Pith. sign in

REVIEW 3 major objections 5 minor 22 references

Multi-task facial affect analysis improves when valence–arousal, expression, and action-unit predictions are given their own fusion, temporal, and calibration strategies rather than a shared architecture.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A task-adaptive system that mixes two frozen visual features with per-task fusion and temporal strategies scores 1.6341 on the ABAW11 validation set, beating its own shared multi-task baselines.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Useful, honest ablation study of task-specific downstream configs for ABAW11, but the comparison against shared MTL is confounded by missing temporal/calibration components and the validation selection shows instability the authors themselves exposed. the 3 major comments →

arxiv 2607.13986 v1 pith:DHDTSPFX submitted 2026-07-15 cs.CV

Task-Specific Feature Fusion Method for Multi-Task Affective Behavior Analysis

Jiajun Sun , Zhe Gao This is my paper
classification cs.CV
keywords multi-task learningaffective behavior analysistask-adaptive feature fusionfacial expression recognitionaction unit detectionvalence-arousal estimationfrozen featurestemporal modeling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 a single shared multi-task architecture is the wrong default for the three facial-affect tasks in the official benchmark: valence–arousal regression, categorical expression recognition, and action-unit detection. The proposed alternative is to adapt two pretrained visual backbones (one transformer-based, one convolutional) on external expression data, freeze them, and then let each task select its own downstream expert from a pool of fusion, temporal, and calibration options. On the official validation split, the task-selected system achieves an overall score of 1.6341 (expression macro-F1 0.4222, action-unit macro-F1 0.5402, valence-arousal mean CCC 0.6717), well above the best shared-model baseline of 1.3598. A video-level five-fold cross-validation gives 1.6166±0.0394, again stronger than the shared baselines, suggesting that task-adaptive fusion of frozen features is a simple and effective default strategy.

Core claim

The central discovery is that valence–arousal, expression, and action-unit tasks favor different feature sources, temporal strategies, fusion mechanisms, and calibration procedures even though they share the same facial frames. Concatenating the transformer and convolutional features with post-hoc smoothing helps expression recognition; late fusion of independently trained logits helps action units; and valence-arousal gains most from residual fusion on the official split, though a temporal convolutional model is more stable across video-level folds. The paper demonstrates this by comparing frame-level heads, temporal convolutional heads, post-hoc smoothing, tree-based predictors, feature co

What carries the argument

The engine is the frozen-feature expert pool. Two pretrained visual backbones — a transformer and a convolutional network — are adapted on an external expression dataset and then frozen to emit 1024-dimensional frame descriptors from the benchmark images. A menu of per-task choices is placed on top: frame-level neural heads, temporal convolutional heads, post-hoc mean/median smoothing, tree-based classifiers, feature concatenation, gated and residual fusion, late logit fusion, and independently tuned per-action-unit thresholds. The central move is to grant each task the best validation-scoring recipe rather than forcing all three through one shared structure.

Load-bearing premise

The assumption that the official validation split is representative enough of the unseen test distribution that every choice tuned on it — per-action-unit thresholds, smoothing windows, fusion weights, and the per-task expert — will transfer.

