Pith. sign in

REVIEW 5 major objections 3 minor 1 cited by

HLFormer: Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning

T0 review · 5 major / 3 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 HLFormer's thesis is that the hierarchy of untrimmed videos is better captured in hyperbolic space, so a Lorentz attention branch plus a cone-based loss that keeps text inside its paired video's entailment cone improves partially relevant…

desk verdict Solid first hyperbolic PRVR with a credible architecture, but the SOTA claim is underdetermined by Table 1 until two cited recent baselines and variance numbers are added. read the letter →

arxiv 2507.17402 v2 pith:CARHEXKL submitted 2025-07-23 cs.CV cs.IRcs.MM

Jun Li , Jinpeng Wang , Chaolei Tan , Niu Lian , Long Chen , Yaowei Wang , Min Zhang , Shu-Tao Xia
show 1 more author
Bin Chen
This is my paper
classification cs.CVcs.IRcs.MM
keywords partiallyrelevantvideoretrievalhyperboliclearningLorentzmodelentailmentconevideo-texthybrid-spaceattentionpartialorderpreservation
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

Partially relevant video retrieval asks a model to match a short text query to an untrimmed video in which the described moment appears somewhere, without knowing where. HLFormer's thesis is that the Euclidean geometry used by previous models distorts the natural hierarchy of video content—coarse moments subsume fine frames—and that moving part of the representation into hyperbolic space fixes this. The paper builds a dual attention network whose Lorentz branch encodes video features on the hyperboloid, fuses them with a Euclidean branch through an adaptive interaction module, and adds a Partial Order Preservation Loss that geometrically confines each text embedding inside a cone anchored at its paired video. On ActivityNet Captions, Charades-STA, and TVR, the reported SumR scores (154.9, 78.7, 187.7) exceed the strongest numbers shown for prior PRVR methods (147.6, 74.7, 179.9). If these results hold, hyperbolic structure becomes a useful inductive bias for retrieval from long, untrimmed video.

What carries the argument

The load-bearing object is the Lorentz model of hyperbolic space (the upper sheet of a hyperboloid with the Lorentzian inner product), used through exponential and logarithmic maps for a Lorentz Attention Block that adds a Gaussian multiscale prior to hyperbolic self-attention. The second mechanism is the entailment cone of the Partial Order Preservation Loss, whose half-aperture $HA(v) = \arcsin(2c/\|v_s\|)$ and exterior angle $EA(v,t)$ penalize text embeddings that fall outside the cone. The Mean-Guided Adaptive Interaction Module computes per-timestep aggregation weights from globally pooled features via cross-attention, dynamically blending the Euclidean and Lorentz branches.

What would settle it

Run GMMFormer v2 and Ambiguity-Restrained on the same three datasets, splits, and features used in Table 1; if either matches or exceeds HLFormer's SumR (154.9 on ActivityNet Captions, 78.7 on Charades-STA, 187.7 on TVR), the headline SOTA claim is not supported.

Watch

Extended reading notes

Core claim

The paper claims that hyperbolic geometry is not just a mathematical alternative but the right geometry for PRVR, because untrimmed videos form a hierarchy from frames to moments to whole video, and in flat Euclidean space distant levels can land close together. HLFormer encodes video in both spaces: Lorentz Attention Blocks apply hyperbolic self-attention with Gaussian multiscale priors to expose hierarchical moment structure, while Euclidean Attention Blocks preserve fine-grained local detail, and the Mean-Guided Adaptive Interaction Module weighs their outputs per timestep. The Partial Order Preservation Loss then formalizes the entailment relation 'text ≺ video' by requiring the text embedding to fall inside the Lorentzian entailment cone of its paired video, with zero loss inside and a penalty proportional to the exterior angle outside. The reported result is consistent improvement over all tabulated baselines, with hybrid attention outperforming either single geometry and the cone constraint adding a small but consistent accuracy gain.

