Pith. sign in

REVIEW 4 major objections 4 minor 26 references

Team RAS in 9th ABAW Competition: Multimodal Compound Expression Recognition Approach

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A zero-shot six-modality pipeline recognizes compound emotions at levels comparable to supervised systems.

desk verdict A promising zero-shot CER pipeline whose headline numbers don't match its own table and whose scene-model training is ambiguous enough to threaten the zero-shot claim. read the letter →

arxiv 2507.02205 v2 pith:CXPTYDDR submitted 2025-07-02 cs.CV

classification cs.CV
keywords compoundexpressionrecognitionzero-shotlearningmultimodalemotionvision-languagemodelmulti-headprobabilityfusionaffectivecomputingscenedescriptionfacial
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 tries to establish that compound expressions—emotional states such as Happily Surprised or Sadly Fearful—can be recognized without any training on compound-expression labels. The authors assemble a six-modality pipeline (static and dynamic face, scene description, scene-label matching, audio, and text) whose components are either pretrained or trained only on eight basic-emotion labels, then fuse them with a learned weighting module and convert basic-emotion outputs into seven compound categories by pair-wise probability or feature aggregation. On the ABAW benchmark they report F1 of 46.95% on AffWild2, 49.02% on AFEW, and 34.85% on C-EXPR-DB, which they claim is comparable to supervised systems trained on target data. If correct, this would make compound-emotion recognition practical for new domains without collecting new labels, and would identify scene descriptions as the most informative modality for this task.

What carries the argument

The load-bearing machinery has three parts. First, Qwen-VL 2.5 turns each four-second video segment into a free-form textual scene description, which is then encoded by CLIP, DistilRoBERTa, or Jina embeddings and classified into eight basic-emotion labels; this scene modality alone produces the best C-EXPR-DB result. Second, the Multi-Head Probability Fusion (MHPF) module learns class-specific convex weights across modalities to combine probability distributions from the face, scene, audio, text, and label-matching models. Third, the Compound Expressions transformation maps eight basic-emotion outputs onto seven compound labels either by Pair-Wise Probability Aggregation (PPA), which sums the probabilities of the constituent basic emotions, or by Pair-Wise Feature Similarity Aggregation (PFSA), which averages basic-emotion feature prototypes in the latent space and matches test features by cosine similarity with temperature-scaled softmax.

What would settle it

Inspect the training logs and corpus annotations used in Section 3.3: if any C-EXPR-DB compound labels appear among the descriptions' training targets, the 34.85% F1 is not zero-shot. Separately, on a corpus with ground-truth basic and compound labels, compute whether the PPA pair-wise sum of predicted basic emotion probabilities matches the compound labels substantially better than a random or learned mapping would.

Watch

Extended reading notes

Core claim

The central claim is that a zero-shot pipeline can match supervised compound-expression recognizers. The best full pipeline achieves 46.95% F1 on AffWild2 and 49.02% on AFEW by fusing face, scene, audio, text, and scene-label modalities through the Multi-Head Probability Fusion module, while the best score on C-EXPR-DB (34.85% macro-F1) comes from a single scene model: Qwen-VL-generated descriptions of each four-second segment, encoded by a DistilRoBERTa-based emotion model, with pair-wise feature similarity aggregation used to form compound predictions. Against earlier work, the authors report that their scene model outperforms the zero-shot visual baseline (32.43%) and the zero-shot audio-visual-text baseline (25.91%), and trails only a curriculum-learning system that was fine-tuned on target compound data (60.63%).

Load-bearing premise

The claim stands or falls on the assumption that the scene-description model was trained without ever seeing target compound-expression labels, combined with the assumption that compound emotions are accurately formed by pair-wise sums of basic-emotion probabilities or feature prototypes.

Editorial extensions

If this is right

  • A new corpus with new compound labels can be handled by changing only the label text and prompt, without fine-tuning on target data.
  • Scene descriptions carry enough emotional signal that a single vision-language model can beat a full six-modality fusion on C-EXPR-DB.
  • Existing basic-emotion classifiers can be reused for compound recognition through the PPA/PFSA transformation, avoiding retraining on compound corpora.
  • Multimodal fusion helps most on large acted and in-the-wild corpora such as AffWild2 and AFEW, while its benefit on small compound-labeled corpora is less clear.

