Pith. sign in

REVIEW 5 major objections 6 minor 62 references

Knowledge-Aligned Counterfactual-Enhancement Diffusion Perception for Unsupervised Cross-Domain Visual Emotion Recognition

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a frozen diffusion model, guided by captions and subject-verb-object knowledge triples, can recognize emotions across visually distant domains, and reports a 12-point accuracy gain over the previous best method.

desk verdict Strong task formulation and benchmark, but the headline gains are overstated by a baseline mismatch and the knowledge-triple pipeline is unvalidated; worth a serious referee, not acceptance yet. read the letter →

arxiv 2505.19694 v1 pith:RWTWH3OA submitted 2025-05-26 cs.CV

Wen Yin , Yong Wang , Guiduo Duan , Dongyang Zhang , Xin Hu , Yuan-Fang Li , Tao He This is my paper
classification cs.CV
keywords visualemotionrecognitionunsuperviseddomainadaptationcross-domaingeneralizationdiffusionmodelsknowledge-guidedcross-attentioncounterfactualcontrastivelearningpseudo-labelingmixtureofexperts
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

The paper introduces Unsupervised Cross-Domain Visual Emotion Recognition (UCDVER), where a model trained on labeled realistic images must recognize emotions in an unlabeled target domain such as stickers, abstract paintings, or art photos. It identifies two obstacles: emotional expression variability and affective distribution shift. To overcome them, it proposes KCDP, which turns image captions into subject-verb-object knowledge triples, uses those triples to guide a frozen latent diffusion model through low-rank cross-attention, and combines the resulting visual features with text knowledge in a mixture-of-experts predictor. A counterfactual contrastive alignment step generates pseudo-labels for the target domain. On a six-dataset benchmark, the paper reports an average accuracy of 46.16% versus 34.00% for the previous best method.

What carries the argument

The load-bearing mechanism is knowledge-guided cross-attention (KGCA) inside a latent diffusion U-Net: knowledge embeddings, formed by concatenating the CLIP encodings of the caption and the triples, are injected as keys and values into cross-attention layers, and only low-rank LoRA matrices are trained so the pretrained U-Net's stored knowledge is preserved. Multi-level attention maps and hidden features are merged by a feature pyramid network into the enhanced visual representation that feeds a mixture-of-experts predictor, which routes between visual-only and knowledge-augmented features. The second mechanism, CLIEA, builds a causal graph over the image, prompt, knowledge, visual feature, and prediction; it then creates counterfactual prompts, estimates the total indirect effect of the prompt on alignment, and uses contrastive learning on that effect to generate the target-domain pseudo-labels.

What would settle it

On the abstract-painting subset, count how many images yield knowledge triples that contain an emotion-related verb or object; then ablate the method by replacing the triples with random noun phrases as the diffusion condition. If accuracy on that domain does not drop, or if most triples are empty, the claim that knowledge triples drive the transfer is not supported.

Watch

Extended reading notes

Core claim

The central claim is that a domain-agnostic knowledge space—captions and subject-verb-object triples extracted by a vision-language captioner and a semantic-role parser—can serve as the bridge that lets a pretrained text-to-image diffusion model perceive emotion in visually alien domains. The paper demonstrates this by guiding the denoising U-Net's cross-attention with knowledge embeddings while keeping the U-Net frozen and adapting only low-rank matrices, then fusing multi-level attention maps into an enhanced visual feature. A counterfactual contrastive alignment estimates the total indirect effect of emotional prompts on image-text similarity, yielding higher-quality pseudo-labels for the unlabeled target domain. The paper reports that this framework outperforms all compared vision-emotion, domain-adaptation, and vision-language models on the new UCDVER benchmark.

Load-bearing premise

The entire approach assumes the automatic captioner and semantic-role parser produce emotion-relevant subject-verb-object triples for every domain; if those triples are noisy or meaningless for abstract or stylized images, both the diffusion guidance and the pseudo-labels are corrupted and the reported cross-domain gains would not hold.

Editorial extensions

If this is right

  • A model trained only on realistic photos can label stickers, cartoons, and other stylized content without any target-domain labels, making emotion recognition for messaging and social media cheaper to deploy.
  • Because the knowledge is textual subject-verb-object triples, errors become traceable: a mistake can be linked to a specific cue rather than to a black-box visual feature.
  • Since the diffusion backbone stays frozen and only low-rank matrices are tuned, adapting to a new target domain requires small training cost and few epochs.
  • The benchmark provides six source-to-target combinations across realistic, sticker, abstract, art-photo, and universal domains, giving future work a standard evaluation for unsupervised cross-domain emotion recognition.

