Pith. sign in

REVIEW 5 major objections 5 minor 60 references

Seeding DETR queries with object-aware and CLIP-derived semantic priors improves HOI detection, raising rare-category mAP on HICO-Det from 29.25 to 32.79.

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 ↗

DQEN enhances DETR-based HOI detection by initializing object and interaction queries with encoder features and CLIP text-derived semantics, improving HICO-Det mAP but not V-COCO.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid incremental HOI work with real HICO-Det gains, but the V-COCO numbers undercut the abstract's 'competitive on both datasets' claim and several reporting issues need fixing. the 5 major comments →

arxiv 2508.18896 v1 pith:LOWDBJVM submitted 2025-08-26 cs.CV

DQEN: Dual Query Enhancement Network for DETR-based HOI Detection

Zhehao Li , Chong Wang , Yi Chen , Yinghao Lu , Jiangbo Qian , Jiong Wang , Jiafei Wu This is my paper
classification cs.CV
keywords human-object interaction detectionDETRtransformer queriesCLIPsemantic feature fusionzero-shot detectionHICO-DetV-COCO
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

The paper argues that in DETR-style end-to-end human-object interaction (HOI) detection, the decoder's randomly initialized queries are a bottleneck, and that replacing their initialization with two informed priors improves accuracy without changing the detection backbone. It proposes DQEN, which enhances object queries with the most confident object-classified encoder features and enhances interaction queries with semantic features distilled from the CLIP-ranked top-K 'a photo of a person verb-ing an object' text matches. On HICO-Det this raises full mAP from 33.75 to 35.23 and rare-category mAP from 29.25 to 32.79 relative to its GEN-VLKT baseline, and it also improves zero-shot unseen-combination performance. On the denser V-COCO benchmark the gains are much smaller, which the paper attributes to multiple co-occurring interactions diluting whole-image CLIP matching. If correct, the contribution is a general recipe: query initialization, not just feature extraction, is a place to inject external and interactive priors.

Core claim

DQEN's central claim is that both halves of the HOI query pair benefit from semantic seeding. For object queries, the top-N encoder tokens chosen by object-classifier score are added to randomly initialized object embeddings; this shifts attention onto the humans interacting with those objects, helping HOI accuracy even though plain object detection slightly worsens. For interaction queries, CLIP computes similarity between the image and every HOI text label; the top-K matching triplets become candidates, and an Interaction Semantic Fusion module separates their verb and object word embeddings, reweights them through a learned correlation matrix, and sums them into one semantic vector that i

What carries the argument

The load-bearing mechanism is query re-initialization from priors. Object Query Enhancement (OQE) selects encoder features with the top-N object-classifier logits and adds them to object queries, with gradients stopped on the encoder features to avoid over-focus on objects. Interaction Query Enhancement (IQE) uses CLIP's image-text similarity to rank HOI text labels, keeps the top-K as HOI triplet candidates, and fuses their verb and object word embeddings through an Interaction Semantic Fusion module into a single interaction semantic feature that is repeated, averaged with human and object features, and fed to the interaction decoder. The Auxiliary Prediction Unit (APU) turns that same sem

Load-bearing premise

The method assumes CLIP's whole-image similarity between the image and a label like 'a photo of a person riding a horse' ranks the true interactions highly enough that the top 16 text matches are reliable semantic priors; in dense images with several interactions this assumption visibly weakens.

What would settle it

On images that contain several different human-object interactions, compute whether the ground-truth HOI labels appear in CLIP's top 16 text matches; if coverage is low in exactly the cases where IQE helps, the link between CLIP ranking and query enhancement is broken. A second check: replace the CLIP-selected top-K triplets with random or shuffled HOI labels; if HICO-Det mAP does not fall, the semantic fusion module is not doing the claimed work.

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

If this is right

  • Rare HOI categories, which have few training examples, benefit most from semantic query seeding: rare mAP rises 3.54 points on HICO-Det default.
  • The approach works without changing the backbone or the two-decoder DETR structure, so the recipe can be dropped into other one-stage DETR-based HOI models.
  • Because the semantic priors come from CLIP text matching, the method inherits zero-shot ability: unseen-category mAP improves in the RF-UC and UV settings.
  • Training converges faster: with the query enhancements the model reaches better HICO-Det accuracy in 60 epochs than the baseline in 90.

