Pith. sign in

REVIEW 4 major objections 4 minor 39 references

AffectFlow-DINO: Uncertainty-Aware Multi-Task Affect Estimation via Conditional Rectified Flow

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

Pith's one-line read A conditional rectified-flow head over a 22-dimensional affect vector learns the joint distribution of valence-arousal, expressions, and action units, and the paper shows that Monte-Carlo decoding, backbone fine-tuning, and post-hoc thresho

desk verdict Solid engineering with an honest ablation study, but the 'consistently improves' claim is contradicted by the paper's own fine-tuned tables—the flow's real role is a regularizer. read the letter →

arxiv 2607.13250 v1 pith:JBGACSDA submitted 2026-07-14 cs.CV

classification cs.CV
keywords rectifiedflowmulti-taskaffectestimationvalence-arousalfacialexpressionrecognitionactionunitdetectionuncertaintythresholdcalibrationABAW
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 argues that modeling the full joint distribution of facial affect—continuous valence-arousal, eight expression classes, and twelve action units—as a 22-dimensional conditional rectified flow, rather than as a point estimate, captures the inherent ambiguity of in-the-wild faces. On the multi-task affect benchmark, the authors show that Monte-Carlo flow decoding improves valence-arousal concordance by +0.058 over deterministic decoding with a frozen backbone, and that the flow objective acts as a structure-aware regularizer once the backbone is fine-tuned. After retuning the flow head during fine-tuning and applying post-hoc per-AU and per-expression threshold calibration, the system reaches a composite score of 1.177 on the validation split, more than 2.6 times the official baseline of 0.45. A key practical claim is that severe class imbalance suppresses, but does not erase, rare-class signal: calibration alone lifts Fear F1 from 3.8% to 33.1% without any retraining.

What carries the argument

The central object is a conditional rectified flow: a learned straight-line transport map from Gaussian noise to the 22-dimensional affect target, conditioned on the image embedding. The joint vector concatenates valence, arousal, one-hot expression, and binary AU scores; categorical and binary dimensions are treated as real-valued during integration and discretized only at the end. The masked flow loss plus deterministic task losses form the overall objective, with flow weight β. Inference averages N=16 Euler trajectories (T=30 steps) for a Monte Carlo estimate. The other key mechanism is post-hoc per-AU and per-class threshold calibration, which tunes decision boundaries on the validation

What would settle it

Train the same architecture on only the 36.6% of frames with full three-task annotation, and compare the flow model's predictions and per-class calibration against the masked model on a held-out fully-annotated set; if the masked model's rare-class margins are biased relative to the fully-supervised model, the missingness-independence assumption fails and the claimed one-to-many posterior is not the true posterior.

Watch

Extended reading notes

Core claim

The central claim is that a conditional rectified-flow head over the joint affect vector learns a meaningful distribution p(y|x), not just a point predictor. The evidence is that with a frozen DINOv3 backbone, flow decoding outperforms deterministic decoding for the same jointly-trained model, particularly for valence and arousal (+0.058 CCC-V), and flow-only training without deterministic task losses already achieves P_MTL=0.773, showing the flow head alone captures useful structure. Conversely, the two objectives are complementary: decoding a det-only model with the flow head collapses to 0.402, and decoding a flow-only model deterministically collapses to 0.408. The paper's final best con

Load-bearing premise

The load-bearing premise is that annotation missingness is independent of true affect values, so masking missing labels during flow training yields an unbiased estimate of the full joint distribution; the authors explicitly note this is not fully met in practice.