Load-bearing premise

The claim that HLFormer consistently surpasses all baselines rests on Table 1 being a complete and fair comparison, yet two recent PRVR methods cited in the paper (GMMFormer v2 [60] and Ambiguity-Restrained [9]) are not listed, and DL-DKD's Charades-STA numbers are absent.

Editorial extensions

If this is right

  • If the reported numbers are correct, HLFormer sets the current best SumR on ActivityNet Captions (154.9 vs 147.6), Charades-STA (78.7 vs 74.7), and TVR (187.7 vs 179.9).
  • The ablation shows hybrid Euclidean–Lorentz attention outperforms either geometry alone, implying the two spaces carry complementary information for moment extraction.
  • The POP loss ablation moves SumR by roughly one point on each dataset, so the main accuracy gain comes from hyperbolic temporal modeling, while the cone loss mainly enforces the desired text-within-video semantics.
  • Removing either the gaze frame-level branch or glance clip-level branch lowers performance across datasets, so multi-granularity input remains necessary even with hyperbolic encoding.

Reading between the lines

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

  • Inference: the same text-inside-video cone constraint could be transferred to other tasks built on entailment, such as video moment retrieval or composed image retrieval, without needing moment annotations.
  • Inference: the modest gain from the POP loss (about one SumR point) suggests the load-bearing part of the method is the Lorentz attention; a testable extension is to replace the cone loss with a simpler hyperbolic margin loss and check whether the same hierarchy is preserved.
  • Inference: the paper's hierarchy argument predicts that gains over Euclidean baselines should grow as datasets contain longer videos with deeper semantic nesting; running HLFormer on longer-form video datasets would test that 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

5 major / 3 minor

Summary. The manuscript proposes HLFormer, a hyperbolic-learning framework for partially relevant video retrieval (PRVR). It encodes video with parallel Lorentz and Euclidean attention blocks, fuses them with a mean-guided adaptive interaction module, and adds a partial-order-preservation loss that confines text embeddings in an entailment cone defined by the corresponding video embedding. Experiments on ActivityNet Captions, Charades-STA, and TVR report SumR values of 154.9, 78.7, and 187.7, which the authors interpret as state-of-the-art. An ablation study and visualizations are included, and code is released.

Significance. If the empirical comparisons are complete, this is a creditable contribution: it is the first hyperbolic formulation for PRVR, the hybrid-space attention design is plausible, the POP loss is a principled adaptation of hyperbolic entailment cones, the ablations support the main design choices, and the code release aids reproducibility. The arithmetic in the tables is coherent, and the reported SumR margins over the listed baselines are substantial. The main uncertainty is whether the state-of-the-art claim is established given two omitted recent PRVR baselines, a per-metric inconsistency on Charades-STA, and the absence of statistical detail; these are fixable issues rather than flaws in the central derivation.

