Pith. sign in

REVIEW 4 major objections 7 minor 33 references

Prompt Guidance and Human Proximal Perception for HOT Prediction with Regional Joint Loss

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read P3HOT combines text-prompt guidance, depth-based proximal perception, and a regional joint loss to achieve state-of-the-art human-object contact prediction on four metrics across two benchmarks.

desk verdict Useful application of text and depth cues to HOT, but the RJLoss as written cannot backpropagate, so the SOTA claim rests on an unexplained gain. read the letter →

arxiv 2507.01630 v2 pith:4XUVCW7I submitted 2025-07-02 cs.CV cs.AI

Yuxiao Wang , Yu Lei , Zhenao Wei , Weiying Xue , Xinyu Jiang , Nan Zhuang , Qi Liu This is my paper · ORCID
classification cs.CVcs.AI
keywords human-objectcontactdetectiontextpromptguidancedepth-basedproximityregionaljointlossadaptiveaccuracymulti-modalsegmentationconnected-componentregularization
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 P3HOT, a framework for Human-Object conTact (HOT) prediction that fuses text-prompt guidance, a depth-based human proximal perception module, and a new Regional Joint Loss. Its central claim is that combining a CLIP-based text prompt with a learnable depth-range filter around each person removes improbable contact regions and yields state-of-the-art scores on four metrics across the HOT-Annotated and HOT-Generated benchmarks. It also proposes a new evaluation metric, AD-Acc., that penalizes maps that over-predict the human body as a single contact class. If the results hold, the method establishes a strong multi-modal baseline for HOT and a more reliable way to measure it.

What carries the argument

The argument runs on three linked mechanisms. The first is an image-text similarity score $S$ (Eq. 1) computed between a ResNet-50 attention-pooled image feature and a frozen text encoder's embeddings of the template 'A [body part] of the human body is in contact with an object.' The second is the Human Proximal Perception (HPP) module, which normalizes a depth map, averages depth inside each SAM-generated human mask, then builds a differentiable filter $FM$ from a learnable margin $\tau$ (Eqs. 8-10) that gates which decoder features survive. The third is Regional Joint Loss, composed of a Local Joint Loss $L_L$ that counts wrong-class pixels inside ground-truth regions and a Global Joint Loss $L_G$ that uses connected components of the inverted prediction to find and penalize foreign classes enclosed within predicted regions; the proposed AD-Acc. metric reweights accuracy by subtracting errors on negative human-body samples.

What would settle it

Run the Table 3 ablation on HOT-Annotated with the RJLoss term detached from the graph (stop-gradient). If SC-Acc. does not drop from 46.0 to roughly 44.5, then the reported +1.5 gain is not caused by RJLoss as a learning signal; alternatively, a direct autograd check that the gradient of $L_L + L_G$ with respect to the network output is nonzero would confirm trainability.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that semantic text prompts and pseudo-3D depth proximity cues jointly solve the over-segmentation and category-inconsistency failures of single-image HOT models. By using human masks and a learnable parameter $\tau$ to keep only depth ranges near the body, the Human Proximal Perception module suppresses background and far-object interference; the text prompt forces the decoder to attend to specific body parts; and the Regional Joint Loss penalizes the appearance of foreign classes both inside ground-truth regions (local) and inside predicted connected regions (global). The result is a model that improves SC-Acc., mIoU, wIoU, and AD-Acc. by 0.7, 2.0, 1.6, and 11.0 points respectively on HOT-Annotated, and scores 35.2/18.0/23.1/30.6 on HOT-Generated, beating both DHOT and PIHOT.

Load-bearing premise

The jump in performance attributed to Regional Joint Loss in Table 3 rests on that loss being differentiable end-to-end, but the paper does not show how gradients flow through the hard binary maps and connected-component routine used in Eqs. 14-19.