Editorial extensions

If this is right

  • Flow decoding improves continuous valence-arousal estimates more than categorical or binary tasks, because continuous targets benefit most from distributional averaging.
  • The flow objective and deterministic task losses are complementary: removing either collapses decoding performance, while jointly they yield the best frozen-backbone result.
  • Post-hoc threshold calibration recovers rare-class signal (Fear 3.8% to 33.1%, Sadness 17.1% to 28.2%, AU15 0% to 10.8%, AU23 0.5% to 18.9%) without retraining, indicating imbalanced training suppresses but does not erase discriminative features.
  • Backbone fine-tuning is the largest performance lever, but the flow head must be retuned jointly; flow retuning with β=1.0 improves deterministic decoding to P_MTL=1.073 (calibrated 1.123).
  • The final combined per-AU and per-class calibration yields P_MTL=1.177, versus the official baseline of 0.45.

Reading between the lines

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

  • If the missingness-independence assumption is violated—rare expressions being systematically under-annotated—the learned flow marginal is biased; a testable extension is training on only fully-annotated frames and comparing the flow's per-class calibration to the masked model.
  • The success of post-hoc threshold calibration suggests decision-boundary optimization could be folded into training via learned per-class temperature parameters, potentially improving rare-class F1 further without retraining.
  • Since flow decoding saturates at N=8 samples and T=10 steps, the generative head may be a cheap uncertainty estimator for affect; a natural extension is using the spread of sampled trajectories as an input-ambiguity measure for active learning or rejection.
  • The observation that fine-tuning favors deterministic decoding while the flow objective still helps as a regularizer suggests the generative head's value may shift from decoder to regularizer as the backbone adapts; future work could decouple these roles explicitly.
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 AffectFlow-DINO, a multi-task affective computing system for the ABAW 2026 multi-task learning challenge. It combines a frozen DINOv3 ViT-S/16 backbone with three deterministic heads (valence-arousal regression, expression classification, AU detection) and a conditional rectified-flow head that models a 22-dimensional joint affect vector. The authors claim that the flow head learns the conditional distribution p(y|x), enabling uncertainty-aware one-to-many predictions via Monte-Carlo decoding, and that this decoding 'consistently improves' deterministic prediction, especially for valence-arousal. They also propose post-hoc per-AU and per-class expression threshold calibration, reporting that this recovers rare-class performance (e.g., Fear F1 from 3.8% to 33.1%) and, together with backbone fine-tuning and flow retuning, yields P_MTL=1.177 on the validation split versus an official baseline of 0.450. The paper includes extensive ablations over 26 configurations, including frozen-backbone objective isolation, flow-loss weight, inference efficiency, fine-tuning, flow retuning, and calibration strategies.

Significance. If fully supported, the paper would be a useful contribution: it is, to the authors' knowledge, the first application of conditional rectified flow to heterogeneous multi-task affect estimation, and it provides a systematic ablation study with useful sanity checks (det-only decoding with flow collapses to 0.402, flow-only decoding with det collapses to 0.408). The post-hoc calibration idea is practically interesting. However, the central claim of consistency is contradicted by the paper's own fine-tuned experiments, the final headline number is obtained by deterministic decoding after validation-set calibration, and the masked-flow validity assumption is acknowledged to be violated. These issues prevent acceptance in the current form.

