain is 'positive-within-noise' and that the calibration gain is dev-fit optimistic. These are legitimate concerns about selection optimism and generalization, but they do not reduce any reported result to an input by construction, and no quoted equation makes a fitted statistic equal to the claimed gain. Therefore, no significant circularity is present.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The ledger is small and honest: the paper introduces no new entities or mediators; the shared affect-latent is inherited from companion work [1]. Load-bearing free choices are the LoRA geometry (rank/scaling), the dev-fit per-AU thresholds, and the unquantified admission boundary of the rule itself. The main hidden assumptions are the transfer of the MSE-based decomposition to macro-F1/CCC metrics and the representativeness of the 50-video validation split used for both admission and evaluation.

free parameters (3)
  • LoRA rank r and scaling alpha = r=16, alpha=32
    Chosen per §II.D/§III.A; jointly control expert strength-vs-diversity and are not swept or justified. The 0.91 decorrelation and 0.4762 F1 results depend on them.
  • per-AU decision thresholds = fit per unit on validation to maximize F1
    §II.E/§III.A; explicitly fitted on the validation split; acknowledged optimistic; recovers only about a third of apparent gain under cross-fit.
  • near-peer / decorrelated admission boundary = not quantified
    The strength-parity test has no stated thresholds for 'near-peer' or 'decorrelated'; membership is judged qualitatively (§II.B), so the rule is not mechanically applied.
assumptions (5)
  • standard math Eq. 2 accuracy-diversity decomposition (Krogh-Vedelsby) with equal member variance and pairwise-correlation-only error structure
    Cited to [15] in §II.B; correct under i.i.d. equal-variance errors, but the paper does not verify equal variance across members whose F1s range 0.4416–0.4762.
  • domain assumption The MSE-based decomposition governs ensemble behavior of macro-F1 and CCC metrics
    §II.B applies Eq. (2) to expression (macro-F1) and VA (CCC) ensembles without derivation; averaging posteriors then thresholding is not MSE-optimal, so the 'only if' admission rule is heuristic for these metrics.
  • ad hoc to paper Companion-work base system (Ref [1]) achieves the claimed 1.6669 and its design is sound
    §II.A adopts the SAL two-backbone system and its score wholesale; the incremental claim 1.6669→1.6949 is measured against a number from a co-authored companion paper that is not audited here.
  • domain assumption Validation-split statistics (correlations, macro-F1, bootstrap) are representative of held-out test behavior
    All admission, early stopping, and calibration use train+val (§III.A); whether these 50-video validation measurements transfer to the test set is not tested.
  • domain assumption AffectNet/EmotioNet pretraining does not violate the 'no Aff-Wild2 pretraining' rule and transfers positively
    The curriculum and experts rely on AffectNet (§II.C/D). The rule is respected as stated; transfer is empirically assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Strength-Parity Ensembling with Parameter-Isolated Experts for Multi-Task Affect Recognition." pith.science (2026). https://pith.science/paper/YNQJINFF

@misc{pith2026260716290,
  author       = {Pith},
  title        = {Pith review of: Strength-Parity Ensembling with Parameter-Isolated Experts for Multi-Task Affect Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YNQJINFF}},
  note         = {Machine review of arXiv:2607.16290}
}
read the original abstract

Leading entries on the multi-task track of the 11th ABAW challenge rely on heavy ensembling, yet which member is worth adding to an already strong ensemble is rarely made explicit. We study this question for joint valence-arousal estimation, 8-way expression recognition, and 12-way action-unit detection from a single unconstrained face, under partial, long-tailed labels and a rule that forbids pretraining on Aff-Wild2. Building on a shared affect-latent that marginalizes the missing labels across two affect-supervised backbones, we propose a strength-parity rule: an added member lowers the ensemble error only when it is both decorrelated from the current members and a near-peer of them in individual accuracy. The rule exposes a concrete obstacle, as on a single backbone re-seeding and even distinct fine-tuning curricula re-converge to a prediction correlation of 0.98 and add no diversity. Parameter-isolation removes it: confining each adaptation to a disjoint low-rank subspace of a shared backbone yields experts that stay decorrelated at 0.91 while remaining near-peers, the strongest of them an AffectNet-adapted expert. The resulting system raises the overall validation score to 1.6949, against the organizers ConvNeXt-with-MixAugment baseline of 0.45; with per-AU calibration and by pooling the shared-latent heads valence-arousal byproduct as a further near-peer, the strongest configuration reaches 1.7259. Source code are available at https://github.com/cprl-team/MTL-ABAW-11th.