Reading between the lines

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

  • A natural next step, not tested in the paper, would be swapping the semantic-role parser for an emotion-specific parser that handles metaphors and non-literal content, which the abstract-painting results suggest would be the most fragile link.
  • The counterfactual contrastive objective could be applied to other abstract visual-linguistic judgments, such as humor or aesthetics, where CLIP-style similarity is known to be brittle; the paper does not explore this.
  • If the captioner is upgraded to a stronger vision-language model, the same pipeline might transfer to even more extreme domains such as memes or emoji-only images without architectural change.
  • The causal-graph framing predicts that pseudo-labels improve only when the prompt's indirect effect is the dominant path; in domains where visual features alone carry most of the emotion signal, the contrastive term would add little.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper introduces a new task, Unsupervised Cross-Domain Visual Emotion Recognition (UCDVER), and proposes the Knowledge-aligned Counterfactual-enhancement Diffusion Perception (KCDP) framework to address it. KCDP combines a Knowledge-Alignment Diffusion Affective Perception (KADAP) module, which extracts image captions and knowledge triples to guide a frozen diffusion denoising U-Net via LoRA-based cross-attention and then predicts emotions with a Mixture-of-Experts classifier, with a Counterfactual-Enhanced Language-Image Emotional Alignment (CLIEA) module that generates pseudo-labels for the target domain through a causal-graph-inspired contrastive objective. Experiments are reported on Emoset, EmotionROI, SER30K, abstract paintings, art photos, and Emo8, with claims of large improvements over prior VER, domain-adaptation, and vision-language baselines, including a 12% gain over TGCA-PVT in the DA setting.

Significance. If the results are reproducible, the paper makes a useful contribution by defining and benchmarking the UCDVER task and by exploring structured knowledge and diffusion models for emotion recognition across very different visual domains. The idea of using domain-agnostic knowledge triples as conditioning for a diffusion-based perception model is interesting and worth investigating. However, the evidence as presented is weakened by several load-bearing issues: the headline comparison in Table 1 mixes protocols (source-only baselines vs. a method trained with unlabeled target data), Eq. (10) does not match the stated TIE objective, Table 4 numerically conflicts with Table 1, and the knowledge-extraction pipeline is never validated despite being essential to both KADAP and CLIEA. No error bars, repeated-run statistics, or code are provided. These issues must be addressed before the claimed state-of-the-art results can be accepted; the framework's novelty currently depends on an unmeasured component.