major comments (4)
  1. [Abstract; §4.3, §4.6, §4.9; Tables 2, 4] The claim that 'rectified-flow decoding consistently improves deterministic prediction' is contradicted by the manuscript's own fine-tuning results. Table 2 shows that after low-LR fine-tuning deterministic decoding reaches P_MTL=1.045 versus flow decoding 0.928. Table 4 shows for both flow-retuning variants that deterministic decoding (1.062 for β=0.5; 1.073 for β=1.0) outperforms flow decoding (0.956 and 0.931). Section 4.9 explicitly admits that 'deterministic decoding continues to outperform flow decoding in all fine-tuned variants.' The improvement from flow decoding is confined to the frozen-backbone regime (Table 1: 0.826 vs 0.802, CCC-V +0.058). Moreover, the best reported P_MTL=1.177 is obtained with deterministic decoding of a flow-retrained model after calibration, not via Monte-Carlo sampling. The abstract and introduction should be revised: in the final configuration, the fl
  2. [§4.4, §4.7, §4.9; Tables 3, 26] The headline result P_MTL=1.177 is obtained by tuning per-AU thresholds and per-class expression logit weights directly on the validation split and then reporting the same validation split. Section 4.9 concedes 'All results in this paper are on the local s-Aff-Wild2 validation split.' This is validation-set optimization, not an unbiased evaluation. The Fear recovery (0.038→0.331) and the +0.054 expression-calibration gain are in-sample fits. Without a held-out test split or nested cross-validation, the reported 1.177 cannot be regarded as a reliable estimate of generalization. The authors should explicitly disclose this in the abstract and results, and ideally re-evaluate on the challenge test set or report cross-validated calibrated scores.
  3. [§3.5, Eq. (8)] The masked-flow objective assumes that annotation missingness is independent of true affect values, so that each masked step is a valid update for the full joint p(y|x). The authors immediately note that 'This assumption is not fully met in practice' because rare classes are likely under-annotated. This is not a minor caveat: the paper's novelty claim — that the flow head learns the true conditional distribution enabling uncertainty-aware one-to-many predictions — depends on the masked marginals being unbiased. If under-annotation is correlated with true affect, the learned flow is biased and the 'uncertainty' is not a calibrated posterior. The paper needs either (a) empirical evidence that missingness is approximately ignorable, (b) a missingness model or imputation, or (c) a re-framing that does not claim posterior validity.
  4. [Abstract; Section 2; Table 1; Table 5] The official baseline is reported inconsistently: the abstract, Table 1, and Table 5 state P_MTL=0.450, while Section 2 states 'the official challenge baseline is P_MTL=0.34.' Additionally, Section 2 compares to the leading 7th ABAW entry (P_MTL=1.529) from a hidden test set, whereas Section 4.9 states that such scores are 'not directly comparable.' The paper should use one consistent baseline and explicitly separate validation results from hidden test comparisons.
minor comments (4)
  1. [§4.2/§4.3; Table 6] Section 4.3 refers to 'the best frozen-backbone result of 0.831,' but the main Table 5 does not show this value; it appears only in the unabridged Table 6 as the β=1, T=50 configuration. Please clarify the reference.
  2. [Figure 2] Figure 2 uses experiment IDs (E1, E3, E9, E22, E27) that are not defined in the main text or appendix. Either define them or replace with the configuration names used in Tables 5/6.
  3. [§4.4] The claim that per-AU calibration is 'theoretically principled rather than an ad-hoc leaderboard trick' is supported only by an intuition about class-imbalanced BCE. Either provide a formal argument or soften the wording.
  4. [General] No code, checkpoints, or reproducibility statement is provided. Given the number of ablations and the calibration tuning, a reproducibility section or link would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all load-bearing claims are empirical ablations; the single self-citation is not load-bearing, and the reported calibration gains are fitted validation scores, not definitional identities.

full rationale

The paper contains no derivation chain in which a claimed output is defined in terms of the quantity it purportedly predicts. The central comparison — flow decoding versus deterministic decoding — is an empirical ablation (Tables 1, 2, 4): the frozen-backbone gain in CCC-V (+0.058) is a measured training outcome, not a constructional identity with the flow loss in Eq. (8). The paper's own fine-tuned results contradict the abstract's 'consistently improves' wording (Table 2: deterministic 1.045 vs flow 0.928; Table 4: deterministic 1.062/1.073 vs flow 0.956/0.931, with Section 4.9 explicitly stating deterministic decoding continues to outperform flow decoding), but this is an internal-consistency and framing problem, not circularity. The per-AU and per-class threshold calibration (Sections 4.4 and 4.7) tunes free thresholds/weights on the validation split and then reports validation performance; that is fitted reporting and limits the strength of the 'recovers rare classes' claim, but the calibrated F1 values are not defined to equal the training signal or the loss — they are optimizable parameters, so the step is not circular. The only self-citation is reference [2] (Bekhouche et al.), used in Related Work as a general pointer that conflict-aware fusion motivates richer predictive models; it is not load-bearing for any equation, result, or forced choice. The masked-flow validity assumption in Section 3.5 is an explicitly admitted statistical limitation of the learned p(y|x), not a circular reduction: the missingness-independence assumption fails to make the training objective equal the target posterior. No quoted equation or parameter is equivalent by construction to the paper's headline scores. Per the standard that a 0 score is the normal, honest finding for empirically self-contained work, the circularity score is 0.