What would settle it

On the held-out test split, compare the final task-adaptive system (concat+smoothing for expression, late fusion for action units, temporal convolutional for VA) against the best shared multi-task baseline and against the validation-selected residual-fusion VA expert; if the shared baseline or the residual VA expert scores higher, the central claim fails. Equivalently, replace the per-AU thresholds tuned on the validation split with fixed 0.5 thresholds: if the tuned thresholds do not improve test macro-F1, the calibration step is overfit to the split.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Divergent preferences across the three tasks imply that forcing a single shared network costs about 0.27 of the overall score (1.6341 vs 1.3598) on this benchmark.
  • Strong frozen features from pretrained backbones can support competitive affect prediction without end-to-end fine-tuning, which simplifies training and reduces compute.
  • Per-action-unit threshold calibration is a meaningful post-processing gain for action-unit detection and should be a default calibration step.
  • Video-level cross-validation can reverse a choice made on the official split (the VA expert), so model selection should be validated across video partitions, not just the single official split.
  • The per-task match between task type (regression, single-label, multi-label) and fusion/temporal strategy offers a reusable design heuristic for other multi-task affective systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension would be to fine-tune the two backbones jointly on the benchmark frames while keeping the same expert pool; if the frozen-feature system already saturates the available signal, gains should be small.
  • The official validation split appears to over-favor the residual fusion for VA relative to video-level folds, so the headline 1.6341 may shrink on the unseen test set; a public test release would settle which expert truly generalizes.
  • The manual per-task selection is effectively a small architecture search; an automatic weighting or gating over experts might match or exceed 1.6341 without hand-picking recipes per task.
  • The same frozen-feature menu could be applied to other multi-task face benchmarks to check whether the divergence pattern (continuous vs single-label vs multi-label tasks) generalizes beyond this dataset.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a task-adaptive feature fusion system for the ABAW11 multi-task affective behavior analysis challenge. Two externally adapted frozen backbones (DINOv2 ViT-L and DINOv3 ConvNeXt-base) are used to extract 1024-d frame features from s-Aff-Wild2. On top of these features, the authors systematically compare frame-level heads, temporal convolutional heads, post-hoc smoothing, LightGBM predictors, feature-level and logit-level fusion, threshold calibration, and shared MTL structures. The validation-selected configuration — concatenation with smoothing for EXPR, late logit fusion with per-AU threshold calibration for AU, and residual fusion with smoothing for VA — achieves an official validation score of 1.6341 (Table 4), while the revised cross-validated configuration using a ViT-L temporal VA expert achieves 1.6166±0.0394 (Table 5). The central claim is that task-adaptive selection over frozen features outperforms shared multi-task structures on ABAW11.

Significance. If the central claim holds, the paper provides a simple and effective default strategy for ABAW-style multi-task affective behavior analysis: rather than forcing all tasks into one shared architecture, choose per-task feature sources, temporal processing, fusion, and calibration on a validation split. The paper is careful in several respects: both backbones are frozen after external adaptation so all downstream comparisons use identical visual inputs; the ablation tables cover a broad set of alternatives; the appendices give implementation details (architectures, losses, smoothing windows, LightGBM settings); and Section 4.7 honestly reports that the official-validation-selected VA expert is less stable under video-level cross-validation than a temporal expert. These are genuine strengths. However, the main performance comparison is confounded, and the validation-selection procedure raises circularity concerns that weaken the central attribution.