Editorial extensions

If this is right

  • If P3HOT's gains are real, multi-modal image-plus-text cues become a standard ingredient for HOT, since the text prompt alone contributes roughly 1.4 SC-Acc. points over the fine-decoder baseline (40.3 vs 38.9 in Table 2).
  • Depth-based proximal filtering gives a cheap quasi-3D inductive bias that reduces background over-segmentation, which is the main failure mode of the prior DHOT model.
  • AD-Acc. could replace C-Acc. in future HOT benchmarks, since it penalizes a trivial all-contact prediction that C-Acc. rewards at 100%.
  • The Regional Joint Loss offers a general recipe for enforcing intra-region category consistency in any segmentation task, independent of the HOT setting.
  • P3HOT is lighter and faster than the depth-restoration-based PIHOT, suggesting that prompt plus depth-range cues achieve comparable or better precision without heavy object inpainting.

Reading between the lines

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

  • Because the paper's RJLoss equations use hard 0/1 masks and a non-differentiable connected-component routine, a natural next step is to reformulate $L_L$ and $L_G$ with soft region maps or a differentiable connected-component surrogate; if such a surrogate recovers or exceeds the reported +1.5 SC-Acc. gain, it would resolve the open question of how much the loss itself trains the network.
  • The learnable margin $\tau$ could be inspected per image: if it consistently shrinks on cluttered scenes and grows on isolated people, the HPP module is implicitly learning scene-depth priors, a behavior the paper does not analyze.
  • The AD-Acc. metric's negative-sample term depends on the human mask; extending it to include object-side negatives or confidence-weighted masks would make it applicable beyond HOT to general contact or affordance prediction.
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 / 7 minor

Summary. The paper proposes P3HOT, a Human-Object conTact (HOT) prediction framework that combines a CLIP-based text-prompt attention mechanism, a depth-based human proximal perception module with a learnable range parameter, multi-level decoder feature fusion, and a newly proposed Regional Joint Loss (RJLoss). It also introduces a new evaluation metric, AD-Acc., which the authors argue better handles negative samples than C-Acc. The central claim, stated in the abstract and Section 4.4, is that the method achieves state-of-the-art results on SC-Acc., mIoU, wIoU, and AD-Acc. on both HOT-Annotated and HOT-Generated, with the largest reported gain on the self-proposed AD-Acc. metric.

Significance. If the results hold, the paper makes several useful contributions: it is the first to integrate textual prompts into HOT, it provides a pseudo-3D depth-range mechanism with a learnable threshold, and it releases code. The ablations are internally organized, and the authors explicitly address the differentiability of the HPP mask in Eqs. (8)-(10), which is a positive sign. However, the significance is currently capped by two issues: the RJLoss as written in Section 3.5 is not differentiable under standard backpropagation, and the empirical SOTA claim relies on single-run scores without error bars, including a new self-defined metric. The paper's core scientific claim is therefore unverified until the RJLoss gradient path is specified or replaced with a differentiable surrogate and the experiments are repeated with variance estimates.