major comments (5)
  1. [§5.3, Table 1] The headline comparison is not apples-to-apples. The VER baselines MDAN*, LORA-V*, and TGCA-PVT* are explicitly marked as trained only on the source domain, whereas KCDP is trained on both source and unlabeled target data in the DA setting. The abstract's "12% improvement over TGCA-PVT" therefore reflects a difference in data access, not just method. Please rerun these baselines under the same DA protocol (or clearly separate source-only generalization from DA) and report the DA-protocol gap to UniMoS separately; the current presentation overstates the contribution.
  2. [§4.2, Eq. (10)] The loss does not implement the stated TIE objective. Eq. (8) defines TIE as the difference between factual and counterfactual prompt outcomes, and the text says CCL "pushes the feature of p away from p*". However, Eq. (10) is a conventional softmax contrastive loss over the K emotional classes; no counterfactual prompt p* appears in the loss, and the symbol y(i) is undefined. Please either rewrite Eq. (10) to include counterfactual samples and derive it from Eq. (8), or revise the description to match the actual loss.
  3. [§5.4, Table 4] The final row ("✓ ✓ MoEv+k") reports S=72.84 and E=37.77 for the S→E DA setting, but Table 1 reports KCDP S→E accuracy as 41.29. Since both tables are supposed to describe the same DA setting, this is a direct numerical conflict. Please clarify which configuration produced Table 1 and correct the discrepancy; without this, the ablation results do not validate the full model.
  4. [§4.1, §4.2, Table 4] The load-bearing premise that InstructBLIP captions plus Allennlp SRL triples provide domain-agnostic, emotionally relevant knowledge is not validated. In particular, abstract paintings are described as having "no contextual content" (Section 5.1), so SRL triples derived from captions of such images are likely to be noisy or emotion-irrelevant. Table 4 shows that adding triples over captions alone improves target accuracy by only +0.82 (35.95 to 36.77) in S→E, which is far smaller than the narrative suggests and may be within noise. Please report a quantitative evaluation of triple extraction (e.g., human agreement or a small labeled validation set) and ablations that isolate whether the gains come from the knowledge stream or from the pseudo-labeling/MoE components; otherwise the proposed mechanism is not supported by the evidence.
  5. [§5.1, §5.3, §5.4] No error bars, confidence intervals, or repeated-run statistics are reported. Several target datasets are very small (abstract: 279 images; artphoto: 806 images), and reported margins such as the +0.82 triple gain and the ~3% UC gains could be within run-to-run variance. Please report mean and standard deviation over at least three runs and, for the key comparisons (KCDP vs UniMoS and KCDP vs TGCA-PVT), a significance test or confidence interval.
minor comments (6)
  1. [§4.1, Eq. (3)-(4)] The LoRA description contains a typo: "finetuning the W(i)_K and W(i)_K" should be "W(i)_K and W(i)_V". Also, Eq. (4) only shows the update for W_K; the corresponding update for W_V should be stated explicitly.
  2. [Eq. (2)] The norm has mismatched parentheses: "∥ϵ−ϵ θ(zt, t, c∥2 2" should be "∥ϵ − ϵθ(zt, t, c)∥_2^2".
  3. [§4.2, Eq. (10)] The symbol y(i) is used in the denominator but never defined, and K is used both as the number of classes here and as the knowledge parser output in Section 4.1. Please disambiguate these notations.
  4. [§5.2] The hyperparameter sentence "We set the initial learning rate, batch size, decay interval, λ1, and λ2 to 1e-5, 16, 0.01, 1, and 1" does not say what "decay interval" refers to (weight decay? learning-rate schedule?) and gives no values for N, TopK, LoRA rank r, or temperature τ. Please complete the implementation details.
  5. [Figure 2] The symbols "🔥" and "❄" are not explained in the caption or text; please add a legend or state in the caption which components are frozen and which are trainable.
  6. [§1, §4.1, §5.3, References] Several prose and formatting issues: "off-the-shell" should be "off-the-shelf" (Section 4.1); "in the form ofsubject-verb-object" is missing a space (Section 1); "on average14% better" should be "12.1 percentage points" or "35% relative improvement" to match Table 1; and reference [4]'s page range "107 – 91" appears reversed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity is found: the central results are empirical, externally benchmarked, and the pseudo-labeling loop is a standard self-training component rather than a definitional reduction.

full rationale

KCDP's contribution is an empirical framework evaluated on external datasets (EmoSet, SER30K, abstract, artphoto, Emo8). The derivation chain from knowledge extraction (BLIP/InstructBLIP plus Allennlp SRL) to knowledge-guided diffusion (KGCA/LoRA) and MoE classification is not equivalent to its inputs: source labels enter through L_s_ce, target pseudo-labels through L_t_ce, and the counterfactual contrastive loss through L_ccl. No equation defines the reported target accuracy as the same quantity used for fitting. CLIEA pseudo-labels (Eq. 11) are computed from visual and knowledge features and used as training targets (Eq. 12); this is a standard self-training design, and the final reported accuracy is measured against held-out ground-truth target labels, so it is not forced by construction. The paper's self-citations ([12], [13], [40]) appear only in related-work context and are not load-bearing. The comparison caveat that VER baselines are source-only while KCDP uses target data is a benchmark-fairness concern, not circularity; it does not make the derivation self-referential.

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

The central claim rests on the reliability of off-the-shelf captioning and knowledge parsing, the validity of the causal-graph story, and the accuracy of CLIEA pseudo-labels; the paper provides no independent validation of these premises, so they are treated as unpaid assumptions.