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

No new physical or architectural entities are introduced: the 22-dimensional affect vector is a design choice, and DINOv3/DINOv2, rectified flow, and threshold calibration are existing components. The load-bearing empirical choices are the validation-tuned thresholds/class weights and the flow/task loss weights, all of which were selected on the same validation split used for reporting.

free parameters (5)
  • Flow loss weight beta = 1.0 (default; plateau 0.5-1.0)
    Ablated over {0.25,0.5,1.0,2.0} on validation (Appendix C.1); affects balance of generative and deterministic objectives.
  • Task loss weights lambda_VA, lambda_EXPR, lambda_AU = 1.0 each (default)
    Swept individually; chosen by validation P_MTL. E.g., lambda_VA=2 gives +0.004, lambda_EXPR>1 hurts, lambda_AU global scaling no effect.
  • Per-AU thresholds tau_k = Not reported; grid-searched per AU on validation
    Section 4.4; directly produces the +0.06 P_AU calibration gain and part of the headline 1.177.
  • Per-class expression logit weights = Not reported; tuned on validation to maximize macro-F1
    Section 4.7; produces the Fear 3.8%->33.1% and Sadness 17.1%->28.2% gains used in the headline.
  • Inference sampling count N and Euler steps T = N=16, T=30
    Saturation sweep on validation (Appendix C.2); N=8/T=10 already saturate, but defaults chosen conservatively.
assumptions (5)
  • standard math Rectified flow provides a straight-line transport from noise to data that can be integrated with few Euler steps (Liu et al. 2023).
    Core of the flow head in Section 3.5, Eqs. (7)-(10); the saturation at N=8/T=10 is attributed to this property.
  • domain assumption Annotation missingness is independent of true affect values.
    Section 3.5: 'under the mild assumption that annotation missingness is independent of true affect values, each masked step is a valid update for the full joint p(y|x)'. The authors explicitly state the assumption 'is not fully met in practice'.
  • domain assumption s-Aff-Wild2 labels are ground truth for VA, expression, and AUs.
    All training and validation scores are computed against these labels; no label-noise model is used.
  • ad hoc to paper Embedding discrete categorical and binary targets into continuous R^22 and allowing off-manifold intermediate flow values is harmless because straight-line trajectories arrive on the target manifold at t=1.
    Section 3.2 explicitly calls this a 'deliberate design choice' with 'known cost of trajectories visiting off-manifold intermediate values'; no empirical validation of the manifold claim is given.
  • domain assumption DINOv3/DINOv2 features provide a suitable frozen visual representation.
    The backbone is frozen in most experiments; if the pretrained features are not affect-discriminative, the flow-head gains would not transfer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AffectFlow-DINO: Uncertainty-Aware Multi-Task Affect Estimation via Conditional Rectified Flow." pith.science (2026). https://pith.science/paper/JBGACSDA

@misc{pith2026260713250,
  author       = {Pith},
  title        = {Pith review of: AffectFlow-DINO: Uncertainty-Aware Multi-Task Affect Estimation via Conditional Rectified Flow},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JBGACSDA}},
  note         = {Machine review of arXiv:2607.13250}
}
abstract