Reading between the lines

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

  • The authors do not test whether the PPA's additive assumption is true; a natural follow-up is to compare the fixed pair-wise sum against a learned composition function on a corpus annotated with both basic and compound labels.
  • The C-EXPR-DB success of the scene model suggests that descriptions may be a more transferable emotion signal than faces across domains; an ablation that feeds the same descriptions to a supervised classifier could separate the contribution of description content from the label-matching step.
  • If the zero-shot claim holds across more corpora, compound-expression recognition could be deployed for new emotion taxonomies, such as clinical or cultural categories, by editing the prompt and label set with no new annotation campaign.
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 describes a multimodal compound expression recognition (CER) pipeline for the ABAW competition setting. It combines static and dynamic facial models, CLIP-based label matching, Qwen-VL scene descriptions, audio, and text, and it introduces a Multi-Head Probability Fusion (MHPF) module plus two compound-expression mapping methods, PPA and PFSA. The authors report that the pipeline achieves macro-F1 scores of 46.95% on AffWild2, 49.02% on AFEW, and 34.85% on C-EXPR-DB without fine-tuning on target compound-expression labels, and they compare these results with prior supervised and zero-shot CER methods.

Significance. If the central claim holds, this is a useful result for affective computing: it would show that compound expressions can be recognized competitively without training on target compound labels, which matters for low-resource CER settings. The paper's strengths include a public code release, a broad set of modalities, and an explicit comparison with prior zero-shot and fine-tuned methods. However, the significance is conditional on resolving a training-protocol ambiguity for the scene model and on reconciling the reported headline numbers with the experimental table.

major comments (4)
  1. [Abstract and §5 vs. Table 1] The abstract and conclusion report F1 scores of 46.95% on AffWild2 and 49.02% on AFEW, but Table 1 row 17 reports macro-F1 values of 46.50 and 47.14 for those datasets. No row in Table 1 matches the abstract/conclusion values. Please state exactly which configuration and metric produce 46.95 and 49.02, or correct the text to match the table.
  2. [§3.3 and §4.1] The training protocol for the scene models is ambiguous and directly affects the zero-shot claim. The text says generated descriptions are 'aligned with compound emotion labels via corpus annotations' and that prompt 2 outputs one of eight basic or seven compound labels, and a linear classification head is added to each encoder. If that head is trained on C-EXPR-DB compound labels, the reported 34.85 macro-F1 on C-EXPR-DB is a supervised result, not a zero-shot result. Please specify the exact training corpus and label space for each scene model and clarify how corpus annotations are used.
  3. [§3.7] The PPA and PFSA mappings assume compound emotions are additive combinations of basic emotions, either at the probability level or in averaged feature space. This assumption is not validated against ground truth. Please add an ablation or analysis comparing PPA/PFSA with a simple learned mapping on the validation set, and report per-class performance for the compound classes, so the reader can see whether the method works uniformly or is driven by a few classes.
  4. [§4.2 and Table 1] No error bars or repeated runs are reported, and only the top three fusion combinations from an unspecified search are shown. The difference between the best C-EXPR-DB result (34.85) and the prior zero-shot baseline (32.43) may not be meaningful without variance information. Please report the number of seeds, the search space considered, and the performance distribution over those configurations.
minor comments (4)
  1. [Table 2] Table 2 lists 'Ours ZH V' with a score of 34.86, but Table 1 row 7 reports 34.85, and the model actually uses scene descriptions and a text encoder in addition to visual input, so the modality label 'V' is misleading.
  2. [§3.4 and Table 1] The audio model is described as 'U-Net-aware' in the text but is labeled 'U-aware' in Table 1 row 9; please use consistent terminology.
  3. [Figure 1] The pipeline figure is very dense, and the labels for individual modality models are hard to read; a larger figure or a tabular summary of each model's input, backbone, and classification head would improve clarity.
  4. [§1 and §5] The abstract lists five modality groups but the conclusion says six modalities, with 'probabilistic fusion' counted as a modality; please clarify whether fusion is a modality or a processing stage.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: compound mapping is not fitted to target labels; scene-model wording is ambiguous but not a demonstrated reduction.