Where Pith is reading between the lines

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

  • OQE's observed behavior — improved HOI but slightly worse object detection — suggests encoder features already encode interaction context, so the object query is acting as an interaction cue rather than a cleaner object detector; ablating OQE on images where the interacting human is far from the object would clarify the mechanism.
  • The V-COCO weakness implies a natural extension: replace whole-image CLIP similarity with region-level or pairwise visual-language matching so that multiple interactions per image do not dilute the top-K candidate set.
  • If the recipe generalizes, the same dual seeding could improve other set-prediction tasks where one query type carries relational semantics, such as scene graph generation or referring expression grounding.
  • The paper's own ablation suggests OQE and APU only cooperate when IQE is present to connect object and action spaces; an explicit test would be to feed action features into OQE or object features into APU and measure whether the conflict the authors observe disappears.
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

5 major / 5 minor

Summary. The paper proposes DQEN, a one-stage DETR-based HOI detector with two query-enhancement mechanisms: Object Query Enhancement (OQE) augments randomly initialized object queries with Top-N encoder features selected by an object classifier, and Interaction Query Enhancement (IQE) uses CLIP image-text similarity to select Top-K HOI triplet candidates, fuses their word embeddings through an Interaction Semantic Fusion module, and uses the resulting semantic feature to initialize interaction queries. An Auxiliary Prediction Unit adds a verb-classification branch. Experiments are reported on HICO-Det and V-COCO. On HICO-Det default, DQEN reaches 35.23 Full / 32.79 Rare / 35.96 Non-Rare without training-free scoring and 35.34 / 33.10 / 36.01 with it, compared with the published GEN-VLKT baseline of 33.75 / 29.25 / 35.10. On V-COCO, DQEN obtains 60.88 AP(S1) / 63.47 AP(S2), which is below the reproduced GEN-VLKT baseline (62.41 / 64.46) and below HOICLIP (63.5 / 64.8).

Significance. The core idea—initializing DETR-style queries with semantic and object-aware priors—is plausible and of interest to the HOI community, especially for rare-category improvement. The paper ships source code, reports efficiency metrics, and includes several targeted ablations (gradient truncation, word embedding initialization, module contributions), which strengthen reproducibility. However, the empirical support is uneven: the headline claim of 'competitive performance on both datasets' is contradicted by the V-COCO results in Table II, and the HICO-Det ablations are computed against a modified baseline that is weaker than the published GEN-VLKT. These issues affect the central claim and must be addressed before the results can be interpreted as stated.