We present \textbf{AffectFlow-DINO}, a multi-task learning system for the 11th ABAW challenge that extends a standard deterministic architecture with a conditional rectified-flow head to model the inherent ambiguity of in-the-wild facial behavior. Instead of predicting a single affect estimate, the model learns a conditional generative distribution, enabling uncertainty-aware one-to-many predictions through Monte Carlo sampling. The system jointly estimates continuous valence-arousal, classifies eight facial expressions, and detects twelve Action Units from static face images. Built on a frozen DINOv3 ViT-S/16 backbone, extensive ablation studies show that rectified-flow decoding consistently improves deterministic prediction, particularly for valence-arousal estimation (CCC-V $+0.058$). We further show that post-hoc threshold calibration effectively recovers performance on severely imbalanced rare classes (e.g., Fear: $3.8\% \rightarrow 33.1\%$) without retraining. Combined with backbone fine-tuning and flow retuning, the final model achieves $\mathbf{P_{MTL}=1.177}$, substantially outperforming the official challenge baseline of $P_{MTL}=0.45$.

Figures

Figures reproduced from arXiv: 2607.13250 by the authors.

Figure 1
Figure 1. Overview of AffectFlow-DINO. A frozen DINOv3 ViT-S/16 backbone encodes the face into a shared embedding z. Three deterministic task heads produce point predictions for VA, expression, and AUs. The conditional rectified-flow head receives z together with a linearly interpolated noised target yt and a sinusoidal time embedding γ(t), and learns to predict the transport velocity. At inference, N trajectories are integra… view at source ↗
Figure 2
Figure 2. PMT L progression across key configurations. Color encodes training stage: frozen backbone (light blue), fine-tuned (dark blue), fine-tuned with both calibrations (red). The dotted line marks the official baseline (PMT L = 0.450). Backbone fine-tuning is the largest single jump; post-hoc per-AU and per-class calibration provide a further +0.132 at zero retraining cost [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗
Figure 3
Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 6 linked inside Pith

  1. [1]

    Barros, P., Churamani, N., Lakomkin, E., Siqueira, H., Sutherland, A., Wermter, S.: Omg-emotion behavior dataset. In: Int. Joint Conf. Neural Netw. pp. 1–7 (2018) ABAW 2026 Multi-Task Affect Analysis 15

  2. [2]

    arXiv preprint arXiv:2603.15818 (2026)

    Bekhouche, S.E., Telli, H., Benlamoudi, A., Herrouz, S.E., Taleb-Ahmed, A., Ha- did, A.: Conflict-aware multimodal fusion for ambivalence and hesitancy recogni- tion. arXiv preprint arXiv:2603.15818 (2026)

  3. [3]

    In: IEEE Conf

    Benitez-Quiroz, C.F., Srinivasan, R., Martinez, A.M.: Emotionet: An accurate, real-time algorithm for the automatic annotation of a million facial expressions in the wild. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5562–5570 (2016)

  4. [4]

    Cabacas-Maso, J., Ortega-Beltran, E., Benito-Altamirano, I., Ventura, C.: Enhanc- ing facial expression recognition through dual-direction attention mixed feature networks: Application to 7th abaw challenge (2024)

  5. [5]

    In: IEEE Conf

    He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 16000– 16009 (2022)

  6. [6]

    In: ACM Int

    Jiang, X., Zong, Y., Zheng, W., Tang, C., Xia, W., Lu, C., Liu, J.: Dfew: A large- scale database for recognizing dynamic facial expressions in the wild. In: ACM Int. Conf. Multimedia. pp. 2881–2889 (2020)

  7. [7]

    In: IEEE Conf

    Kollias, D.: Abaw: Valence-arousal estimation, expression recognition, action unit detection and multi-task learning challenges. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 2328–2336 (2022)

  8. [8]

    Kollias, D.: Abaw: Learning from synthetic data and multi-task learning challenges. In: Eur. Conf. Comput. Vis. pp. 157–172. Springer (2023)

Show all 39 references
  1. [9]

    In: IEEE Conf

    Kollias, D.: Multi-label compound expression recognition: C-expr database and network. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 5589–5598 (2023)

  2. [10]