full rationale

The paper's derivation chain is self-contained: basic-emotion classifiers are trained on AFEW and AffWild2 (§4.1), and compound outputs are obtained by PPA (summing basic-emotion probabilities) or PFSA (cosine similarity to prototypes averaged from basic-emotion prototypes) in §3.7. Neither PPA nor PFSA uses C-EXPR-DB compound labels during training; the compound prototypes are built from validation features of correctly classified samples, and no target-corpus labels enter the mapping. The §3.3 phrase 'aligned with compound emotion labels via corpus annotations' is ambiguous, but §4.1 states that models are trained on AFEW and AffWild2 and that C-EXPR-DB is used only for evaluation, and §3.7 explicitly operates on eight basic emotion labels before conversion. Thus the best C-EXPR-DB score is not shown to be a supervised result renamed as zero-shot. Self-citations to prior RAS work ([21], [22]) are contextual or component-level and are not load-bearing; no uniqueness theorem or ansatz is smuggled in via citation. The paper's main weakness is a reporting ambiguity, not circularity.

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

The central CER result rests on several assumptions about how basic emotions compose into compound emotions, the faithfulness of LLM-generated scene descriptions, and the label-alignment between corpora. The additive/mixture assumption behind PPA and PFSA is not validated with target labels, so it carries much of the interpretive load.

free parameters (4)
  • PFSA softmax temperature = not reported
    Temperature scaling in Section 3.7 controls sharpness of compound-emotion distribution; its value is not given and it is presumably tuned on validation, directly affecting the 34.85 PFSA score.
  • MHPF fusion weights = learned (not reported)
    Section 3.6 defines learnable per-head and per-class weights; these are fit to training data and determine the 46.95/49.02 headline numbers.
  • AFEW 'other' placeholder class = added as 8th class
    Section 4.1 introduces a placeholder 'other' label for AFEW to align with AffWild2; this hand-chosen mapping affects multi-corpus training and downstream CE mapping.
  • Segment window and sampling = 4 s window, 2 s overlap, 20 fps
    Section 3 states the window and frame downsampling; these are design choices from prior studies that affect temporal emotion modeling for all modalities.
assumptions (5)
  • domain assumption Compound emotions decompose as additive combinations of basic emotion probabilities (PPA).
    Section 3.7 states compound probabilities are sums of basic emotion probabilities; unvalidated against ground-truth compound labels.
  • domain assumption Compound emotion prototypes can be approximated by averaging basic-emotion prototypes in feature space (PFSA).
    Section 3.7 constructs compound prototypes by averaging basic prototypes after L2 normalization; no evidence this holds in the learned latent space.
  • domain assumption Qwen-VL generated scene descriptions faithfully capture emotional content.
    Section 3.3 uses Qwen-VL descriptions as input modality; no quality check or human evaluation of generated descriptions is reported.
  • domain assumption The 8-class label spaces of AffWild2 and AFEW are compatible after adding the 'other' placeholder.
    Section 4.1 aligns the two corpora by adding a placeholder class, assuming no semantic mismatch for the remaining classes.
  • domain assumption CLIP cosine similarity between frames and emotion label texts provides meaningful zero-shot emotion probabilities.
    Section 3.2 relies on CLIP for zero-shot classification; no calibration or validation of CLIP's emotion semantics is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Team RAS in 9th ABAW Competition: Multimodal Compound Expression Recognition Approach." pith.science (2026). https://pith.science/paper/CXPTYDDR

@misc{pith2026250702205,
  author       = {Pith},
  title        = {Pith review of: Team RAS in 9th ABAW Competition: Multimodal Compound Expression Recognition Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CXPTYDDR}},
  note         = {Machine review of arXiv:2507.02205}
}
read the original abstract

Compound Expression Recognition (CER), a subfield of affective computing, aims to detect complex emotional states formed by combinations of basic emotions. In this work, we present a novel zero-shot multimodal approach for CER that combines six heterogeneous modalities into a single pipeline: static and dynamic facial expressions, scene and label matching, scene context, audio, and text. Unlike previous approaches relying on task-specific training data, our approach uses zero-shot components, including Contrastive Language-Image Pretraining (CLIP)-based label matching and Qwen-VL for semantic scene understanding. We further introduce a Multi-Head Probability Fusion (MHPF) module that dynamically weights modality-specific predictions, followed by a Compound Expressions (CE) transformation module that uses Pair-Wise Probability Aggregation (PPA) and Pair-Wise Feature Similarity Aggregation (PFSA) methods to produce interpretable compound emotion outputs. Evaluated under multi-corpus training, the proposed approach shows F1 scores of 46.95% on AffWild2, 49.02% on Acted Facial Expressions in The Wild (AFEW), and 34.85% on C-EXPR-DB via zero-shot testing, which is comparable to the results of supervised approaches trained on target data. This demonstrates the effectiveness of the proposed approach for capturing CE without domain adaptation. The source code is publicly available.