major comments (4)
  1. [Section 3.5, Eqs. (14)-(20)] The RJLoss is not a differentiable function of the network parameters as written. Eq. (14) constructs Oc by hard-thresholding the predicted class, and the condition "opq(gtpq)=c" is also self-referential because opq appears on both sides; the resulting map is piecewise-constant with zero derivative almost everywhere. Eq. (17) applies scipy's ConnectedArea to a hard-negated map, an operation with no defined gradient, and Eq. (19) sums hard binary comparisons. No soft surrogate, straight-through estimator, or detach trick is described. Consequently, RJLoss cannot update the network under standard backpropagation, yet Table 3 attributes a +1.5 SC-Acc. gain (44.5 to 46.0) to adding RJLoss. This is load-bearing: without that gain, the method's SC-Acc. (44.5) would not beat PIHOT (45.3) on HOT-Annotated. The authors explicitly solved the analogous differentiability problem for the HPP mask in Eqs. (8)-(10) with a ReLU relaxation; a similar fix, or a precise gradient specification, is needed for RJLoss.
  2. [Section 4.5, Table 5] The text states that the optimal loss weights are α=0.3 and β=0.5, but the best reported row in Table 5 uses α=0.3, β=0.1. The text and table are inconsistent, so the loss-weight claim needs correction and the correct configuration should be stated unambiguously.
  3. [Tables 1 and 3] All results are reported as single runs without error bars or significance tests. The claimed SOTA margins over PIHOT on SC-Acc. are 0.7 and 0.3 on HOT-Annotated and HOT-Generated, respectively, which are small relative to typical run-to-run variation in segmentation training. To support the SOTA claim, the authors should report mean ± std over at least three seeds or provide an equivalent statistical comparison.
  4. [Abstract and Section 4.3] AD-Acc. is introduced and defined in this paper and is then used as one of the four metrics in the SOTA claim. Because it is a self-proposed metric, it should be validated (for example, by reporting its relationship to the existing metrics on all baselines) and computed for all methods in Table 1. As presented, the 'four metrics' claim is not directly comparable with previously published results, which did not use AD-Acc.
minor comments (7)
  1. [Eq. (14)] The notation in Eq. (14) is self-referential: opq(gtpq) is defined in terms of opq(gtpq)=c. It should be rewritten using a separate symbol for the predicted hard label at position (p,q).
  2. [Table 5] The column headers read 'IoU' and 'mIoU', which is inconsistent with the metrics 'mIoU' and 'wIoU' used in Tables 1 and 4; the headers should be corrected to match the metric definitions in Section 4.3.
  3. [Eq. (24)] The AD-Acc. formula is missing parentheses and is ambiguous as typeset; it should be written as P(GT_B ⊗ O_B)/(P(GT_B)+δ) - P(ζ ⊗ O_B)/(P(ζ)+δ) or with fully explicit grouping.
  4. [Section 3.2] The similarity vector S is defined as S∈R^{1×TN}, but Eq. (13) indexes S[i] for i=1,...,17; the relationship between TN and the 18 output channels (including background) should be clarified.
  5. [Section 4.4] The text refers to 'PIHOI' where it should refer to 'PIHOT'; the typo should be corrected.
  6. [References] The reference 'OpenCat' is listed as '[?]' in Section 2 and should be resolved before publication.
  7. [Abstract and Section 5] There are minor grammatical issues, such as 'The sources code are available' in the abstract and the misspelled heading 'Conlusion' in Section 5; these should be cleaned up.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central comparisons are empirical against externally defined benchmarks, and the newly proposed metric is applied uniformly rather than derived from the model's own outputs.

full rationale

The paper's derivation chain is self-contained: the claimed state-of-the-art results are empirical comparisons on the externally introduced HOT-Annotated and HOT-Generated benchmarks, and the proposed modules (text-prompt attention, HPP depth filtering, decoder feature fusion, and the losses) are specified as concrete network operations with ground-truth objectives. The new AD-Acc. metric is defined as a fixed function of predictions, ground truth, and the human mask, and is then applied uniformly to all baselines; introducing a metric and reporting a gain on it is an evaluation-design choice, not a circular derivation. Several references are to the authors' own prior work, notably PIHOT [26], TED-Net [22], FreeA [24], and the HOI review [25], but these are used as related work and as baselines rather than as the justification for the central claim, so they are not load-bearing self-citations. The RJLoss as written in Eqs. 14, 17, and 19 uses hard thresholding and scipy ConnectedArea, which raises a genuine differentiability and reproducibility concern about the reported +1.5 SC-Acc. gain, but that is a correctness/implementation issue rather than circular reasoning. No step reduces by construction to its own input, and no prediction is merely a renamed fit, so the circularity score is 0.

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

The model relies on off-the-shelf CLIP, SAM, and ZoeDepth as fixed priors, a single global learnable depth half-width tau, and hand-tuned loss weights. The RJLoss as written is not differentiable and is the main unverified assumption. No new physical entities are introduced.