major comments (5)
  1. [Abstract, Table II, Section V] The abstract states the method 'achieves competitive performance on both the HICO-Det and the V-COCO datasets.' Table II shows the opposite for V-COCO: DQEN (60.88 S1 / 63.47 S2) is below the GEN-VLKT baseline reported in the same table (62.41 / 64.46) and below HOICLIP (63.5 / 64.8). Section V acknowledges this degradation but it is not reflected in the abstract or conclusion. The paper should either report V-COCO as a limitation or remove the claim of competitive performance on both datasets.
  2. [Section IV-B, Table IV] The ablation baseline in Table IV is not the published GEN-VLKT model: Section IV-B states that the authors 'use GEN-VLKT as the baseline to remove Position-Guided Embedding, and change the human and object query initialization to the same value.' This modified baseline scores 32.68 Full, whereas the published GEN-VLKT in Table I is 33.75. All module gains, including the 1.54 mAP attributed to IQE, are therefore measured against a weaker baseline. To attribute improvements to the proposed modules, ablations should be run on the unmodified GEN-VLKT baseline (or the baseline changes should be justified and reported as a separate contribution).
  3. [Section IV-C, Table IV] The drop-one-out results quoted in the text do not match Table IV. The text says removing IQE causes drops of 1.48, 3.54, and 0.86 mAP for Full, Rare, and Non-Rare. From Table IV, removing IQE from the complete model (row 8 vs row 6) gives drops of 35.23−33.50=1.73 Full, 32.79−29.47=3.32 Rare, and 35.96−34.71=1.25 Non-Rare. The text's numbers appear to use a different reference point, making the central ablation claim internally inconsistent. Please correct the text or the table and state exactly which configuration is the drop-one-out reference.
  4. [Section III-D, Implementation Details, Fig. 5] The weighting parameter alpha in Eq. (12) is never given a value, although the APU contribution is a claimed component and is ablated. Without alpha, the auxiliary prediction branch is not reproducible. Additionally, the hyperparameters K (Fig. 5) and R are selected using test-set mAP; this is a form of test-set tuning and should be acknowledged. If a validation split is used, please describe it; otherwise, the reported gains may be optimistically biased.
  5. [Section III-C, Eq. (5), Eq. (15)] The same CLIP similarity signal msim is used both to construct the HOI candidate set for interaction-query initialization (Eq. 5) and as a training-free inference term (Eq. 15). This creates a risk of double-counting: the IQE module may partly learn to amplify information that is later added directly to the score. To validate that IQE provides an independent benefit, report an ablation that removes the training-free term (or uses a different candidate-selection signal) and show the contribution of IQE separately from the shared CLIP term.
minor comments (5)
  1. [Section III-A] Typo: 'humam' should be 'human' in the HOI triplet description.
  2. [Eq. (16)] Equation (16) writes S_o ⊙ S_o, which appears to be a typo (likely S_o, or a different combination). Please clarify the intended formula.
  3. [Fig. 3] The caption is incomplete: 'with the HOI accuracy of Top-K matched texts' is not a complete sentence and the figure axis labels are not explained.
  4. [Section IV-D, Table VIII] The 60-epoch results are interesting but are compared only against the 90-epoch published numbers. Please report the 90-epoch DQEN result in the same table or in the text to enable a fair convergence comparison.
  5. [Section IV-C, Table VI] Table VI reports that OQE decreases object-detection mAP from 34.86 to 32.72. This is an important finding, but the text does not discuss why the selected features, despite correct object categories, produce attention maps focused on humans. A short explanation would help.

Circularity Check

0 steps flagged

No significant circularity: the proposed modules are ablated independently, and the CLIP-similarity reuse is an input feature, not a fitted prediction.

full rationale

DQEN is an empirical architecture paper; there is no derivation chain from first principles whose output is equivalent to its input. The central claims (OQE, IQE, APU) are supported by ablations in Table IV: each module is toggled against a stated modified GEN-VLKT baseline, and the full model's HICO-Det gain is reported both with and without the training-free term (Table I), so the IQE improvement is not merely the CLIP similarity term re-entering at inference. Eq. (5) and Eq. (15) both use the same CLIP image-text similarity msim, but Eq. (15) is an explicitly additive, training-free scoring term, and the model's trained interaction score Shoi is not defined as msim; thus no fitted parameter is renamed as a prediction. The self-citations ([7], [41]) appear only in context/loss discussions and are not load-bearing for any claimed result. Hyperparameters K and R are selected from test-set performance curves (Fig. 5 and Section IV-C), which is benchmark tuning rather than equation-level circularity. The paper's own Section V admits reduced effectiveness on V-COCO due to multiple interactions per image; this is a claim-support weakness, not a circularity. No quoted reduction equates a derived result to an input by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method's performance depends on K, N, alpha, and R; K and R are tuned directly on the test set. No new physical or categorical entities are introduced.