Figures

Figures reproduced from arXiv: 2607.16290 by the authors.

Figure 1
Figure 1. System overview. Frozen affect-supervised backbones and the added members (highlighted: parameter-isolated LoRA experts and a curriculum [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Prediction correlation between expression members; lower correlation [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 8 linked inside Pith

  1. [1]

    A shared latent for partially-labeled multi-task facial affect recognition,

    H. H. Nguyen, S. P. Van, S.-H. Kim, and V .-T. Huynh, “A shared latent for partially-labeled multi-task facial affect recognition,” 2026. [Online]. Available: https://arxiv.org/abs/2607.16285

    arXiv 2026

  2. [2]

    7th abaw competition: Multi-task learning and compound expression recognition,

    D. Kollias, S. Zafeiriou, I. Kotsia, A. Dhall, S. Ghosh, C. Shao, and G. Hu, “7th abaw competition: Multi-task learning and compound expression recognition,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 31–45

    2024

  3. [3]

    From affect to complex behavior: Advancing multimodal human-centered ai at the 10th abaw workshop & competition,

    D. Kollias, P. Tzirakis, A. Cowen, S. Zafeiriou, I. Kotsia, E. Granger, M. Pedersoli, S. Bacon, J. Madsen, S. Belharbi, M. H. Aslam, C. Shao, and G. Hu, “From affect to complex behavior: Advancing multimodal human-centered ai at the 10th abaw workshop & competition,” 2026

    2026

  4. [4]

    Deep affect prediction in-the- wild: Aff-wild database and challenge, deep architectures, and beyond,

    D. Kollias, P. Tzirakis, M. A. Nicolaou, A. Papaioannou, G. Zhao, B. Schuller, I. Kotsia, and S. Zafeiriou, “Deep affect prediction in-the- wild: Aff-wild database and challenge, deep architectures, and beyond,” International Journal of Computer Vision, pp. 1–23, 2019

    2019

  5. [5]

    Abaw: Valence-arousal estimation, expression recognition, action unit detection & multi-task learning challenges,

    D. Kollias, “Abaw: Valence-arousal estimation, expression recognition, action unit detection & multi-task learning challenges,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 2328–2336

    2022

  6. [6]

    Abaw: Learning from synthetic data & multi-task learning challenges,

    D. Kollias, “Abaw: Learning from synthetic data & multi-task learning challenges,” inEuropean Conference on Computer Vision. Springer, 2023, pp. 157–172

    2023

  7. [7]

    The 6th affective behavior analysis in-the-wild (abaw) competition,

    D. Kollias, P. Tzirakis, A. Cowen, S. Zafeiriou, I. Kotsia, A. Baird, C. Gagne, C. Shao, and G. Hu, “The 6th affective behavior analysis in-the-wild (abaw) competition,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2024, pp. 4587– 4598

    2024

  8. [8]

    Affective behavior analysis using action unit relation graph and multi- task cross attention,

    D.-K. Nguyen, S. Pant, N.-H. Ho, G.-S. Lee, S.-H. Kim, and H.-J. Yang, “Affective behavior analysis using action unit relation graph and multi- task cross attention,” inComputer Vision – ECCV 2022 Workshops, ser. Lecture Notes in Computer Science. Springer, 2023, pp. 132–142

    2022

Show all 36 references
  1. [9]

    Affective behavior analysis using task- adaptive and AU-assisted graph network,

    X. Li, W. Du, and H. Yang, “Affective behavior analysis using task- adaptive and AU-assisted graph network,” inComputer Vision – ECCV 2024 Workshops, ser. Lecture Notes in Computer Science. Springer, 2025, pp. 393–403