major comments (5)
  1. [Sec. 4.3, Table 1] The headline claim that HLFormer 'consistently surpasses all baselines' is not fully supported because Table 1 omits two PRVR-specific baselines that are cited in the paper itself: GMMFormer v2 [60] and Ambiguity-Restrained [9]. The Charades-STA row for DL-DKD [16] is also listed as '-' rather than with numbers. Without those results under the same feature and split settings, the SumR-based SOTA claim is conditional. Please add the missing comparisons or explicitly justify their exclusion.
  2. [Table 1, Charades-STA] The table shows HLFormer's Charades-STA R@1 of 2.6 is below PEAN's R@1 of 2.7, even though HLFormer's SumR is higher (78.7 vs. 74.7). The statement 'HLFormer consistently surpasses all baselines' is therefore not supported at the per-metric level. The claim should be qualified to the SumR aggregate, or additional evidence should be provided to justify the stronger wording.
  3. [Sec. 3.4, Eq. (18)] The half-aperture formula HA(v) = arcsin(2c/||v_s||) is undefined when ||v_s|| < 2c, which can occur near the origin, including at initialization, since the paper does not state any constraint, clipping, or norm lower bound. The authors should specify how the model avoids this regime and, if a constraint is used, report its effect on optimization and final performance.
  4. [Sec. 4.4, Fig. 5] The visualization evidence that HLFormer learns a 'text ≺ video' hierarchy is partly by construction: the POP loss directly penalizes text embeddings that fall outside the video-defined cone, so Fig. 5(b) largely verifies that the loss is being optimized rather than providing independent evidence that hyperbolic geometry alone creates the hierarchy. A comparison against a Euclidean-cone or margin-based analogue would strengthen the attribution of the observed structure to hyperbolic space.
  5. [Sec. 4.2, Table 1] All reported results are single-run point estimates without variance, confidence intervals, or the number of seeds. This makes it difficult to assess the reliability of the claimed gains, especially on Charades-STA where the R@1 value is below PEAN's. Please report mean ± standard deviation over at least three runs for the main results and key ablations.
minor comments (3)
  1. [Table 1 caption] The word 'faithfull' in the caption should be 'faithful'.
  2. [Figure 3] The axis labels in the provided version of Figure 3 appear garbled; please verify the final PDF rendering so that all axis text is legible.
  3. [Sec. 3.3, Eqs. (12)-(14)] The notation for the Lorentz linear layer is hard to parse; please clarify the dimensions of f_x(M) and the role of the p vector in projecting to the Lorentz manifold.

Circularity Check

1 steps flagged · score 3.0 of 10

One geometric analysis figure restates the POP loss objective by construction; the SOTA retrieval claim is benchmarked externally and is not circular, though Table 1 omits two recent cited PRVR baselines.

  1. self definitional [Sec. 3.4 Eq. (18)-(20); Sec. 4.4 'Visualization of Hyperbolic Space']
    "Lpop(v,t) = max(0,EA(v,t)−HA(v)). ... Glance branch embeddings are positioned closer to the origin than text query embeddings, indicating that clip-level video representations subsume textual queries. This phenomenon can be attributed to Lpop, which enforces the partial order relationship between video and text representations."

    Eq. (20) is the training loss that directly penalizes any text embedding t falling outside the half-aperture cone of its paired video v. The Sec. 4.4 visualization then reports, as a discovered geometric property, exactly the containment relation that this loss was constructed to impose: text embeddings inside or near the video cone, and the w/o Lpop vs. w/ Lpop contrast is a check that the optimizer minimized Eq. (20). The paper itself concedes the construction by saying the phenomenon 'can be attributed to Lpop.' This is not a fitted parameter renamed as a prediction, and it does not drive the main retrieval claim; it is circular only for the qualitative 'entailment' evidence.

full rationale

The central retrieval claim is an empirical benchmark result, not a first-principles derivation: HLFormer is trained with a combination of retrieval loss, diversity loss, and the POP loss, and its SumR scores are compared against external baselines on three standard datasets. That part of the paper is self-contained and not circular. The one genuinely by-construction element is the geometric evidence in Fig. 5: the POP loss (Eq. 20) by definition confines text embeddings within the video-anchored Lorentzian cone, so Fig. 5(b) mostly demonstrates that the loss achieved its stated objective rather than independently confirming a learned hierarchy. The ablation (Table 2) shows the full model only marginally exceeds the w/o Lpop variant (154.9 vs. 154.0 on ActivityNet; 187.7 vs. 186.8 on TVR), so this circular evidence is not the main driver of the reported gains. No load-bearing self-citation chain or imported uniqueness theorem was found; the Lorentz constructions cite prior external work, and MERU/Ganea-style entailment cones are a known building block, so using them is a novelty concern rather than circularity. Separately, the SOTA claim is under-supported empirically because Table 1 omits GMMFormer v2 [60] and Ambiguity-Restrained [9], both cited in Related Work, and leaves DL-DKD Charades-STA entries as '-'; this is an empirical-completeness weakness, not a circular-reasoning defect. Overall, the core retrieval evaluation is independent of its inputs, so the paper is only mildly circular.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on five to six hand-set hyperparameters (α, λ, c, τ, σ, sampling counts) whose values are mostly unreported, and on domain assumptions about video hierarchy and the text-video partial order that are asserted rather than demonstrated. No new physical or conceptual entities are introduced; all components are network modules and loss terms built from existing hyperbolic geometry. The most fragile assumption is the completeness of the baseline comparison in Table 1.