free parameters (4)
  • K (number of HOI triplet candidates) = 16 for HICO-Det, 20 for V-COCO
    Selected by sweeping the final mAP on the test set (Fig. 5); no validation split is reported.
  • N (number of Top-N encoder features for object query enhancement) = 64 (equal to query count Nq)
    Set equal to Nq and not ablated separately; affects OQE behavior.
  • alpha (weighting of auxiliary verb prediction)
    Introduced in Eq. 12 but its value is never reported in the paper.
  • R (training-free top-R text similarity cutoff) = 10
    Used in Eq. 15; set without a validation-based justification.
axioms (4)
  • standard math Standard DETR set-prediction with Hungarian matching and cross-entropy losses is a valid training objective for HOI detection.
    Used as the underlying framework throughout the method section.
  • domain assumption CLIP image-text similarity between the full image and HOI text labels is a usable prior for the true HOI categories.
    Motivates IQE and is partially validated by Fig. 3, but only at a coarse top-8 level.
  • ad hoc to paper The top-N object-classifier scores on encoder features select objects that are involved in interactions with humans.
    This is the premise of OQE; the paper itself finds it degrades pure object detection yet helps HOI detection.
  • ad hoc to paper Gradient truncation through the object query enhancement pathway prevents the model from overfocusing on objects.
    Empirically justified in Section IV-C via Fig. 6, but only for the specific architecture choices made.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DQEN: Dual Query Enhancement Network for DETR-based HOI Detection." pith.science (2026). https://pith.science/paper/LOWDBJVM

@misc{pith2026250818896,
  author       = {Pith},
  title        = {Pith review of: DQEN: Dual Query Enhancement Network for DETR-based HOI Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LOWDBJVM}},
  note         = {Machine review of arXiv:2508.18896}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Human-Object Interaction (HOI) detection focuses on localizing human-object pairs and recognizing their interactions. Recently, the DETR-based framework has been widely adopted in HOI detection. In DETR-based HOI models, queries with clear meaning are crucial for accurately detecting HOIs. However, prior works have typically relied on randomly initialized queries, leading to vague representations that limit the model's effectiveness. Meanwhile, humans in the HOI categories are fixed, while objects and their interactions are variable. Therefore, we propose a Dual Query Enhancement Network (DQEN) to enhance object and interaction queries. Specifically, object queries are enhanced with object-aware encoder features, enabling the model to focus more effectively on humans interacting with objects in an object-aware way. On the other hand, we design a novel Interaction Semantic Fusion module to exploit the HOI candidates that are promoted by the CLIP model. Semantic features are extracted to enhance the initialization of interaction queries, thereby improving the model's ability to understand interactions. Furthermore, we introduce an Auxiliary Prediction Unit aimed at improving the representation of interaction features. Our proposed method achieves competitive performance on both the HICO-Det and the V-COCO datasets. The source code is available at https://github.com/lzzhhh1019/DQEN.

Figures

Figures reproduced from arXiv: 2508.18896 by Chong Wang, Jiafei Wu, Jiangbo Qian, Jiong Wang, Yi Chen, Yinghao Lu, Zhehao Li.

Figure 1
Figure 1. Figure 1: The pipeline of proposed Dual Query Enhancement Network. Previous [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of our DQEN. Compared with the previous works, our method contains two main components: Object Query Enhancement [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of CLIP’s image-text matching capability with the HOI [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The pipeline of the Interaction Semantic Fusion. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of attention maps of encoder with the corresponding [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ours and GEN-VLKT’s convergence curves. TSA denotes the training [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of predictions. The orange rectangles containing scores indicate HOI scores, whereas those with words represent the predicted results. [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of failed prediction in GEN-VLKT [12] and successfully predict in Ours. The red rectangle with words denotes incorrect predictions, [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] 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

60 extracted references · 50 canonical work pages

  1. [1]

    Image captioning: Transforming objects into words,

    S. Herdade, A. Kappeler, K. Boakye, and J. Soares, “Image captioning: Transforming objects into words,” Advances in neural information processing systems, vol. 32, 2019

    work page 2019

  2. [2]

    Cascade seman- tic prompt alignment network for image captioning,

    J. Li, L. Zhang, K. Zhang, B. Hu, H. Xie, and Z. Mao, “Cascade seman- tic prompt alignment network for image captioning,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 7, pp. 5266– 5281, 2023

    work page 2023

  3. [3]