major comments (3)
  1. [§4.5/Table 4 vs. Appendix A.6] The central claim — task-adaptive fusion beats shared MTL structures — is not isolated by the comparison. Appendix A.6 states that all shared MTL baselines 'operate independently at the frame level and do not introduce explicit temporal modeling', while the winning task-adaptive configuration uses median smoothing (window 25) for EXPR, per-AU threshold calibration for AU, and temporal smoothing/residual fusion for VA (A.4, A.5, Tables 2–3). The 0.2743 gap (1.6341 vs. 1.3598) could therefore be due to temporal context and calibration, which are orthogonal to task-adaptivity. Table 2 already shows that a single-feature ViT-L temporal VA expert reaches 0.6367, far above the best shared-MTL VA of 0.4962. Please add controlled shared-MTL baselines that include temporal modeling and/or calibration, or report task-adaptive variants that omit these components, so the effect of task-specific sele
  2. [§4.6/§4.7/Table 5] The official-validation-selected VA expert (residual fusion, 0.6717 in Table 4) is outperformed by the ViT-L temporal expert in all five video-level folds (0.6411±0.0309 vs. 0.5124±0.0378). This is a load-bearing instability: the headline 1.6341 relies on a configuration that the paper itself shows does not generalize across video partitions. The revised configuration (1.6166±0.0394) is presented after inspecting the same data used for selection, so its superiority over the shared baselines is not an independent confirmation. Please provide a cross-validated selection protocol in which all per-task choices (including the VA expert) are made on training folds only, and report the resulting test-fold performance.
  3. [Appendix A.4 / Eq. (2)] Many hyperparameters are selected on the official validation split: the 12 per-AU decision thresholds are each 'selected independently by maximizing its validation F1', the smoothing window size is evaluated over 3–25, and the late-fusion weight α is selected on validation. This repeated validation-set selection inflates the reported scores and makes them non-independent of the held-out set. Table 5 does not state whether the per-AU thresholds and smoothing windows were re-selected for each fold or fixed from the official split. Please clarify and, if possible, quantify the selection bias (e.g., by reporting scores with thresholds fixed on training folds or by nested cross-validation).
minor comments (5)
  1. [Table 3 caption] The caption says 'tuned macro-F1' but the column header is 'AU'. Please clarify what 'tuned' refers to — presumably per-AU threshold calibration — and apply the same terminology consistently in Tables 1–4.
  2. [§4.1, Eq. (3)] The text says 'F1_EXPR denotes the sum of the F1 scores over the eight expression classes', but the notation is confusing because F1 usually denotes a per-class harmonic mean. Please rewrite to avoid ambiguity, e.g., use Σ_c F1_c / 8.
  3. [Table 5] The table lists both VA-Res. and VA-TCN columns and the caption says 'Overall is computed using VA-TCN'. Please state this explicitly in a footnote or in the text to avoid uncertainty about which VA score is used.
  4. [§4.7] The phrase 'During this more detailed evaluation, we find that...' reads like a narrative rather than a result statement. Please rephrase to report the finding objectively.
  5. [Appendix A.5, Eq. (15)] In Eq. (15), the residual gate g_t appears to be shared with Eq. (13); define its argument explicitly (it should be the same G_res([a_t; b_t])) or rename to avoid ambiguity.

Circularity Check

3 steps flagged

Headline validation scores are selected on the same validation split used to report them; the five-fold final score is also selected on its own folds.

specific steps
  1. fitted input called prediction [Appendix A.4 (AU threshold calibration); Section 4.6 / Table 4]
    "For AU threshold calibration, the decision threshold of each AU is selected independently by maximizing its validation F1 score. Thresholds are searched from 0.01 to 0.99 with a step size of 0.01."

    The AU macro-F1 of 0.5402 reported in Table 4 and folded into the headline 1.6341 is computed on the very validation split whose F1 was maximized per AU. The reported AU score is therefore the objective being optimized, not an independent measurement; the 'tuned macro-F1' label in Tables 1-3 even names this.

  2. self definitional [Section 3.4 Eq. (2); Appendix A.4; Section 4.6]
    "Based on the official validation experiments, we construct a validation-selected system using the best-performing configuration for each task. ... where α is selected on the validation set. ... We evaluate window sizes of 3,5,7,9,13,17,and25; the standard EXPR configuration uses median smoothing with a window size of 25."

    The final system is defined as the configuration with the best score on the official validation split, and the fusion weight, smoothing windows, and per-task experts are all chosen by evaluating that same split. Hence the reported validation score 1.6341 is the selection criterion itself; the 'outperforming' claim on this split is partly guaranteed by construction.

  3. fitted input called prediction [Section 4.7 (Cross-Fold Expert Stability)]
    "Based on this observation, our final system retains concatenated features for EXPR and late fusion for AU, while replacing the residual VA expert with the ViT-L temporal convolutional expert. The resulting task-adaptive configuration achieves an overall five-fold score of 1.6166±0.0394."