    In: IEEE Int

    Kollias,D.,Schulc,A.,Hajiyev,E.,Zafeiriou,S.:Analysingaffectivebehaviorinthe first abaw 2020 competition. In: IEEE Int. Conf. Autom. Face Gesture Recognit. pp. 794–800 (2020)

  3. [11]

    arXiv preprint arXiv:2409.17717 (2024)

    Kollias, D., Shao, C., Kaloidas, O., Patras, I.: Behaviour4all: In-the-wild facial behaviour analysis toolkit. arXiv preprint arXiv:2409.17717 (2024)

  4. [12]

    arXiv preprint arXiv:1910.11111 (2019)

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

  5. [13]

    arXiv preprint arXiv:2105.03790 (2021)

    Kollias, D., Sharmanska, V., Zafeiriou, S.: Distribution matching for heterogeneous multi-task learning: A large-scale face study. arXiv preprint arXiv:2105.03790 (2021)

  6. [14]

    In: AAAI

    Kollias, D., Sharmanska, V., Zafeiriou, S.: Distribution matching for multi-task learning of classification tasks: A large-scale study on faces and beyond. In: AAAI. vol. 38, pp. 2813–2821 (2024)

  7. [15]

    Kollias, D., Tzirakis, P., Baird, A., Cowen, A., Zafeiriou, S.: Abaw: Valence-arousal estimation, expression recognition, action unit detection and emotional reaction in- tensityestimationchallenges.In:IEEEConf.Comput.Vis.PatternRecog.Worksh. pp. 5888–5897 (2023)

  8. [16]

    In: IEEE Conf

    Kollias, D., Tzirakis, P., Cowen, A., Zafeiriou, S., Kotsia, I., Baird, A., Gagne, C., Shao, C., Hu, G.: The 6th affective behavior analysis in-the-wild (abaw) competi- tion. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 4587–4598 (2024)

  9. [17]

    In: IEEE Conf

    Kollias, D., Tzirakis, P., Cowen, A., Zafeiriou, S., Kotsia, I., Granger, E., Pedersoli, M., Bacon, S., Baird, A., Gagne, C., et al.: Advancements in affective and behavior analysis: The 8th abaw workshop and competition. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. ...

  10. [18]

    Bekhouche et al

    Kollias, D., Tzirakis, P., Cowen, A., Zafeiriou, S., Kotsia, I., Granger, E., Pedersoli, M., Bacon, S., Madsen, J., Belharbi, S., Aslam, M.H., Shao, C., Hu, G.: From affect to complex behavior: Advancing multimodal human-centered ai at the 10th abaw workshop and competition (2...

  11. [19]

    Kollias, D., Tzirakis, P., Nicolaou, M.A., Papaioannou, A., Zhao, G., Schuller, B., Kotsia, I., Zafeiriou, S.: Deep affect prediction in-the-wild: Aff-wild database and challenge, deep architectures, and beyond. Int. J. Comput. Vis.127(6–7), 907–929 (2019)

  12. [20]

    arXiv preprint arXiv:1910.04855 (2019)

    Kollias, D., Zafeiriou, S.: Expression, affect, action unit recognition: Aff-wild2, multi-task learning and arcface. arXiv preprint arXiv:1910.04855 (2019)

  13. [21]

    arXiv preprint arXiv:2103.15792 (2021)

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

  14. [22]

    In: IEEE Int

    Kollias, D., Zafeiriou, S.: Analysing affective behavior in the second abaw2 com- petition. In: IEEE Int. Conf. Comput. Vis. Worksh. pp. 3652–3660 (2021)

  15. [23]

    Kollias, D., Zafeiriou, S., Kotsia, I., Dhall, A., Ghosh, S., Shao, C., Hu, G.: 7th abaw competition: Multi-task learning and compound expression recognition. In: Eur. Conf. Comput. Vis. pp. 31–45. Springer (2024)

  16. [24]