    Multi-stream interaction networks for human action recognition,

    H. Wang, B. Yu, J. Li, L. Zhang, and D. Chen, “Multi-stream interaction networks for human action recognition,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 5, pp. 3050–3060, 2021

    work page 2021

  4. [4]

    Understanding atomic hand-object interaction with human intention,

    H. Fan, T. Zhuo, X. Yu, Y . Yang, and M. Kankanhalli, “Understanding atomic hand-object interaction with human intention,” IEEE Transac- tions on Circuits and Systems for Video Technology , vol. 32, no. 1, pp. 275–285, 2021. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11

    work page 2021

  5. [5]

    Transductive learning with prior knowledge for generalized zero-shot action recognition,

    T. Su, H. Wang, Q. Qi, L. Wang, and B. He, “Transductive learning with prior knowledge for generalized zero-shot action recognition,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 34, no. 1, pp. 260–273, 2023

    work page 2023

  6. [6]

    Exploring spatio–temporal graph convolution for video-based human–object interaction recognition,

    N. Wang, G. Zhu, H. Li, M. Feng, X. Zhao, L. Ni, P. Shen, L. Mei, and L. Zhang, “Exploring spatio–temporal graph convolution for video-based human–object interaction recognition,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 10, pp. 5814–5827, 2023

    work page 2023

  7. [7]

    Feature reconstruc- tion with disruption for unsupervised video anomaly detection,

    C. Tao, C. Wang, S. Lin, S. Cai, D. Li, and J. Qian, “Feature reconstruc- tion with disruption for unsupervised video anomaly detection,” IEEE Transactions on Multimedia , 2024

    work page 2024

  8. [8]

    Evcap: Element-aware video captioning,

    S. Liu, A. Li, Y . Zhao, J. Wang, and Y . Wang, “Evcap: Element-aware video captioning,” IEEE Transactions on Circuits and Systems for Video Technology, 2024

    work page 2024

  9. [9]

    Coun- terfactual samples synthesizing for robust visual question answering,

    L. Chen, X. Yan, J. Xiao, H. Zhang, S. Pu, and Y . Zhuang, “Coun- terfactual samples synthesizing for robust visual question answering,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 10 800–10 809

    work page 2020

  10. [10]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in European conference on computer vision . Springer, 2020, pp. 213– 229

    2020

  11. [11]

    Mining the benefits of two-stage and one-stage hoi detection,

    A. Zhang, Y . Liao, S. Liu, M. Lu, Y . Wang, C. Gao, and X. Li, “Mining the benefits of two-stage and one-stage hoi detection,” Advances in Neural Information Processing Systems , vol. 34, pp. 17 209–17 220, 2021

    work page 2021

  12. [12]

    Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection,

    Y . Liao, A. Zhang, M. Lu, Y . Wang, X. Li, and S. Liu, “Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 20 123–20 132

    work page 2022

  13. [13]

    Hoiclip: Efficient knowledge trans- fer for hoi detection with vision-language models,

    S. Ning, L. Qiu, Y . Liu, and X. He, “Hoiclip: Efficient knowledge trans- fer for hoi detection with vision-language models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 23 507–23 517

    work page 2023

  14. [14]

    Qpic: Query-based pairwise human-object interaction detection with image-wide contextual informa- tion,

    M. Tamura, H. Ohashi, and T. Yoshinaga, “Qpic: Query-based pairwise human-object interaction detection with image-wide contextual informa- tion,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 10 410–10 419

    work page 2021

  15. [15]

    Deformable detr: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to-end object detection,”arXiv preprint arXiv:2010.04159, 2020

    Pith/arXiv arXiv 2010

  16. [16]

    Efficient detr: improving end-to- end object detector with dense prior,

    Z. Yao, J. Ai, B. Li, and C. Zhang, “Efficient detr: improving end-to- end object detector with dense prior,” arXiv preprint arXiv:2104.01318, 2021

    Pith/arXiv arXiv 2021

  17. [17]

    Dino: Detr with improved denoising anchor boxes for end-to- end object detection,

    H. Zhang, F. Li, S. Liu, L. Zhang, H. Su, J. Zhu, L. M. Ni, and H.-Y . Shum, “Dino: Detr with improved denoising anchor boxes for end-to- end object detection,” arXiv preprint arXiv:2203.03605 , 2022