free parameters (4)
  • tau (learnable depth range half-width) = learned, value not reported
    Sec. 3.3, Eq. 5-10: a global scalar defining the depth window around each person's average depth, learned end-to-end. The paper does not report its converged value or sensitivity.
  • alpha (Local Joint Loss weight) = 0.3
    Table 5 and Eq. 21: weight selected by comparing final benchmark performance; no validation protocol is described.
  • beta (Global Joint Loss weight) = 0.1
    Table 5 and Eq. 21: weight selected by tuning; the text says 0.5 in one place, but the best row in Table 5 uses 0.1.
  • gamma (image-text BCE weight) = 1.0
    Eq. 21 and Sec. 4.2: weight for the binary cross-entropy image-text matching loss, set by hand.
assumptions (4)
  • domain assumption CLIP image-text similarity scores are a valid per-channel attention signal for pixel-level contact classes.
    Sec. 3.2-3.4: Eq. 1 computes global image-text similarity and Eq. 13 multiplies segmentation channels by S; this assumes global image-text alignment improves local contact segmentation.
  • domain assumption Off-the-shelf SAM human masks and ZoeDepth depth maps are accurate enough that the depth-window filter improves contact segmentation.
    Sec. 3.3: the HPP module relies on M and D from external models; no error analysis or robustness study is provided.
  • ad hoc to paper The Regional Joint Loss can be optimized by gradient descent as written.
    Sec. 3.5: Eq. 14 uses hard predicted labels, Eq. 17 uses scipy ConnectedArea, and Eq. 19 sums binary maps; none of these operations provides a gradient path, and the paper omits a soft surrogate or stop-gradient trick.
  • domain assumption A single global learnable tau represents the interaction depth range for all images and people.
    Sec. 3.3, Eq. 5-10: one scalar tau is applied to every person and every image, with no per-image or per-person adaptation beyond the computed average depth.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Prompt Guidance and Human Proximal Perception for HOT Prediction with Regional Joint Loss." pith.science (2026). https://pith.science/paper/4XUVCW7I

@misc{pith2026250701630,
  author       = {Pith},
  title        = {Pith review of: Prompt Guidance and Human Proximal Perception for HOT Prediction with Regional Joint Loss},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4XUVCW7I}},
  note         = {Machine review of arXiv:2507.01630}
}
abstract

The task of Human-Object conTact (HOT) detection involves identifying the specific areas of the human body that are touching objects. Nevertheless, current models are restricted to just one type of image, often leading to too much segmentation in areas with little interaction, and struggling to maintain category consistency within specific regions. To tackle this issue, a HOT framework, termed \textbf{P3HOT}, is proposed, which blends \textbf{P}rompt guidance and human \textbf{P}roximal \textbf{P}erception. To begin with, we utilize a semantic-driven prompt mechanism to direct the network's attention towards the relevant regions based on the correlation between image and text. Then a human proximal perception mechanism is employed to dynamically perceive key depth range around the human, using learnable parameters to effectively eliminate regions where interactions are not expected. Calculating depth resolves the uncertainty of the overlap between humans and objects in a 2D perspective, providing a quasi-3D viewpoint. Moreover, a Regional Joint Loss (RJLoss) has been created as a new loss to inhibit abnormal categories in the same area. A new evaluation metric called ``AD-Acc.'' is introduced to address the shortcomings of existing methods in addressing negative samples. Comprehensive experimental results demonstrate that our approach achieves state-of-the-art performance in four metrics across two benchmark datasets. Specifically, our model achieves an improvement of \textbf{0.7}$\uparrow$, \textbf{2.0}$\uparrow$, \textbf{1.6}$\uparrow$, and \textbf{11.0}$\uparrow$ in SC-Acc., mIoU, wIoU, and AD-Acc. metrics, respectively, on the HOT-Annotated dataset. The sources code are available at https://github.com/YuxiaoWang-AI/P3HOT.

Figures

Figures reproduced from arXiv: 2507.01630 by the authors.

Figure 1
Figure 1. Current method vs Ours. HOI focuses on detecting humans, objects, and their inter￾actions. It does not specify which parts of the human body come into contact with objects. However, identifying spe￾cific contact points on the human is required in the HOT task. This technology can be applied in various fields such as human-computer interaction [22], virtual reality [12], and gesture recognition [6]. HOT divides into … view at source ↗
Figure 2
Figure 2. Overall architecture. The image encoder and text encoder are used to extract features from images and text, respectively. In [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The calculation schematic of regions of other categories enclosed within category [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of DHOT [5], PIHOT [26], and our proposed method. Model HOT-Annotated HOT-Generated SC-Acc. C-Acc. mIoU wIoU AD-Acc. SC-Acc. C-Acc. mIoU wIoU AD-Acc. ResNet+UperNet [28] 35.1 62.6 19.5 22.7 - 21.1 42.7 8.0 11.6 - ResNet+PPM [31] 34.6 61.1 20.1 23.3 - 21.2…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 26 canonical work pages

  1. [1]