free parameters (6)
  • α_f, α_c (similarity fusion weights) = not reported (α_f + α_c = 1)
    Eq 9 weights frame-level and clip-level similarity; set by hand, with no reported values or sensitivity analysis.
  • λ1, λ2 (loss weights) = not reported
    Eq 21 balances Lsim, Ldiv, and Lpop; λ2 directly controls the POP loss highlighted in the abstract, yet no values or sweeps appear in Sec 4.2.
  • c (cone boundary constant) = 0.1
    Eq 18 defines the half-aperture of the entailment cone near the origin; fixed at 0.1 with no sensitivity analysis.
  • τ (MAIM temperature) = not reported
    Eq 17 softmax temperature for block aggregation weights; unspecified in Sec 4.2.
  • Gaussian variance schedule σ^2 = 2^1 .. 2^{NL-1} and ∞ per block
    Sec 4.2 sets per-block variances for Euclidean and Lorentz attention; the exact schedule is a hand choice.
  • Mf, Mc (frame and clip sampling counts) = not reported
    Sec 4.2 describes dense frame sampling and sparse clip pooling but does not state Mf or Mc.
assumptions (6)
  • standard math Lorentz model geometry: exponential/log maps (Eq 4-5), Lorentzian centroid (Eq 6), and squared Lorentzian distance d^2_L(a,b) = -2 - 2⟨a,b⟩_L are used as given.
    Sec 3.1 and Sec 3.3 rely on standard hyperbolic geometry results from Law et al. and Nickel and Kiela; assumed correct for curvature K=-1.
  • domain assumption Untrimmed videos exhibit a semantic hierarchy (frames, moments, whole video) that benefits PRVR when modeled geometrically.
    Sec 1 and Fig 1(a) motivate the entire hyperbolic design; the hierarchy is asserted as a belief, not measured on the datasets.
  • domain assumption In PRVR, the text query is semantically subsumed by its paired video: text ≺ video.
    Sec 3.4 asserts that "each video fully entails its corresponding text" to justify the POP loss (Eq 18-20); the direction is asserted rather than derived, and its measured impact is small.
  • domain assumption Max-pooled cosine similarity over sampled frames and clips (Eq 8) is a sufficient score for partial relevance.
    Sec 3.2 adopts the max-over-embeddings similarity from MS-SL; it presumes at least one sampled unit is semantically comparable to the query.
  • domain assumption Pre-extracted I3D, ResNet152, and RoBERTa features from prior works are fixed and adequate; no end-to-end video or text training is needed.
    Sec 4.2 uses features from [31, 66, 45, 14]; retrieval performance is bounded by these frozen representations.
  • ad hoc to paper Applying the Gaussian neighborhood prior inside Lorentzian-distance attention (Eq 15) preserves the multiscale benefit of GMMFormer's Gaussian attention.
    The multiplicative combination Mg_σ ⊙ exp(-d^2_L) inside softmax is specific to this paper; its numerical behavior with Lorentz distances is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HLFormer: Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning." pith.science (2026). https://pith.science/paper/CARHEXKL

@misc{pith2026250717402,
  author       = {Pith},
  title        = {Pith review of: HLFormer: Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CARHEXKL}},
  note         = {Machine review of arXiv:2507.17402}
}
read the original abstract