free parameters (4)
  • lambda_1, lambda_2 (loss weights) = 1, 1
    Set manually in Eq (13); no sensitivity analysis; central training loss.
  • tau (temperature in L_ccl) = not reported
    Appears in Eq (10); hand-set; affects pseudo-label sharpness.
  • LoRA rank r = not reported
    Eq (4); controls adaptation capacity of cross-attention.
  • N and TopK (MoE experts) = not reported
    Eqs (6)-(7); router architecture unspecified.
assumptions (4)
  • domain assumption Knowledge triplets extracted from BLIP captions and SRL are domain-agnostic and emotion-relevant.
    Section 4.1 Textual Knowledge Extraction; no evaluation of triplet noise or emotion relevance across domains.
  • domain assumption Emotional Causality theory [4] describes the source-target shift (external event vs emotional state/physiological response).
    Section 1, Figure 1; motivates why knowledge triples should bridge domains.
  • ad hoc to paper Pseudo-labels from CLIEA are accurate enough to be used as target ground truth in Eq (12).
    Section 4.3; no pseudo-label confidence threshold or noise modeling.
  • ad hoc to paper Frozen UNet plus LoRA preserves generalization while adapting to emotion.
    Section 4.1; empirical conjecture in Figure 4, no theoretical support.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Knowledge-Aligned Counterfactual-Enhancement Diffusion Perception for Unsupervised Cross-Domain Visual Emotion Recognition." pith.science (2026). https://pith.science/paper/RWTWH3OA

@misc{pith2026250519694,
  author       = {Pith},
  title        = {Pith review of: Knowledge-Aligned Counterfactual-Enhancement Diffusion Perception for Unsupervised Cross-Domain Visual Emotion Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RWTWH3OA}},
  note         = {Machine review of arXiv:2505.19694}
}
read the original abstract

Visual Emotion Recognition (VER) is a critical yet challenging task aimed at inferring emotional states of individuals based on visual cues. However, existing works focus on single domains, e.g., realistic images or stickers, limiting VER models' cross-domain generalizability. To fill this gap, we introduce an Unsupervised Cross-Domain Visual Emotion Recognition (UCDVER) task, which aims to generalize visual emotion recognition from the source domain (e.g., realistic images) to the low-resource target domain (e.g., stickers) in an unsupervised manner. Compared to the conventional unsupervised domain adaptation problems, UCDVER presents two key challenges: a significant emotional expression variability and an affective distribution shift. To mitigate these issues, we propose the Knowledge-aligned Counterfactual-enhancement Diffusion Perception (KCDP) framework. Specifically, KCDP leverages a VLM to align emotional representations in a shared knowledge space and guides diffusion models for improved visual affective perception. Furthermore, a Counterfactual-Enhanced Language-image Emotional Alignment (CLIEA) method generates high-quality pseudo-labels for the target domain. Extensive experiments demonstrate that our model surpasses SOTA models in both perceptibility and generalization, e.g., gaining 12% improvements over the SOTA VER model TGCA-PVT. The project page is at https://yinwen2019.github.io/ucdver.

Figures

Figures reproduced from arXiv: 2505.19694 by the authors.

Figure 1
Figure 1. Illustration of two significant challenges of UCDVER [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of KCDP, comprising KADAP (§4.1, green box), CLIEA (§4.2, pink box) and Optimization (§4.3, gray box). Specifi￾cally, KADAP uses a BLIP to obtain captions and then employs a knowledge parser to extract knowledge triples. We leverage the CLIP text encoder to encode the knowledge, whose embeddings guide the denoising network via a cross-attention component. The final visual and knowledge representations are f… view at source ↗
Figure 3
Figure 3. Detail of our CLIEA causal graph. (a) Factual causality [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Visualization of cross-domain embedding versus base [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 54 canonical work pages

  1. [1]

    One-shot unsupervised domain adaptation with personalized diffusion models

    Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalo- geiton, and St ´ephane Lathuili `ere. One-shot unsupervised domain adaptation with personalized diffusion models. In CVPR, pages 698–708, 2023. 3

    work page 2023

  2. [2]

    Uniemox: Cross- modal semantic-guided large-scale pretraining for universal scene emotion perception, 2024

    Chuang Chen, Xiao Sun, and Zhi Liu. Uniemox: Cross- modal semantic-guided large-scale pretraining for universal scene emotion perception, 2024. 6

    work page 2024

  3. [3]

    TGCA-PVT: Topic-guided context-aware pyramid vision transformer for sticker emotion recognition

    Jian Chen, Wei Wang, Yuzhu Hu, Junxin Chen, Han Liu, and Xiping Hu. TGCA-PVT: Topic-guided context-aware pyramid vision transformer for sticker emotion recognition. InACM MM, 2024. 6, 7

    work page 2024

  4. [4]