    The final configuration is chosen after observing the five-fold results ('outperforms residual fusion in every fold'), and the reported mean is computed over exactly those folds. Thus 1.6166±0.0394 is also the score of the best configuration selected on the same folds, not an unbiased estimate for an unseen split, though the paper transparently reports the losing expert too.

full rationale

The central experimental pipeline is not circular: backbones are adapted on external AffectNet/RAF-DB data without ABAW11 annotations, downstream models are trained on official training labels, and the frozen-feature comparisons in Tables 1-3 are genuine measurements. There are no load-bearing self-citations or imported uniqueness theorems; the cited methods (DINOv2/v3, TCN, LightGBM, MTL) are independent external work. The circularity is localized to evaluation. Appendix A.4 fits each AU threshold to maximize validation F1, Eq. (2) selects α on the validation set, A.4 selects smoothing windows on the same split, and Section 4.6 defines the final system as the best validation configuration; the headline 1.6341 is therefore the selection objective, not an independent prediction. Section 4.7's five-fold check is more stable, but the final five-fold configuration is chosen after inspecting the same five-fold scores, so 1.6166±0.0394 also inherits selection bias. The additional confound (shared MTL baselines lack temporal modeling and calibration, Appendix A.6) is a validity threat to the attribution, not circularity. Net: partial circularity in the reported validation numbers, with independent feature/ablation content, so 6.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

No new physical or conceptual entities are postulated. The four-expert mixture-of-experts head (A.2), task-token model, and residual-relation modules (A.6) are internal model components whose behavior is fully specified by the paper; they make no claims about the world outside the model, so independent empirical evidence is not applicable.

free parameters (7)
  • 12 per-AU decision thresholds = searched 0.01-0.99, step 0.01, per AU
    A.4: each AU threshold is selected to maximize its validation F1; part of the headline AU score 0.5402.
  • Smoothing window size w = EXPR uses median smoothing, w=25; 7 candidates evaluated
    A.4: windows 3,5,7,9,13,17,25 compared; EXPR configuration uses median with w=25; contributes to EXPR 0.4222.
  • Late-fusion weight alpha = selected on validation set
    Eq. 2 in Section 3.4: used by the AU and VA late-fusion experts; chosen on the validation set.
  • Per-task expert selection = concat+median (EXPR), late fusion+calibration (AU), residual->TCN (VA)
    Sections 4.6 and 4.7: each task's configuration is the highest-scoring one among the evaluated pool on the same split used for reporting.
  • Loss weights lambda_mse=0.5, lambda_ccc=1; MTL lambdas=1 = 0.5, 1, 1, 1, 1
    Eq. 5 and Eq. 18: hand-set, no sensitivity analysis reported.
  • Residual-gate bias init -4 = -4
    A.5: bias initialized so training starts near the ViT-L branch; hand-chosen to control fusion behavior.
  • Backbone adaptation design and LightGBM hyperparameters = e.g., LightGBM: 300 estimators, lr 0.03, 31 leaves, subsample 0.9, feature-fraction 0.8
    A.2/A.3: frozen first 8 ViT blocks, 4-expert MoE head, and tree-model parameters chosen by hand; no ablation or sensitivity analysis reported.
axioms (5)
  • domain assumption The official ABAW11 MTL score S (Eq. 3) and per-task metrics (macro-F1, mean CCC per Eq. 4) are as stated and computed correctly on the official validation annotations.
    Used to construct all reported scores; cannot be verified without the official evaluation code.
  • domain assumption External adaptation on AffectNet+RAF-DB transfers to the s-Aff-Wild2 facial distribution.
    Sections 3.2/A.2: backbones are adapted on external expression sets then frozen; no ABAW11 fine-tuning, so transfer is assumed.
  • domain assumption Frozen 1024-d DINOv2/DINOv3 features contain sufficient task-relevant information for VA, EXPR, and AU.
    Section 3.2: features are extracted offline and never updated; no measurement of information loss from freezing is reported.
  • domain assumption Samples with missing labels are handled correctly by validity masks, and validation-based selection is a fair estimate of generalization.
    Sections 3.1/4.1 and 5: 'all conclusions are based on validation experiments'; Section 4.7 shows validation selection is unstable for the VA expert.
  • standard math Standard ML background: backpropagation, CCC formula, macro-F1, ViT/ConvNeXt architectures as published.
    Implicit reliance on published models [6,13,16,19] and standard losses; not proven in the paper.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Task-Specific Feature Fusion Method for Multi-Task Affective Behavior Analysis." pith.science (2026). https://pith.science/paper/DHDTSPFX