    2024

  2. [10]

    HSEmotion team at the 7th ABAW challenge: Multi-task learning and compound facial expression recognition,

    A. V . Savchenko, “HSEmotion team at the 7th ABAW challenge: Multi-task learning and compound facial expression recognition,”arXiv preprint arXiv:2407.13184, 2024

    2024 arXiv

  3. [11]

    Affective be- haviour analysis via progressive learning,

    C. Liu, W. Zhang, F. Qiu, L. Li, D. Wang, and X. Yu, “Affective be- haviour analysis via progressive learning,” inComputer Vision – ECCV 2024 Workshops, ser. Lecture Notes in Computer Science. Springer, 2025, pp. 366–379

    2024

  4. [12]

    A multi-task mean teacher for semi-supervised facial affective behavior analysis,

    L. Wang, S. Wang, J. Qi, and K. Suzuki, “A multi-task mean teacher for semi-supervised facial affective behavior analysis,” 2021

    2021

  5. [13]

    SS- MFAR: Semi-supervised multi-task facial affect recognition,

    D. Gera, B. N. S. Kumar, B. V . R. Kumar, and S. Balasubramanian, “SS- MFAR: Semi-supervised multi-task facial affect recognition,” 2022

    2022

  6. [14]

    Semi- supervised learning with deep generative models,

    D. P. Kingma, D. J. Rezende, S. Mohamed, and M. Welling, “Semi- supervised learning with deep generative models,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 27, 2014

    2014

  7. [15]

    Neural network ensembles, cross validation, and active learning,

    A. Krogh and J. Vedelsby, “Neural network ensembles, cross validation, and active learning,” inAdvances in Neural Information Processing Systems (NIPS), vol. 7, 1995, pp. 231–238

    1995

  8. [16]

    Diversity creation methods: a survey and categorisation,

    G. Brown, J. Wyatt, R. Harris, and X. Yao, “Diversity creation methods: a survey and categorisation,”Information Fusion, vol. 6, no. 1, pp. 5–20, 2005

    2005

  9. [17]

    Simple and scalable predictive uncertainty estimation using deep ensembles,

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,” inAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  10. [18]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inInternational Conference on Learning Representations (ICLR), 2022

    2022

  11. [19]

    LoRA ensembles for large language model fine-tuning,

    X. Wang, L. Aitchison, and M. Rudolph, “LoRA ensembles for large language model fine-tuning,”arXiv preprint arXiv:2310.00035, 2023

    2023 arXiv

  12. [20]

    FSFM: A generalizable face security foundation model via self-supervised facial representation learning,

    G. Wang, F. Lin, T. Wu, Z. Liu, Z. Ba, and K. Ren, “FSFM: A generalizable face security foundation model via self-supervised facial representation learning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  13. [21]

    Multi-modal facial affective analysis based on masked autoencoder,

    W. Zhang, B. Ma, F. Qiu, and Y . Ding, “Multi-modal facial affective analysis based on masked autoencoder,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023

    2023

  14. [22]

    AffectNet: A database for facial expression, valence, and arousal computing in the wild,

    A. Mollahosseini, B. Hasani, and M. H. Mahoor, “AffectNet: A database for facial expression, valence, and arousal computing in the wild,”IEEE Transactions on Affective Computing, vol. 10, no. 1, pp. 18–31, 2019

    2019

  15. [23]

    DINOv2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanniet al., “DINOv2: Learning robust visual features without supervision,”Transactions on Machine Learning Research (TMLR), 2024

    2024

  16. [24]

    EmotioNet: An accurate, real-time algorithm for the automatic annotation of a million facial expressions in the wild,

    C. F. Benitez-Quiroz, R. Srinivasan, and A. M. Martinez, “EmotioNet: An accurate, real-time algorithm for the automatic annotation of a million facial expressions in the wild,” inIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2016, pp. 5562–5570

    2016

  17. [25]