    Pith/arXiv arXiv 2022

  18. [18]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PMLR, 2021, pp. 8748–8763

    2021

  19. [19]

    Category-aware transformer network for better human-object inter- action detection,

    L. Dong, Z. Li, K. Xu, Z. Zhang, L. Yan, S. Zhong, and X. Zou, “Category-aware transformer network for better human-object inter- action detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 19 538–19 547

    work page 2022

  20. [20]

    Inter- activeness field in human-object interactions,

    X. Liu, Y .-L. Li, X. Wu, Y .-W. Tai, C. Lu, and C.-K. Tang, “Inter- activeness field in human-object interactions,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 20 113–20 122

    work page 2022

  21. [21]

    Mining cross- person cues for body-part interactiveness learning in hoi detection,

    X. Wu, Y .-L. Li, X. Liu, J. Zhang, Y . Wu, and C. Lu, “Mining cross- person cues for body-part interactiveness learning in hoi detection,” in European Conference on Computer Vision . Springer, 2022, pp. 121– 136

    work page 2022

  22. [22]

    Spatially conditioned graphs for detecting human-object interactions,

    F. Z. Zhang, D. Campbell, and S. Gould, “Spatially conditioned graphs for detecting human-object interactions,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 13 319–13 327

    work page 2021

  23. [23]

    Exploring structure-aware transformer over interaction proposals for human-object interaction detection,

    Y . Zhang, Y . Pan, T. Yao, R. Huang, T. Mei, and C.-W. Chen, “Exploring structure-aware transformer over interaction proposals for human-object interaction detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 19 548–19 557

    work page 2022

  24. [24]

    Cascaded human-object interaction recognition,

    T. Zhou, W. Wang, S. Qi, H. Ling, and J. Shen, “Cascaded human-object interaction recognition,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 4263–4272

    work page 2020

  25. [25]

    Hoi analysis: Integrating and decomposing human-object interaction,

    Y .-L. Li, X. Liu, X. Wu, Y . Li, and C. Lu, “Hoi analysis: Integrating and decomposing human-object interaction,” Advances in Neural Infor- mation Processing Systems , vol. 33, pp. 5011–5022, 2020

    work page 2020

  26. [26]

    Vsgnet: Spatial attention network for detecting human object interactions using graph convolu- tions,

    O. Ulutan, A. Iftekhar, and B. S. Manjunath, “Vsgnet: Spatial attention network for detecting human object interactions using graph convolu- tions,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 13 617–13 626

    work page 2020

  27. [27]

    Detecting human-object interactions with object-guided cross-modal calibrated semantics,

    H. Yuan, M. Wang, D. Ni, and L. Xu, “Detecting human-object interactions with object-guided cross-modal calibrated semantics,” in Proceedings of the AAAI Conference on artificial intelligence , vol. 36, no. 3, 2022, pp. 3206–3214

    work page 2022

  28. [28]

    Polysemy deciphering network for human-object interaction detection,

    X. Zhong, C. Ding, X. Qu, and D. Tao, “Polysemy deciphering network for human-object interaction detection,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XX 16 . Springer, 2020, pp. 69–85

    work page 2020

  29. [29]

    Hotr: End-to-end human-object interaction detection with transformers,

    B. Kim, J. Lee, J. Kang, E.-S. Kim, and H. J. Kim, “Hotr: End-to-end human-object interaction detection with transformers,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 74–83

    work page 2021

  30. [30]

    Relational context learning for human- object interaction detection,

    S. Kim, D. Jung, and M. Cho, “Relational context learning for human- object interaction detection,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2023, pp. 2925–2934

    work page 2023

  31. [31]

    Category query learning for human-object interaction classification,

    C. Xie, F. Zeng, Y . Hu, S. Liang, and Y . Wei, “Category query learning for human-object interaction classification,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 15 275–15 284

    work page 2023

  32. [32]

    Towards hard-positive query mining for detr-based human-object interaction detection,