@misc{pith2026260713986,
  author       = {Pith},
  title        = {Pith review of: Task-Specific Feature Fusion Method for Multi-Task Affective Behavior Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DHDTSPFX}},
  note         = {Machine review of arXiv:2607.13986}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The 11th Affective Behavior Analysis in-the-wild (ABAW11) Multi-Task Learning Challenge requires a unified system to predict valence-arousal, categorical expressions, and facial action units from the official s-Aff-Wild2 images. Although these tasks are naturally related through facial behavior, our validation experiments show that they benefit from different visual features, temporal processing strategies, fusion mechanisms, and calibration procedures. In this paper, we study task-adaptive feature fusion for ABAW11 multi-task affective behavior analysis. We first adapt two pretrained visual backbones, DINOv2 ViT-L and DINOv3 ConvNeXt-base, on an external expression-oriented facial image set and then freeze them to extract complementary frame-level features from the official ABAW11 data. On top of these frozen features, we systematically compare frame-level prediction heads, temporal convolutional heads, post-hoc temporal smoothing, LightGBM models, feature concatenation, gated fusion, residual fusion, late logit fusion, threshold calibration, and shared MTL structures. The final system selects task-specific fusion and prediction strategies rather than forcing all tasks to share a single architecture. On the ABAW11 validation set, the selected system achieves an EXPR macro-F1 of 0.4222, an AU macro-F1 of 0.5402, and a mean VA CCC of 0.6717, resulting in an overall validation score of 1.6341. The results suggest that task-adaptive fusion of frozen visual features is a simple and effective strategy for ABAW-style multi-task affective behavior analysis.

Figures

Figures reproduced from arXiv: 2607.13986 by Jiajun Sun, Zhe Gao.

Figure 1
Figure 1. Figure 1: presents the overall workflow of our system. The pipeline consists of three main stages: external-data backbone adaptation, frozen feature extraction on s-Aff-Wild2, and task-specific expert construction. Instead of training a sin￾gle end-to-end architecture for all three ABAW11 tasks, we first establish two complementary visual representations and then construct separate downstream configurations for VA, … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 9 linked inside Pith

  1. [1]

    11th workshop and competition on affective and behavior analysis in-the- wild.https://affective-behavior-analysis-in-the-wild.github.io/11th/ (2026), accessed: 2026-07-14

    2026

  2. [2]

    arXiv preprint arXiv:1702.01992 (2017)

    Arevalo, J., Solorio, T., Montes-y Gómez, M., González, F.A.: Gated multimodal units for information fusion. arXiv preprint arXiv:1702.01992 (2017)

    Pith/arXiv arXiv 2017

  3. [3]

    arXiv preprint arXiv:1803.01271 (2018)

    Bai, S., Kolter, J.Z., Koltun, V.: An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271 (2018)

    Pith/arXiv arXiv 2018

  4. [4]

    In: Proceedings of the ACM International Conference on Multimodal Interaction (2016)

    Barsoum, E., Zhang, C., Ferrer, C.C., Zhang, Z.: Training deep networks for facial expression recognition with crowd-sourced label distribution. In: Proceedings of the ACM International Conference on Multimodal Interaction (2016)

    2016

  5. [5]

    Machine Learning28, 41–75 (1997)

    Caruana, R.: Multitask learning. Machine Learning28, 41–75 (1997)

    1997

  6. [6]

    In: International Conference on Learning Representations (2021)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: International Conference on Learning Representations (2021)

    2021

  7. [7]

    In: International Conference on Neural Information Processing (2013)

    Goodfellow, I.J., Erhan, D., Carrier, P.L., Courville, A., Mirza, M., Hamner, B., Cukierski, W., Tang, Y., Thaler, D., Lee, D.H., et al.: Challenges in representation learning: A report on three machine learning contests. In: International Conference on Neural Information Processing (2013)

    2013

  8. [8]