Partially Relevant Video Retrieval (PRVR) addresses the critical challenge of matching untrimmed videos with text queries describing only partial content. Existing methods suffer from geometric distortion in Euclidean space that sometimes misrepresents the intrinsic hierarchical structure of videos and overlooks certain hierarchical semantics, ultimately leading to suboptimal temporal modeling. To address this issue, we propose the first hyperbolic modeling framework for PRVR, namely HLFormer, which leverages hyperbolic space learning to compensate for the suboptimal hierarchical modeling capabilities of Euclidean space. Specifically, HLFormer integrates the Lorentz Attention Block and Euclidean Attention Block to encode video embeddings in hybrid spaces, using the Mean-Guided Adaptive Interaction Module to dynamically fuse features. Additionally, we introduce a Partial Order Preservation Loss to enforce "text < video" hierarchy through Lorentzian cone constraints. This approach further enhances cross-modal matching by reinforcing partial relevance between video content and text queries. Extensive experiments show that HLFormer outperforms state-of-the-art methods. Code is released at https://github.com/lijun2005/ICCV25-HLFormer.

Figures

Figures reproduced from arXiv: 2507.17402 by the authors.

Figure 1
Figure 1. (a) Modeling the semantic hierarchy in untrimmed videos [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of HLFormer. (a) The sentence embedding [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The influence of different attention blocks, with default [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The UMAP [42] visualization displays the learned frame embeddings from a video in TVR. Data points of the same color correspond to the same moment. As shown in Tab. 2, the worst performance occurs when only Lsim is used. Comparing Variant (5) with Variant (3), adding L…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. H3Former: Hypergraph-based Semantic-Aware Aggregation via Hyperbolic Hierarchical Contrastive Loss for Fine-Grained Visual Classification

    cs.CV 2025-11 conditional novelty 5.0 of 10

    A hypergraph-based token-to-region aggregation plus a hyperbolic hierarchical contrastive loss yields reported state-of-the-art fine-grained classification accuracy on four benchmarks.

Reference graph

Works this paper leans on

68 extracted references · 59 canonical work pages · cited by 1 Pith paper

  1. [60]

    Gmmformer v2: An uncertainty-aware framework for partially relevant video retrieval, 2024

    Yuting Wang, Jinpeng Wang, Bin Chen, Tao Dai, Ruisheng Luo, and Shu-Tao Xia. Gmmformer v2: An uncertainty-aware framework for partially relevant video retrieval, 2024. 1, 2

    work page 2024

  2. [9]

    Ambiguity-restrained text-video representa- tion learning for partially relevant video retrieval

    Cheol-Ho Cho, WonJun Moon, Woojin Jun, MinSeok Jung, and Jae-Pil Heo. Ambiguity-restrained text-video representa- tion learning for partially relevant video retrieval. InProceed- ings of the AAAI Conference on Artificial Intelligence, pages 2500–2508, 2025. 2

    work page 2025

  3. [16]

    Dual learning with dynamic knowledge distillation for partially relevant video retrieval

    Jianfeng Dong, Minsong Zhang, Zheng Zhang, Xianke Chen, Daizong Liu, Xiaoye Qu, Xun Wang, and Baolong Liu. Dual learning with dynamic knowledge distillation for partially relevant video retrieval. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 11302– 11312, 2023. 2, 6, 7

    work page 2023

  4. [1]

    Hyperbolic image segmen- tation

    Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmen- tation. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 4453–4462, 2022. 3

    work page 2022

  5. [2]

    Quo vadis, action recognition? a new model and the kinetics dataset

    Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. Inpro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299–6308, 2017. 6

    2017

  6. [3]

    Horopca: Hyperbolic dimensionality reduction via horospher- ical projections, 2021

    Ines Chami, Albert Gu, Dat Nguyen, and Christopher R ´e. Horopca: Hyperbolic dimensionality reduction via horospher- ical projections, 2021. 8

    work page 2021

  7. [4]