Figures

Figures reproduced from arXiv: 2507.02205 by the authors.

Figure 1
Figure 1. Pipeline of the proposed multimodal CER approach. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 20 canonical work pages

  1. [1]

    Wavlm: Large-scale self- supervised pre-training for full stack speech processing

    Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, et al. Wavlm: Large-scale self- supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing , 16 (6):1505–1518, 2022. 4

  2. [2]

    Transformers are ssms: gener- alized models and efficient algorithms through structured state space duality

    Tri Dao and Albert Gu. Transformers are ssms: gener- alized models and efficient algorithms through structured state space duality. In International Conference on Machine Learning, 2024. 3

  3. [3]

    ABAW: Valence-arousal estimation, ex- pression recognition, action unit detection & multi-task learning challenges

    Dimitrios Kollias. ABAW: Valence-arousal estimation, ex- pression recognition, action unit detection & multi-task learning challenges. In CVPR, pages 2328–2336, 2022. 5

  4. [4]

    ABAW: Learning from synthetic data & multi-task learning challenges

    Dimitrios Kollias. ABAW: Learning from synthetic data & multi-task learning challenges. In ECCV, pages 157–172,

  5. [5]

    Multi-label compound expression recog- 6 nition: C-EXPR database & network

    Dimitrios Kollias. Multi-label compound expression recog- 6 nition: C-EXPR database & network. In CVPR, pages 5589– 5598, 2023. 1, 5

  6. [6]

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

    Dimitrios Kollias and Stefanos Zafeiriou. Affect analysis in-the-wild: Valence-arousal, expressions, action units and a unified framework. arXiv preprint arXiv:2103.15792, 2021. 5

  7. [7]

    Analysing affec- tive behavior in the second ABAW2 competition

    Dimitrios Kollias and Stefanos Zafeiriou. Analysing affec- tive behavior in the second ABAW2 competition. In CVPR, pages 3652–3660, 2021. 5

  8. [8]

    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. In International Conference on Automatic Face and Gesture Recognition (FG), pages 794–800, 2020. 5

Show all 26 references
  1. [9]

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

    Dimitrios Kollias, Viktoriia Sharmanska, and Stefanos Zafeiriou. Distribution matching for heterogeneous multi- task learning: a large-scale face study. arXiv preprint arXiv:2105.03790, 2021. 5

  2. [10]

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

    Dimitrios Kollias, Panagiotis Tzirakis, Alice Baird, Alan Cowen, and Stefanos Zafeiriou. ABAW: Valence-arousal estimation, expression recognition, action unit detection & emotional reaction intensity estimation challenges. InCVPR, pages 5888–5897, 2023. 5

  3. [11]

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

    Dimitrios Kollias, Viktoriia Sharmanska, and Stefanos Zafeiriou. Distribution matching for multi-task learning of classification tasks: a large-scale study on faces & beyond. In AAAI Conference on Artificial Intelligence , pages 2813– 2821, 2024. 5

  4. [12]

    The 6th affective behav- ior analysis in-the-wild (abaw) competition

    Dimitrios Kollias, Panagiotis Tzirakis, Alan Cowen, Ste- fanos Zafeiriou, Irene Kotsia, Alice Baird, Chris Gagne, Chunchang Shao, and Guanyu Hu. The 6th affective behav- ior analysis in-the-wild (abaw) competition. In CVPR, pages 4587–4598, 2024. 2, 5

  5. [13]

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

    Dimitrios Kollias, Stefanos Zafeiriou, Irene Kotsia, Abhinav Dhall, Shreya Ghosh, Chunchang Shao, and Guanyu Hu. 7th abaw competition: Multi-task learning and compound expression recognition. arXiv preprint arXiv:2407.03835 , pages 1–14, 2024. 5