    In: Advances in Neural Information Processing Systems

    Ke,G.,Meng,Q.,Finley,T.,Wang,T.,Chen,W.,Ma,W.,Ye,Q.,Liu,T.Y.:Light- GBM: A highly efficient gradient boosting decision tree. In: Advances in Neural Information Processing Systems. vol. 30 (2017)

    2017

  9. [9]

    International Journal of Computer Vision (2019)

    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. International Journal of Computer Vision (2019)

    2019

  10. [10]

    arXiv preprint arXiv:1811.07770 (2018)

    Kollias,D.,Zafeiriou,S.:Aff-wild2:Extendingtheaff-wilddatabaseforaffectrecog- nition. arXiv preprint arXiv:1811.07770 (2018)

    Pith/arXiv arXiv 2018

  11. [11]

    arXiv preprint arXiv:2407.03835 (2024)

    Kollias, D., Zafeiriou, S., Kotsia, I., Dhall, A., Ghosh, S., Shao, C., Hu, G.: 7th abaw competition: Multi-task learning and compound expression recognition. arXiv preprint arXiv:2407.03835 (2024)

    Pith/arXiv arXiv 2024

  12. [12]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017)

    Li, S., Deng, W., Du, J.: Reliable crowdsourcing and deep locality-preserving learn- ing for expression recognition in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017)

    2017

  13. [13]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (2022)

    2022

  14. [14]

    In: IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops (2010)

    Lucey, P., Cohn, J.F., Kanade, T., Saragih, J., Ambadar, Z., Matthews, I.: The ex- tendedcohn-kanadedataset(ck+):Acompletedatasetforactionunitandemotion- specified expression. In: IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops (2010)

    2010

  15. [15]

    In: IEEE Transactions on Affective Computing (2017)

    Mollahosseini, A., Hasani, B., Mahoor, M.H.: Affectnet: A database for facial ex- pression, valence, and arousal computing in the wild. In: IEEE Transactions on Affective Computing (2017)

    2017

  16. [16]

    Transactions on Machine Learning Research (2024) 10 J

    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. Transactions on Machine Learning Research (2024) 10 J. Sun and Z. Gao

    2024

  17. [17]

    Information Fusion37, 98–125 (2017)

    Poria, S., Cambria, E., Bajpai, R., Hussain, A.: A review of affective comput- ing: From unimodal analysis to multimodal fusion. Information Fusion37, 98–125 (2017)

    2017

  18. [18]

    arXiv preprint arXiv:2303.09162 (2023)

    Savchenko, A.V.: Emotieffnet facial features in uni-task emotion recognition in video at abaw-5 competition. arXiv preprint arXiv:2303.09162 (2023)

    Pith/arXiv arXiv 2023

  19. [19]

    arXiv preprint arXiv:2508.10104 (2025)

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khali- dov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., et al.: Dinov3. arXiv preprint arXiv:2508.10104 (2025)

    Pith/arXiv arXiv 2025

  20. [20]

    arXiv preprint arXiv:1906.00295 (2019)

    Tsai, Y.H.H., Bai, S., Liang, P.P., Kolter, J.Z., Morency, L.P., Salakhutdinov, R.: Multimodal transformer for unaligned multimodal language sequences. arXiv preprint arXiv:1906.00295 (2019)

    Pith/arXiv arXiv 1906

  21. [21]

    arXiv preprint arXiv:1707.07250 (2017)

    Zadeh, A., Chen, M., Poria, S., Cambria, E., Morency, L.P.: Tensor fusion network for multimodal sentiment analysis. arXiv preprint arXiv:1707.07250 (2017)

    Pith/arXiv arXiv 2017

  22. [22]

    Zhang, Z., An, L., Cui, Z., Xu, A., Dong, T., Jiang, Y., Shi, J., Liu, X., Sun, X., Wang, M.: Facial affect recognition based on transformer encoder and audiovisual fusion for the abaw5 challenge. arXiv preprint arXiv:2303.09158 (2023) Task-Specific Feature Fusion 11 A Implementation Details A.1 Overall Pipeline and Feature Extraction Figure 1 presents th...

    Pith/arXiv arXiv 2023

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.