    Hy- perbolic uncertainty aware semantic segmentation.IEEE Transactions on Intelligent Transportation Systems, 25(2): 1275–1290, 2023

    Bike Chen, Wei Peng, Xiaofeng Cao, and Juha R¨oning. Hy- perbolic uncertainty aware semantic segmentation.IEEE Transactions on Intelligent Transportation Systems, 25(2): 1275–1290, 2023. 3

    work page 2023

  8. [5]

    Fine-grained video-text retrieval with hierarchical graph reasoning

    Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10638–10647, 2020. 1, 2, 6, 7

    work page 2020

Show all 68 references
  1. [6]

    Fully hyperbolic neural networks.arXiv preprint arXiv:2105.14686, 2021

    Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks.arXiv preprint arXiv:2105.14686, 2021. 3, 5

    2021 arXiv

  2. [7]

    Joint searching and grounding: Multi- granularity video content retrieval

    Zhiguo Chen, Xun Jiang, Xing Xu, Zuo Cao, Yijun Mo, and Heng Tao Shen. Joint searching and grounding: Multi- granularity video content retrieval. InProceedings of the 31st ACM International Conference on Multimedia, pages 975–983, 2023. 2, 6, 7

    2023

  3. [8]

    Transferable dual multi-granularity semantic excavating for partially relevant video retrieval.Image and Vision Comput- ing, 149:105168, 2024

    Dingxin Cheng, Shuhan Kong, Bin Jiang, and Qiang Guo. Transferable dual multi-granularity semantic excavating for partially relevant video retrieval.Image and Vision Comput- ing, 149:105168, 2024. 2

    2024

  4. [10]

    Hyperbolic Image- Text Representations

    Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Ramakrishna Vedantam. Hyperbolic Image- Text Representations. InProceedings of the International Conference on Machine Learning, 2023. 2, 3, 5

    2023

  5. [11]

    Predicting visual features from text for image and video caption retrieval

    Jianfeng Dong, Xirong Li, and Cees GM Snoek. Predicting visual features from text for image and video caption retrieval. IEEE Transactions on Multimedia, 20(12):3377–3388, 2018. 1, 2

    2018

  6. [12]

    Dual encoding for zero- example video retrieval

    Jianfeng Dong, Xirong Li, Chaoxi Xu, Shouling Ji, Yuan He, Gang Yang, and Xun Wang. Dual encoding for zero- example video retrieval. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9346–9355, 2019

    2019

  7. [13]

    Dual encoding for video retrieval by text.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4065–4080, 2021

    Jianfeng Dong, Xirong Li, Chaoxi Xu, Xun Yang, Gang Yang, Xun Wang, and Meng Wang. Dual encoding for video retrieval by text.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4065–4080, 2021. 1, 6, 7

    2021

  8. [14]

    Partially relevant video retrieval

    Jianfeng Dong, Xianke Chen, Minsong Zhang, Xun Yang, Shujie Chen, Xirong Li, and Xun Wang. Partially relevant video retrieval. InProceedings of the 30th ACM International Conference on Multimedia, pages 246–257, 2022. 1, 2, 4, 6, 7

    2022

  9. [15]

    Reading-strategy in- spired visual representation learning for text-to-video retrieval

    Jianfeng Dong, Yabing Wang, Xianke Chen, Xiaoye Qu, Xirong Li, Yuan He, and Xun Wang. Reading-strategy in- spired visual representation learning for text-to-video retrieval. IEEE Transactions on Circuits and Systems for Video Tech- nology, 32(8):5680–5694, 2022. 1, 2, 6, 7

    2022

  10. [17]

    Hyperbolic vision trans- formers: Combining improvements in metric learning

    Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision trans- formers: Combining improvements in metric learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7399–7409, 2022. 2, 3

    2022

  11. [18]