    X. Zhong, C. Ding, Z. Li, and S. Huang, “Towards hard-positive query mining for detr-based human-object interaction detection,” in European Conference on Computer Vision . Springer, 2022, pp. 444–460

    work page 2022

  33. [33]

    Multi-scale human-object interaction detector,

    Y . Cheng, Z. Wang, W. Zhan, and H. Duan, “Multi-scale human-object interaction detector,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 4, pp. 1827–1838, 2022

    work page 2022

  34. [34]

    Rr-net: Relation reasoning for end-to-end human-object interaction detection,

    D. Yang, Y . Zou, C. Zhang, M. Cao, and J. Chen, “Rr-net: Relation reasoning for end-to-end human-object interaction detection,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 32, no. 6, pp. 3853–3865, 2021

    work page 2021

  35. [35]

    Hodn: Disentangling human-object feature for hoi detection,

    S. Fang, Z. Lin, K. Yan, J. Li, X. Lin, and R. Ji, “Hodn: Disentangling human-object feature for hoi detection,” IEEE Transactions on Multi- media, 2023

    work page 2023

  36. [36]

    Ted-net: Dispersal attention for perceiving interaction region in indirectly-contact hoi detection,

    Y . Wang, Q. Liu, and Y . Lei, “Ted-net: Dispersal attention for perceiving interaction region in indirectly-contact hoi detection,”IEEE Transactions on Circuits and Systems for Video Technology , 2024

    work page 2024

  37. [37]

    Exploiting scene graphs for human-object interaction detection,

    T. He, L. Gao, J. Song, and Y .-F. Li, “Exploiting scene graphs for human-object interaction detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 15 984–15 993

    work page 2021

  38. [38]

    Rlip: Relational language-image pre-training for human-object inter- action detection,

    H. Yuan, J. Jiang, S. Albanie, T. Feng, Z. Huang, D. Ni, and M. Tang, “Rlip: Relational language-image pre-training for human-object inter- action detection,” Advances in Neural Information Processing Systems , vol. 35, pp. 37 416–37 431, 2022

    work page 2022

  39. [39]

    Visual genome: Connecting language and vision using crowdsourced dense image annotations,

    R. Krishna, Y . Zhu, O. Groth, J. Johnson, K. Hata, J. Kravitz, S. Chen, Y . Kalantidis, L.-J. Li, D. A. Shammaet al., “Visual genome: Connecting language and vision using crowdsourced dense image annotations,” International journal of computer vision , vol. 123, pp. 32–73, 2017

    work page 2017

  40. [40]

    Distilling the knowledge in a neural network,

    G. Hinton, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531, 2015

    Pith/arXiv arXiv 2015

  41. [41]

    Restructuring the teacher and student in self-distillation,

    Y . Zheng, C. Wang, C. Tao, S. Lin, J. Qian, and J. Wu, “Restructuring the teacher and student in self-distillation,” IEEE Transactions on Image Processing, 2024

    work page 2024

  42. [42]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  43. [43]

    Learning to detect human-object interactions,

    Y .-W. Chao, Y . Liu, X. Liu, H. Zeng, and J. Deng, “Learning to detect human-object interactions,” in 2018 ieee winter conference on applications of computer vision (wacv) . IEEE, 2018, pp. 381–389

    work page 2018

  44. [44]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

  45. [45]

    Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer,

    F. Z. Zhang, D. Campbell, and S. Gould, “Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 20 104–20 112

    work page 2022

  46. [46]

    End-to-end human object interaction detection with hoi transformer,

    C. Zou, B. Wang, Y . Hu, J. Liu, Q. Wu, Y . Zhao, B. Li, C. Zhang, C. Zhang, Y . Weiet al., “End-to-end human object interaction detection with hoi transformer,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 11 825–11 834

    work page 2021

  47. [47]

    Open-category human-object interaction pre-training via language modeling framework,

    S. Zheng, B. Xu, and Q. Jin, “Open-category human-object interaction pre-training via language modeling framework,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 19 392–19 402

    work page 2023

  48. [48]