    In: IEEE Int

    Kollias, D., Zafeiriou, S., Kotsia, I., Slabaugh, G., Senadeera, D.C., Zheng, J., Ya- dav, K.K.K., Shao, C., Hu, G.: From emotions to violence: Multimodal fine-grained behavior analysis at the 9th abaw. In: IEEE Int. Conf. Comput. Vis. Worksh. pp. 1–12 (2025)

  17. [25]

    In: IEEE Int

    Kossaifi, J., Tzimiropoulos, G., Todorovic, S., Pantic, M.: Afew-va database for va- lence and arousal estimation in-the-wild. In: IEEE Int. Conf. Autom. Face Gesture Recognit. pp. 327–333 (2017)

  18. [26]

    In: IEEE Trans

    Kossaifi, J., Walecki, R., Panagakis, Y., Shen, J., Schmitt, M., Ringeval, F., Han, J., Pandit, V., Toisoul, A., Schuller, B., et al.: Sewa db: A rich database for audio- visual emotion and sentiment research in the wild. In: IEEE Trans. Pattern Anal. Mach. Intell. (2019)

  19. [27]

    In: IEEE Conf

    Li, S., Deng, W., Du, J.: Reliable crowdsourcing and deep locality-preserving learn- ing for expression recognition in the wild. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 2852–2861 (2017)

  20. [28]

    Li, X., Du, W., Yang, H.: Affective behavior analysis using task-adaptive and au- assisted graph network (2024)

  21. [29]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object de- tection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 2980–2988 (2017)

  22. [30]

    Liu, C., Zhang, W., Qiu, F., Li, L., Wang, D., Yu, X.: 7abaw-affective behaviour analysis via progressive learning (2024)

  23. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019)

    Liu, S., Johns, E., Davison, A.J.: End-to-end multi-task learning with attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2019)

  24. [32]

    In: International Conference on Learning Representations (2023)

    Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: International Conference on Learning Representations (2023)

  25. [33]

    In: IEEE Int

    Mavadati, S.M., Mahoor, M.H., Bartlett, K., Trinh, P., Cohn, J.F.: Disfa: A spon- taneous facial action intensity database. In: IEEE Int. Conf. Autom. Face Gesture Recognit. pp. 1–8 (2013)

  26. [34]

    IEEE Trans

    Mollahosseini, A., Hasani, B., Mahoor, M.H.: Affectnet: A database for facial ex- pression, valence, and arousal computing in the wild. IEEE Trans. Affect. Comput. 10(1), 18–31 (2017)

  27. [35]

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. In: arXiv preprint arXiv:2304.07193 (2023) ABAW 2026 Multi-Task Affect Analys...

  28. [36]

    In: IEEE Conf

    Wang, Y., Sun, Y., Huang, Y., Liu, Z., Gao, S., Zhang, W., Ge, W., Zhang, W.: Ferv39k: A large-scale multi-scene dataset for facial expression recognition in videos. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 20922–20931 (2022)

  29. [37]

    Yu, J., Zhang, Y., Lu, X., Zheng, Y., Wang, Y., Zhu, L.: Solution for 8th compe- tition on affective and behavior analysis in-the-wild (2025)

  30. [38]

    In: IEEE Conf

    Zafeiriou, S., Kollias, D., Nicolaou, M.A., Papaioannou, A., Zhao, G., Kotsia, I.: Aff-wild: Valence and arousal in-the-wild challenge. In: IEEE Conf. Comput. Vis. Pattern Recog. Worksh. pp. 1980–1987 (2017)

  31. [39]

    Image Vis

    Zhang, X., Yin, L., Cohn, J.F., Canavan, S., Reale, M., Horowitz, A., Liu, P., Girard, J.M.: Bp4d-spontaneous: A high-resolution spontaneous 3d dynamic facial expression database. Image Vis. Comput.32(10), 692–706 (2014) A Full Results Summary Table 6 is the unabridged version...

Pith tools

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