    Vse++: Improving visual-semantic embeddings with hard negatives.arXiv preprint arXiv:1707.05612, 2017

    Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives.arXiv preprint arXiv:1707.05612, 2017. 1, 2

    2017 arXiv

  12. [19]

    Grounding language with vision: A conditional mutual information cali- brated decoding strategy for reducing hallucinations in lvlms

    Hao Fang, Changle Zhou, Jiawei Kong, Kuofeng Gao, Bin Chen, Tao Liang, Guojun Ma, and Shu-Tao Xia. Grounding language with vision: A conditional mutual information cali- brated decoding strategy for reducing hallucinations in lvlms. arXiv preprint arXiv:2505.19678, 2025. 2

    2025

  13. [20]

    Linguistic hallucination for text-based video retrieval

    Sheng Fang, Tiantian Dang, Shuhui Wang, and Qingming Huang. Linguistic hallucination for text-based video retrieval. IEEE Transactions on Circuits and Systems for Video Tech- nology, 34(10):9692–9705, 2024. 6, 7

    2024

  14. [21]

    Hy- perbolic entailment cones for learning hierarchical embed- dings

    Octavian Ganea, Gary Becigneul, and Thomas Hofmann. Hy- perbolic entailment cones for learning hierarchical embed- dings. InProceedings of the 35th International Conference on Machine Learning, pages 1646–1655. PMLR, 2018

    2018

  15. [22]

    Hy- perbolic neural networks.Advances in neural information processing systems, 31, 2018

    Octavian Ganea, Gary B´ecigneul, and Thomas Hofmann. Hy- perbolic neural networks.Advances in neural information processing systems, 31, 2018. 3

    2018

  16. [23]

    Tall: Temporal activity localization via language query

    Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on com- puter vision, pages 5267–5275, 2017. 2, 6

    2017

  17. [24]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 6

    2016

  18. [25]

    Lorentzian residual neural networks.arXiv preprint arXiv:2412.14695, 2024

    Neil He, Menglin Yang, and Rex Ying. Lorentzian residual neural networks.arXiv preprint arXiv:2412.14695, 2024. 3

    2024 arXiv

  19. [26]

    Con- quer: Contextual query-aware ranking for video corpus mo- ment retrieval

    Zhijian Hou, Chong-Wah Ngo, and Wing Kwong Chan. Con- quer: Contextual query-aware ranking for video corpus mo- ment retrieval. InProceedings of the 29th ACM International Conference on Multimedia, pages 3900–3908, 2021. 6, 7

    2021

  20. [27]

    Progressive event alignment network for partial relevant video retrieval

    Xun Jiang, Zhiguo Chen, Xing Xu, Fumin Shen, Zuo Cao, and Xunliang Cai. Progressive event alignment network for partial relevant video retrieval. In2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1973–

    1973

  21. [28]

    Hyperbolic image embeddings

    Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3

    2020

  22. [29]

    Dense-captioning events in videos

    Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on com- puter vision, pages 706–715, 2017. 2, 6

    2017

  23. [30]

    Lorentzian distance learning for hyperbolic representations

    Marc Law, Renjie Liao, Jake Snell, and Richard Zemel. Lorentzian distance learning for hyperbolic representations. InProceedings of the 36th International Conference on Ma- chine Learning, pages 3672–3681. PMLR, 2019. 2

    2019

  24. [31]

    Tvr: A large-scale dataset for video-subtitle moment retrieval

    Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16, pages 447–463. Springer, 2020. 2, 6, 7

    2020

  25. [32]

    Beyond euclidean: Dual- space representation learning for weakly supervised video violence detection

    Jiaxu Leng, Zhanjie Wu, Mingpi Tan, Yiran Liu, Ji Gan, Haosheng Chen, and Xinbo Gao. Beyond euclidean: Dual- space representation learning for weakly supervised video violence detection. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3

    2024

  26. [33]