    Efficient adaptive human-object interaction detection with concept-guided memory,

    T. Lei, F. Caba, Q. Chen, H. Jin, Y . Peng, and Y . Liu, “Efficient adaptive human-object interaction detection with concept-guided memory,” in JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 12 Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 6480–6490

    work page 2021

  49. [49]

    Exploring predicate visual context in detecting of human-object interactions,

    F. Z. Zhang, Y . Yuan, D. Campbell, Z. Zhong, and S. Gould, “Exploring predicate visual context in detecting of human-object interactions,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 10 411–10 421

    work page 2023

  50. [50]

    Clip4hoi: towards adapting clip for practical zero-shot hoi detection,

    Y . Mao, J. Deng, W. Zhou, L. Li, Y . Fang, and H. Li, “Clip4hoi: towards adapting clip for practical zero-shot hoi detection,” Advances in Neural Information Processing Systems , vol. 36, pp. 45 895–45 906, 2023

    work page 2023

  51. [51]

    Exploring conditional multi- modal prompts for zero-shot hoi detection,

    T. Lei, S. Yin, Y . Peng, and Y . Liu, “Exploring conditional multi- modal prompts for zero-shot hoi detection,” in European Conference on Computer Vision . Springer, 2024, pp. 1–19

    work page 2024

  52. [52]

    Unseen no more: Unlocking the potential of clip for generative zero-shot hoi detection,

    Y . Guo, Y . Liu, J. Li, W. Wang, and Q. Jia, “Unseen no more: Unlocking the potential of clip for generative zero-shot hoi detection,” in Proceedings of the 32nd ACM International Conference on Multimedia , 2024, pp. 1711–1720

    work page 2024

  53. [53]

    Ernet: An efficient and reliable human-object interaction detection network,

    J. Lim, V . M. Baskaran, J. M.-Y . Lim, K. Wong, J. See, and M. Tistarelli, “Ernet: An efficient and reliable human-object interaction detection network,” IEEE Transactions on Image Processing , vol. 32, pp. 964– 979, 2023

    work page 2023

  54. [54]

    Learning from easy to hard pairs: Multi-step reasoning network for human-object interaction detection,

    Y . Zhou, G. Tan, M. Li, and C. Gou, “Learning from easy to hard pairs: Multi-step reasoning network for human-object interaction detection,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 4368–4377

    work page 2023

  55. [55]

    A plug-and-play method for rare human-object interactions detection by bridging domain gap,

    L. Zhang, W. Suo, P. Wang, and Y . Zhang, “A plug-and-play method for rare human-object interactions detection by bridging domain gap,” in Proceedings of the 32nd ACM International Conference on Multimedia , 2024, pp. 8613–8622

    work page 2024

  56. [56]

    Visual semantic role labeling,

    S. Gupta and J. Malik, “Visual semantic role labeling,” arXiv preprint arXiv:1505.04474, 2015

    Pith/arXiv arXiv 2015

  57. [57]

    Visual compositional learning for human-object interaction detection,

    Z. Hou, X. Peng, Y . Qiao, and D. Tao, “Visual compositional learning for human-object interaction detection,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16 . Springer, 2020, pp. 584–600

    work page 2020

  58. [58]

    Affordance transfer learning for human-object interaction detection,

    Z. Hou, B. Yu, Y . Qiao, X. Peng, and D. Tao, “Affordance transfer learning for human-object interaction detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 495–504

    work page 2021

  59. [59]

    Detecting human-object interaction via fabricated compositional learning,

    Z. Hou, B. Yu, Y . Qiao, and D. Tao, “Detecting human-object interaction via fabricated compositional learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 14 646–14 655

    work page 2021

  60. [60]

    Decoupled weight decay regularization,

    I. Loshchilov, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017. VII. B IOGRAPHY SECTION Zhehao Li received his B.S. degree from Wenzhou University in 2023. He is currently pursuing a mas- ter’s degree in the Faculty of Electrical Engineering and Computer Science, Ningbo University, Ningbo, China. His main research interests ...

    Pith/arXiv arXiv 2017

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.