    Perceiving causality in character perception: A metaphorical study of causation in film.Metaphor and Symbol, 31:107 – 91, 2016

    Maarten Co ¨egnarts and Peter Kravanja. Perceiving causality in character perception: A metaphorical study of causation in film.Metaphor and Symbol, 31:107 – 91, 2016. 1, 2

    work page 2016

  5. [5]

    Instructblip: towards general-purpose vision-language models with instruction tuning

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: towards general-purpose vision-language models with instruction tuning. InNeurIPS, Red Hook, NY , USA, 2024. Curran Associates Inc. 1, 4, 6, 7

    work page 2024

  6. [6]

    Domain-agnostic mutual prompting for unsu- pervised domain adaptation

    Zhekai Du, Xinyao Li, Fengling Li, Ke Lu, Lei Zhu, and Jingjing Li. Domain-agnostic mutual prompting for unsu- pervised domain adaptation. InCVPR, pages 23375–23384. IEEE, 2024. 2, 3, 5, 6, 7

    work page 2024

  7. [7]

    Boosting object detection with zero-shot day-night domain adaptation

    Zhipeng Du, Miaojing Shi, and Jiankang Deng. Boosting object detection with zero-shot day-night domain adaptation. InCVPR, pages 12666–12676. IEEE, 2024. 2, 3

    work page 2024

  8. [8]

    Csda: Learning category-scale joint feature for domain adaptive object detection

    Changlong Gao, Chengxu Liu, Yujie Dun, and Xueming Qian. Csda: Learning category-scale joint feature for domain adaptive object detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 11421– 11430, 2023. 2

    work page 2023

Show all 62 references
  1. [9]

    Liu, Matthew Peters, Michael Schmitz, and Luke S

    Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew Peters, Michael Schmitz, and Luke S. Zettlemoyer. Allennlp: A deep se- mantic natural language processing platform. 2017. 4

    2017

  2. [10]

    Kingdom: Knowledge-guided domain adaptation for senti- ment analysis.ACL, pages 3198–3210, 2020

    Deepanway Ghosal, Devamanyu Hazarika, Abhinaba Roy, Navonil Majumder, Rada Mihalcea, and Soujanya Poria. Kingdom: Knowledge-guided domain adaptation for senti- ment analysis.ACL, pages 3198–3210, 2020. 2

    2020

  3. [11]

    Prompting diffusion representations for cross-domain semantic segmentation.arXiv preprint arXiv:2307.02138, 2023

    Rui Gong, Martin Danelljan, Han Sun, Julio Delgado Man- gas, and Luc Van Gool. Prompting diffusion representations for cross-domain semantic segmentation.arXiv preprint arXiv:2307.02138, 2023. 3

    2023 arXiv

  4. [12]

    One network for multi-domains: Domain adaptive hashing with intersectant generative adversarial net- work.IJCAI, page 2477–2483, 2019

    Tao He, Yuan-Fang Li, Lianli Gao, Dongxiang Zhang, and Jingkuan Song. One network for multi-domains: Domain adaptive hashing with intersectant generative adversarial net- work.IJCAI, page 2477–2483, 2019. 3

    2019

  5. [13]

    To- wards open-vocabulary scene graph generation with prompt- based finetuning

    Tao He, Lianli Gao, Jingkuan Song, and Yuan-Fang Li. To- wards open-vocabulary scene graph generation with prompt- based finetuning. InECCV, pages 56–73, 2022. 5

    2022

  6. [14]

    Denoising diffu- sion probabilistic models.NeurIPS, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models.NeurIPS, 2020. 3

    2020

  7. [15]

    Lora: Low-rank adaptation of large language models.ICLR,

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.ICLR,

  8. [16]

    Busson, ´Alan Livio Guedes, S´ergio Colcher, and Ruy Luiz Milidi´u

    Ivan Jesus, Jessica Cardoso, Antonio Jose G. Busson, ´Alan Livio Guedes, S´ergio Colcher, and Ruy Luiz Milidi´u. A cnn-based tool to index emotion on anime character stickers. InISM, pages 319–3193, 2019. 2

    2019

  9. [17]

    Dgin- style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control

    Yuru Jia, Lukas Hoyer, Shengyu Huang, Tianfu Wang, Luc Van Gool, Konrad Schindler, and Anton Obukhov. Dgin- style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In European Conference on Computer Vision, pages 91–109. Spr...