    Aff-wild: Valence and arousal ‘in-the-wild’ challenge,

    S. Zafeiriou, D. Kollias, M. A. Nicolaou, A. Papaioannou, G. Zhao, and I. Kotsia, “Aff-wild: Valence and arousal ‘in-the-wild’ challenge,” inComputer Vision and Pattern Recognition Workshops (CVPRW), 2017 IEEE Conference on. IEEE, 2017, pp. 1980–1987

    2017

  18. [26]

    Expression, affect, action unit recog- nition: Aff-wild2, multi-task learning and arcface,

    D. Kollias and S. Zafeiriou, “Expression, affect, action unit recog- nition: Aff-wild2, multi-task learning and arcface,”arXiv preprint arXiv:1910.04855, 2019

    1910 arXiv

  19. [27]

    Face behavior a la carte: Expressions, affect and action units in a single network,

    D. Kollias, V . Sharmanska, and S. Zafeiriou, “Face behavior a la carte: Expressions, affect and action units in a single network,”arXiv preprint arXiv:1910.11111, 2019

    1910 arXiv

  20. [28]

    Analysing affective behavior in the first abaw 2020 competition,

    D. Kollias, A. Schulc, E. Hajiyev, and S. Zafeiriou, “Analysing affective behavior in the first abaw 2020 competition,” in2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), 2020, pp. 794–800

    2020

  21. [29]

    Affect analysis in-the-wild: Valence- arousal, expressions, action units and a unified framework,

    D. Kollias and S. Zafeiriou, “Affect analysis in-the-wild: Valence- arousal, expressions, action units and a unified framework,”arXiv preprint arXiv:2103.15792, 2021

    2021 arXiv

  22. [30]

    Analysing affective behavior in the second abaw2 competition,

    D. Kollias and S. Zafeiriou, “Analysing affective behavior in the second abaw2 competition,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 3652–3660

    2021

  23. [31]

    Distribution matching for heterogeneous multi-task learning: a large-scale face study,

    D. Kollias, V . Sharmanska, and S. Zafeiriou, “Distribution matching for heterogeneous multi-task learning: a large-scale face study,”arXiv preprint arXiv:2105.03790, 2021

    2021 arXiv

  24. [32]

    Abaw: Valence-arousal estimation, expression recognition, action unit detection & emotional reaction intensity estimation challenges,

    D. Kollias, P. Tzirakis, A. Baird, A. Cowen, and S. Zafeiriou, “Abaw: Valence-arousal estimation, expression recognition, action unit detection & emotional reaction intensity estimation challenges,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogn...

    2023

  25. [33]

    Distribution matching for multi-task learning of classification tasks: a large-scale study on faces & beyond,

    D. Kollias, V . Sharmanska, and S. Zafeiriou, “Distribution matching for multi-task learning of classification tasks: a large-scale study on faces & beyond,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 3, 2024, pp. 2813–2821

    2024

  26. [34]

    Behaviour4all: in-the- wild facial behaviour analysis toolkit,

    D. Kollias, C. Shao, O. Kaloidas, and I. Patras, “Behaviour4all: in-the- wild facial behaviour analysis toolkit,”arXiv preprint arXiv:2409.17717, 2024

    2024 arXiv

  27. [35]

    Advancements in affective and behavior analysis: The 8th abaw workshop and competi- tion,

    D. Kollias, P. Tzirakis, A. Cowen, S. Zafeiriou, I. Kotsia, E. Granger, M. Pedersoli, S. Bacon, A. Baird, C. Gagneet al., “Advancements in affective and behavior analysis: The 8th abaw workshop and competi- tion,” inProceedings of the Computer Vision and Pattern Recognition Co...

    2025

  28. [36]

    From emotions to violence: Multimodal fine-grained behavior analysis at the 9th abaw,

    D. Kollias, S. Zafeiriou, I. Kotsia, G. Slabaugh, D. C. Senadeera, J. Zheng, K. K. K. Yadav, C. Shao, and G. Hu, “From emotions to violence: Multimodal fine-grained behavior analysis at the 9th abaw,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, ...

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.