  6. [14]

    Dvd: A comprehensive dataset for advanc- ing violence detection in real-world scenarios.arXiv preprint arXiv:2506.05372, pages 1–10, 2025

    Dimitrios Kollias, Damith C Senadeera, Jianian Zheng, Kaushal KK Yadav, Greg Slabaugh, Muhammad Awais, and Xiaoyun Yang. Dvd: A comprehensive dataset for advanc- ing violence detection in real-world scenarios.arXiv preprint arXiv:2506.05372, pages 1–10, 2025. 5

  7. [15]

    Ad- vancements in affective and behavior analysis: The 8th abaw workshop and competition

    Dimitrios Kollias, Panagiotis Tzirakis, Alan Cowen, Ste- fanos Zafeiriou, Irene Kotsia, Eric Granger, Marco Peder- soli, Simon Bacon, Alice Baird, Chris Gagne, et al. Ad- vancements in affective and behavior analysis: The 8th abaw workshop and competition. In CVPR, pages 5572–5583,

  8. [16]

    Compound expression recognition via curriculum learning

    Chen Liu, Feng Qiu, Wei Zhang, Lincheng Li, Dadong Wang, and Xin Yu. Compound expression recognition via curriculum learning. In European Conference on Computer Vision, pages 282–293. Springer, 2025. 2, 6

  9. [17]

    Abaw7 challenge: A facial affect recognition approach based on transformer encoder and multilayer perceptron

    Xuxiong Liu, Kang Shen, Jun Yao, Boyan Wang, Yu Wang, Yujie Guan, Xin Liu, Gengchen Li, Liuwei An, Zishun Cui, et al. Abaw7 challenge: A facial affect recognition approach based on transformer encoder and multilayer perceptron. In European Conference on Computer Vision, pages ...

  10. [18]

    Medi- apipe: A framework for building perception pipelines

    Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, et al. Medi- apipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, pages 1–9, 2019. 3

  11. [19]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In International Conference on Machine Learning...

  12. [20]

    Textualized and feature-based models for compound multimodal emo- tion recognition in the wild

    Nicolas Richet, Soufiane Belharbi, Haseeb Aslam, Meike Emilie Schadt, Manuela Gonz ´alez-Gonz´alez, Gustave Cortal, Alessandro Lameiras Koerich, Marco Pedersoli, Alain Finkel, Simon Bacon, et al. Textualized and feature-based models for compound multimodal emo- tion recognitio...

  13. [21]

    In search of a robust facial expressions recognition model: A large-scale visual cross-corpus study

    Elena Ryumina, Denis Dresvyanskiy, and Alexey Karpov. In search of a robust facial expressions recognition model: A large-scale visual cross-corpus study. Neurocomputing, 514: 435–450, 2022. 1, 3

  14. [22]

    Zero-shot audio-visual compound expression recognition method based on emotion probability fusion

    Elena Ryumina, Maxim Markitantov, Dmitry Ryumin, Hey- sem Kaya, and Alexey Karpov. Zero-shot audio-visual compound expression recognition method based on emotion probability fusion. In CVPRW, pages 4752–4760, 2024. 2

  15. [23]

    Smoothing predictions of multi-task emotinet models for compound facial expression recogni- tion

    Andrey V Savchenko. Smoothing predictions of multi-task emotinet models for compound facial expression recogni- tion. In European Conference on Computer Vision , pages 257–266. Springer, 2025. 2, 6

  16. [24]

    jina-embeddings-v3: Multilingual embeddings with task lora

    Saba Sturua, Isabelle Mohr, Mohammad Kalim Akram, Michael G ¨unther, Bo Wang, Markus Krimmel, Feng Wang, Georgios Mastrapas, Andreas Koukounas, Andreas Kouk- ounas, Nan Wang, and Han Xiao. jina-embeddings-v3: Multilingual embeddings with task lora. arXiv preprint arXiv:2409.10...

  17. [25]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, pages 1–11,

  18. [26]

    Soft focal loss: Evaluating sample quality for dense object detection

    Zhenyuan Wang, Xuemei Xie, Jianxiu Yang, and Guang- ming Shi. Soft focal loss: Evaluating sample quality for dense object detection. Neurocomputing, 480:271–280,

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.