    2025

  10. [18]

    Difnet: Semantic segmentation by diffusion networks.NeurIPS, 31, 2018

    Peng Jiang, Fanglin Gu, Yunhai Wang, Changhe Tu, and Baoquan Chen. Difnet: Semantic segmentation by diffusion networks.NeurIPS, 31, 2018. 3

    2018

  11. [19]

    Podia- 3d: Domain adaptation of 3d generative model across large domain gap using pose-preserved text-to-image diffusion

    Gwanghyun Kim, Ji Ha Jang, and Se Young Chun. Podia- 3d: Domain adaptation of 3d generative model across large domain gap using pose-preserved text-to-image diffusion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22603–22612, 2023. 2

    2023

  12. [20]

    Text-image align- ment for diffusion-based perception

    Neehar Kondapaneni, Markus Marks, Manuel Knott, Rog´erio Guimar ˜aes, and Pietro Perona. Text-image align- ment for diffusion-based perception. InCVPR, pages 13883– 13893. IEEE, 2024. 3

    2024

  13. [21]

    Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation

    Zhengfeng Lai, Noranart Vesdapunt, Ning Zhou, Jun Wu, Cong Phuoc Huynh, Xuelu Li, Kah Kuen Fu, and Chen-Nee Chuah. Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation. InProceedings of the IEEE/CVF International Conference on Computer Vi...

    2023

  14. [22]

    Diffusiongan3d: Boosting text-guided 3d gen- eration and domain adaptation by combining 3d gans and diffusion priors

    Biwen Lei, Kai Yu, Mengyang Feng, Miaomiao Cui, and Xu- ansong Xie. Diffusiongan3d: Boosting text-guided 3d gen- eration and domain adaptation by combining 3d gans and diffusion priors. InCVPR, pages 10487–10497, 2024. 2, 3

    2024

  15. [23]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 1, 2, 6, 7

    2023

  16. [24]

    Split to merge: Unifying separated modalities for unsupervised domain adaptation

    Xinyao Li, Yuke Li, Zhekai Du, Fengling Li, Ke Lu, and Jingjing Li. Split to merge: Unifying separated modalities for unsupervised domain adaptation. InCVPR, pages 23364– 23374, 2024. 2, 3, 5, 6, 7

    2024

  17. [25]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Doll ´ar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. InCVPR, pages 2117–2125,

  18. [26]

    Visual instruction tuning.NeurIPS, 36, 2024

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.NeurIPS, 36, 2024. 1, 7

    2024

  19. [27]

    SER30K: A large-scale dataset for sticker emotion recognition

    Shengzhe Liu, Xin Zhang, and Jufeng Yang. SER30K: A large-scale dataset for sticker emotion recognition. InACM MM, pages 33–41. ACM, 2022. 2, 6, 7

    2022

  20. [28]

    Affective image classi- fication using features inspired by psychology and art theory

    Jana Machajdik and Allan Hanbury. Affective image classi- fication using features inspired by psychology and art theory. InACMMM, pages 83–92, 2010. 4, 6

    2010

  21. [29]

    Schmidt, and Tim Fingscheidt

    Joshua Niemeijer, Manuel Schwonberg, Jan-Aike Term¨ohlen, Nico M. Schmidt, and Tim Fingscheidt. Generalization by adaptation: Diffusion-based domain extension for domain-generalized semantic segmentation. In WACV, pages 2818–2828. IEEE, 2024. 3

    2024

  22. [30]

    Roy-Chowdhury

    Rameswar Panda, Jianming Zhang, Haoxiang Li, Joon- Young Lee, Xin Lu, and Amit K. Roy-Chowdhury. Con- templating visual emotions: Understanding and overcoming dataset bias. InECCV, pages 594–612. Springer, 2018. 2

    2018

  23. [31]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zem- ing Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 32, 2019. 6

    2019

  24. [32]

    Models, reasoning and inference.Cam- bridge, UK: CambridgeUniversityPress, 19(2):3, 2000

    Judea Pearl et al. Models, reasoning and inference.Cam- bridge, UK: CambridgeUniversityPress, 19(2):3, 2000. 5

    2000

  25. [33]

    Flype: Multitask prompt tuning for multimodal human understanding of social media

    Bohua Peng, Chengfu Wu, Wei He, William Thorne, Aline Villavicencio, Yujin Wang, and Aline Paes. Flype: Multitask prompt tuning for multimodal human understanding of social media. InMUWS@ CIKM, pages 18–33, 2023. 2