    ZoeDepth: Zero-shot trans- fer by combining relative and metric depth

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M ¨uller. ZoeDepth: Zero-shot trans- fer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 4

    arXiv 2023

  2. [2]

    Behave: Dataset and method for tracking human object in- teractions

    Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object in- teractions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935– 15946, 2022. 2, 3

    work page 2022

  3. [3]

    End- to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End- to-end object detection with transformers. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 213–229. Springer, 2020. 2

    work page 2020

  4. [4]

    Learning to detect human-object interactions

    Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 381–389, 2018. 2

    work page 2018

  5. [5]

    Detecting human-object contact in images

    Yixin Chen, Sai Kumar Dwivedi, Michael J Black, and Dim- itrios Tzionas. Detecting human-object contact in images. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 17100–17110, 2023. 1, 2, 3, 5, 6, 7

    work page 2023

  6. [6]

    Hand gesture segmentation against complex back- ground based on improved atrous spatial pyramid pooling

    Zhenchao Cui, Yu Lei, Yuxiao Wang, Wenzhu Yang, and Jing Qi. Hand gesture segmentation against complex back- ground based on improved atrous spatial pyramid pooling. Journal of Ambient Intelligence and Humanized Computing, 14(9):11795–11807, 2023. 1, 2

    work page 2023

  7. [7]

    Hierarchical reasoning network for human- object interaction detection

    Yiming Gao, Zhanghui Kuang, Guanbin Li, Wayne Zhang, and Liang Lin. Hierarchical reasoning network for human- object interaction detection. IEEE Transactions on Image Processing, 30:8306–8317, 2021. 2

    work page 2021

  8. [8]

    Detecting and recognizing human-object interactions

    Georgia Gkioxari, Ross Girshick, Piotr Doll ´ar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8359–8367, 2018. 2

    work page 2018

Show all 33 references
  1. [9]

    Visual semantic role la- beling

    Saurabh Gupta and Jitendra Malik. Visual semantic role la- beling. arXiv preprint arXiv:1505.04474, 2015. 6

    2015 arXiv

  2. [10]

    Resolving 3d human pose ambiguities with 3d scene constraints

    Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision , pages 2282– 2292, 2019. 6

    2019

  3. [11]

    Mask R-CNN

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask R-CNN. In Proceedings of the IEEE inter- national conference on computer vision , pages 2961–2969,

  4. [12]

    Learning to pre- dict head pose in remotely-rendered virtual reality

    Gazi Karam Illahi, Ashutosh Vaishnav, Teemu K ¨am¨ar¨ainen, Matti Siekkinen, and Mario Di Francesco. Learning to pre- dict head pose in remotely-rendered virtual reality. In Pro- ceedings of the 14th Conference on ACM Multimedia Sys- tems, pages 27–38, 2023. 1

    2023

  5. [13]

    Estimation of pos- ture and joint angle of human body using foot pressure distri- bution: Morphological computation with human foot

    Yo Kobayashi and Yasutaka Nakashima. Estimation of pos- ture and joint angle of human body using foot pressure distri- bution: Morphological computation with human foot. arXiv preprint arXiv:2401.12464, 2024. 2

    2024 arXiv

  6. [14]

    PastaNet: Toward human activity knowledge en- gine

    Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. PastaNet: Toward human activity knowledge en- gine. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 382–391, 2020. 6

    2020

  7. [15]

    GEN-VLKT: Simplify association and en- hance interaction understanding for hoi detection

    Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. GEN-VLKT: Simplify association and en- hance interaction understanding for hoi detection. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123–20132, 2022. 2

    2022

  8. [16]

    A dataset of relighted 3d interacting hands.Advances in Neural Information Processing Systems, 36, 2024

    Gyeongsik Moon, Shunsuke Saito, Weipeng Xu, Rohan Joshi, Julia Buffalini, Harley Bellan, Nicholas Rosen, Jesse Richardson, Mallorie Mize, Philippe De Bree, et al. A dataset of relighted 3d interacting hands.Advances in Neural Information Processing Systems, 36, 2024. 2

    2024

  9. [17]

    Detecting hands and recognizing physical contact in the wild

    Supreeth Narasimhaswamy, Trung Nguyen, and Minh Hoai Nguyen. Detecting hands and recognizing physical contact in the wild. Advances in neural information processing sys- tems, 33:7841–7851, 2020. 2

    2020

  10. [18]

    Language segment-anything

    paulguerrero. Language segment-anything. https://github.com/paulguerrero/lang-sam, 2024. 4

    2024

  11. [19]

    Expressive body capture: 3d hands, face, and body from a single image

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognitio...