    2023

  26. [34]

    Gallagher

    Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and An- drew C. Gallagher. A mixed bag of emotions: Model, pre- dict, and transfer emotion distributions. InCVPR, pages 860–868. IEEE Computer Society, 2015. 1, 2, 6

    2015

  27. [35]

    Where do emotions come from? predicting the emotion stimuli map

    Kuan-Chuan Peng, Amir Sadovnik, Andrew Gallagher, and Tsuhan Chen. Where do emotions come from? predicting the emotion stimuli map. In2016 IEEE international con- ference on image processing (ICIP), pages 614–618. IEEE,

  28. [36]

    Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents.arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3

    2022 arXiv

  29. [37]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InCVPR, pages 10674– 10685. IEEE, 2022. 2, 3

    2022

  30. [38]

    Rosin, and Liang Wang

    Dongyu She, Jufeng Yang, Ming-Ming Cheng, Yu-Kun Lai, Paul L. Rosin, and Liang Wang. Wscnet: Weakly supervised coupled networks for visual sentiment classification and de- tection.IEEE TMM, 22(5):1358–1371, 2020. 1, 2

    2020

  31. [39]

    Ad-clip: Adapting domains in prompt space using clip

    Mainak Singha, Harsh Pal, Ankit Jha, and Biplab Banerjee. Ad-clip: Adapting domains in prompt space using clip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4355–4364, 2023. 6, 7

    2023

  32. [40]

    Deep discrete hashing with self-supervised pairwise labels

    Jingkuan Song, Tao He, Hangbo Fan, and Lianli Gao. Deep discrete hashing with self-supervised pairwise labels. InMa- chine Learning and Knowledge Discovery in Databases: Eu- ropean Conference, ECML PKDD 2017, Skopje, Macedonia, September 18–22, 2017, Proceedings, Part I 10, pages 223–

    2017

  33. [41]

    SKEP: Sentiment knowl- edge enhanced pre-training for sentiment analysis

    Hao Tian, Can Gao, Xinyan Xiao, Hao Liu, Bolei He, Hua Wu, Haifeng Wang, and Feng Wu. SKEP: Sentiment knowl- edge enhanced pre-training for sentiment analysis. InACL, pages 4067–4076, Online, 2020. 2

    2020

  34. [42]

    Neural discrete representation learning.NeurIPS, 30, 2017

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning.NeurIPS, 30, 2017. 4

    2017

  35. [43]

    Visualizing data using t-sne.JMLR, 9(11), 2008

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.JMLR, 9(11), 2008. 8

    2008

  36. [44]

    Attention is all you need.NeurIPS, 30, 2017

    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, 30, 2017. 5

    2017

  37. [45]

    Metaalign: Coordinating domain alignment and classifica- tion for unsupervised domain adaptation

    Guoqiang Wei, Cuiling Lan, Wenjun Zeng, and Zhibo Chen. Metaalign: Coordinating domain alignment and classifica- tion for unsupervised domain adaptation. InCVPR, pages 16643–16653, 2021. 2, 3

    2021

  38. [46]

    Imagine that! abstract-to-intricate text-to-image syn- thesis with scene graph hallucination diffusion.NeurIPS, 36,

    Shengqiong Wu, Hao Fei, Hanwang Zhang, and Tat-Seng Chua. Imagine that! abstract-to-intricate text-to-image syn- thesis with scene graph hallucination diffusion.NeurIPS, 36,

  39. [47]

    Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using dif- fusion models

    Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using dif- fusion models. InICCV, pages 1206–1217, 2023. 3

    2023

  40. [48]

    Emovit: Revolutionizing emotion insights with vi- sual instruction tuning

    Hongxia Xie, Chu-Jun Peng, Yu-Wen Tseng, Hung-Jen Chen, Chan-Feng Hsu, Hong-Han Shuai, and Wen-Huang Cheng. Emovit: Revolutionizing emotion insights with vi- sual instruction tuning. InCVPR, 2024. 1, 6, 7

    2024

  41. [49]

    MDAN: multi-level dependent attention network for visual emotion analysis

    Liwen Xu, Zhengtao Wang, Bin Wu, and Simon Lui. MDAN: multi-level dependent attention network for visual emotion analysis. InCVPR, pages 9469–9478. IEEE, 2022. 1, 6, 7

    2022

  42. [50]

    Robust emotion recog- nition in context debiasing

    Dingkang Yang, Kun Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, and Lihua Zhang. Robust emotion recog- nition in context debiasing. InCVPR, pages 12447–12457,

  43. [51]

    Stimuli-aware visual emotion analysis.IEEE TIP, 30:7432–7445, 2021

    Jingyuan Yang, Jie Li, Xiumei Wang, Yuxuan Ding, and Xinbo Gao. Stimuli-aware visual emotion analysis.IEEE TIP, 30:7432–7445, 2021. 1

    2021

  44. [52]

    Emoset: A large- scale visual emotion dataset with rich attributes

    Jingyuan Yang, Qirui Huang, Tingting Ding, Dani Lischin- ski, Daniel Cohen-Or, and Hui Huang. Emoset: A large- scale visual emotion dataset with rich attributes. InICCV, pages 20326–20337. IEEE, 2023. 1, 2, 6

    2023

  45. [53]

    Emogen: Emo- tional image content generation with text-to-image diffusion models

    Jingyuan Yang, Jiawei Feng, and Hui Huang. Emogen: Emo- tional image content generation with text-to-image diffusion models. InCVPR, pages 6358–6368. IEEE, 2024. 3, 5

    2024

  46. [54]

    Visual sentiment analysis by attending on local image regions

    Quanzeng You, Hailin Jin, and Jiebo Luo. Visual sentiment analysis by attending on local image regions. InAAAI, pages 231–237. AAAI Press, 2017. 1, 2

    2017

  47. [55]

    Zero-shot emotion recognition via affec- tive structural embedding

    Chi Zhan, Dongyu She, Sicheng Zhao, Ming-Ming Cheng, and Jufeng Yang. Zero-shot emotion recognition via affec- tive structural embedding. InICCV, pages 1151–1160. IEEE,

  48. [56]

    Learning emo- tion representations from verbal and nonverbal communica- tion

    Sitao Zhang, Yimu Pan, and James Z Wang. Learning emo- tion representations from verbal and nonverbal communica- tion. InCVPR, pages 18993–19004, 2023. 1

    2023

  49. [57]

    Pdanet: Polarity-consistent deep attention network for fine-grained visual emotion regression

    Sicheng Zhao, Zizhou Jia, Hui Chen, Leida Li, Guiguang Ding, and Kurt Keutzer. Pdanet: Polarity-consistent deep attention network for fine-grained visual emotion regression. InACM MM, pages 192–201. ACM, 2019. 1, 2

    2019

  50. [58]

    Cycleemotiongan: Emotional semantic consistency preserved cyclegan for adapting image emotions

    Sicheng Zhao, Chuang Lin, Pengfei Xu, Sendong Zhao, Yuchen Guo, Ravi Krishna, Guiguang Ding, and Kurt Keutzer. Cycleemotiongan: Emotional semantic consistency preserved cyclegan for adapting image emotions. InAAAI, pages 2620–2627, 2019. 2

    2019

  51. [59]

    Sticker820k: Empowering interactive retrieval with stickers.CoRR, abs/2306.06870,

    Sijie Zhao, Yixiao Ge, Zhongang Qi, Lin Song, Xiaohan Ding, Zehua Xie, and Ying Shan. Sticker820k: Empowering interactive retrieval with stickers.CoRR, abs/2306.06870,

    arXiv

  52. [60]

    Unleashing text-to-image diffusion models for visual perception

    Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. InICCV, pages 5706–5716. IEEE, 2023. 3

    2023

  53. [61]

    A fa- cial expression-aware multimodal multi-task learning frame- work for emotion recognition in multi-party conversations

    Wenjie Zheng, Jianfei Yu, Rui Xia, and Shijin Wang. A fa- cial expression-aware multimodal multi-task learning frame- work for emotion recognition in multi-party conversations. InACL, pages 15445–15459, 2023. 1

    2023

  54. [62]

    Bridge then begin anew: Generating target-relevant interme- diate model for source-free visual emotion adaptation.AAAI,

    Jiankun Zhu, Sicheng Zhao, Jing Jiang, Wenbo Tang, Zhaopan Xu, Tingting Han, Pengfei Xu, and Hongxun Yao. Bridge then begin anew: Generating target-relevant interme- diate model for source-free visual emotion adaptation.AAAI,

Pith tools

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