    2019

  12. [20]

    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...

    2021

  13. [21]

    Deco: Dense estimation of 3d human-scene contact in the wild

    Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J Black. Deco: Dense estimation of 3d human-scene contact in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8001–8013, 2023. 2

    2023

  14. [22]

    TED-Net: Dispersal at- tention for perceiving interaction region in indirectly-contact hoi detection

    Yuxiao Wang, Qi Liu, and Yu Lei. TED-Net: Dispersal at- tention for perceiving interaction region in indirectly-contact hoi detection. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2

    2024

  15. [23]

    CycleHOI: Improving human-object interaction detection with cycle consistency of detection and generation

    Yisen Wang, Yao Teng, and Limin Wang. CycleHOI: Improving human-object interaction detection with cycle consistency of detection and generation. arXiv preprint arXiv:2407.11433, 2024. 2

    2024 arXiv

  16. [24]

    Freea: Human-object interac- tion detection using free annotation labels

    Yuxiao Wang, Zhenao Wei, Xinyu Jiang, Yu Lei, Weiying Xue, Jinxiu Liu, and Qi Liu. Freea: Human-object interac- tion detection using free annotation labels. arXiv preprint arXiv:2403.01840, 2024. 1, 2

    2024 arXiv

  17. [25]

    A review of human-object interaction de- tection

    Yuxiao Wang, Qiwei Xiong, Yu Lei, Weiying Xue, Qi Liu, and Zhenao Wei. A review of human-object interaction de- tection. arXiv preprint arXiv:2408.10641, 2024. 1, 2

    2024 arXiv

  18. [26]

    Precision-enhanced human-object contact detection via depth-aware perspective interaction and object texture restoration

    Yuxiao Wang, Wenpeng Neng, Zhenao Wei, Yu Lei, Weiy- ing Xue, Nan Zhuang, Yanwu Xu, Xinyu Jiang, and Qi Liu. Precision-enhanced human-object contact detection via depth-aware perspective interaction and object texture restoration. In Proceedings of the AAAI Conference on Arti-...

    2025

  19. [27]

    Watch-n-Patch: Unsupervised understanding of ac- tions and relations

    Chenxia Wu, Jiemi Zhang, Silvio Savarese, and Ashutosh Saxena. Watch-n-Patch: Unsupervised understanding of ac- tions and relations. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 4362– 4370, 2015. 6

    2015

  20. [28]

    Unified perceptual parsing for scene understand- ing

    Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understand- ing. In Proceedings of the European conference on computer vision (ECCV), pages 418–434, 2018. 7

    2018

  21. [29]

    Open-world human-object interaction detection via multi-modal prompts

    Jie Yang, Bingliang Li, Ailing Zeng, Lei Zhang, and Ruimao Zhang. Open-world human-object interaction detection via multi-modal prompts. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 16954–16964, 2024. 2

    2024

  22. [30]

    Learning a contact po- tential field for modeling the hand-object interaction

    Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Junming Zhang, Jiefeng Li, and Cewu Lu. Learning a contact po- tential field for modeling the hand-object interaction. IEEE transactions on pattern analysis and machine intelligence ,

  23. [31]

    Pyramid scene parsing network

    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881–2890, 2017. 7

    2017

  24. [32]

    Towards stable human pose estimation via cross-view fu- sion and foot stabilization

    Li’an Zhuo, Jian Cao, Qi Wang, Bang Zhang, and Liefeng Bo. Towards stable human pose estimation via cross-view fu- sion and foot stabilization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 650–659, 2023. 2

    2023

  25. [33]

    End-to-end human object interaction detection with HOI transformer

    Cheng Zou, Bohan Wang, Yue Hu, Junqi Liu, Qian Wu, Yu Zhao, Boxun Li, Chenguang Zhang, Chi Zhang, Yichen Wei, et al. End-to-end human object interaction detection with HOI transformer. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, page...

    2021

Pith tools

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