    Fully hyper- bolic convolutional neural networks.Research in the Mathe- matical Sciences, 9(4):60, 2022

    Keegan Lensink, Bas Peters, and Eldad Haber. Fully hyper- bolic convolutional neural networks.Research in the Mathe- matical Sciences, 9(4):60, 2022. 3, 5

    2022

  27. [34]

    Hy- perbolic anomaly detection

    Huimin Li, Zhentao Chen, Yunhao Xu, and Junlin Hu. Hy- perbolic anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17511–17520, 2024. 3

    2024

  28. [35]

    W2vv++ fully deep learning for ad-hoc video search

    Xirong Li, Chaoxi Xu, Gang Yang, Zhineng Chen, and Jian- feng Dong. W2vv++ fully deep learning for ad-hoc video search. InProceedings of the 27th ACM international confer- ence on multimedia, pages 1786–1794, 2019. 1, 2

    2019

  29. [36]

    Protecting your video con- tent: Disrupting automated video-based llm annotations

    Haitong Liu, Kuofeng Gao, Yang Bai, Jinmin Li, Jinxiao Shan, Tao Dai, and Shu-Tao Xia. Protecting your video con- tent: Disrupting automated video-based llm annotations. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24056–24065, 2025. 2

    2025

  30. [37]

    Multi-task ranking with user behaviors for text-video search

    Peidong Liu, Dongliang Liao, Jinpeng Wang, Yangxin Wu, Gongfu Li, Shu-Tao Xia, and Jin Xu. Multi-task ranking with user behaviors for text-video search. InCompanion Proceedings of the Web Conference 2022, pages 126–130,

    2022

  31. [38]

    Use what you have: Video retrieval using representations from collaborative experts.arXiv preprint arXiv:1907.13487, 2019

    Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using representations from collaborative experts.arXiv preprint arXiv:1907.13487, 2019. 1, 2, 6, 7

    1907 arXiv

  32. [39]

    Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692,

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettle- moyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692,

    1907 arXiv

  33. [40]

    Searching for actions on the hyperbole

    Teng Long, Pascal Mettes, Heng Tao Shen, and Cees GM Snoek. Searching for actions on the hyperbole. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1141–1150, 2020. 3

    2020

  34. [41]

    Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning

    Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293–304, 2022. 6, 7

    2022

  35. [42]

    Umap: Uniform manifold approximation and projection for dimen- sion reduction.arXiv preprint arXiv:1802.03426, 2018

    Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimen- sion reduction.arXiv preprint arXiv:1802.03426, 2018. 8

    2018 arXiv

  36. [43]

    Evdclip: Improving vision-language retrieval with entity visual descriptions from large language models

    Guanghao Meng, Sunan He, Jinpeng Wang, Tao Dai, Letian Zhang, Jieming Zhu, Qing Li, Gang Wang, Rui Zhang, and Yong Jiang. Evdclip: Improving vision-language retrieval with entity visual descriptions from large language models. In Proceedings of the AAAI Conference on Artificia...

    2025

  37. [44]

    Howto100m: Learning a text-video embedding by watching hundred mil- lion narrated video clips

    Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred mil- lion narrated video clips. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages...

    2019

  38. [45]

    Local-global video-text interactions for temporal grounding

    Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810–10819, 2020. 6

    2020

  39. [46]

    Poincar ´e embeddings for learning hierarchical representations

    Maximillian Nickel and Douwe Kiela. Poincar ´e embeddings for learning hierarchical representations. InAdvances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3

    2017

  40. [47]

    Learning continuous hierarchies in the Lorentz model of hyperbolic geometry

    Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the Lorentz model of hyperbolic geometry. In Proceedings of the 35th International Conference on Machine Learning, pages 3779–3788. PMLR, 2018. 3

    2018

  41. [48]

    Compositional entailment learning for hyperbolic vision-language models

    Avik Pal, Max van Spengler, Guido Maria D’Amely di Me- lendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. InThe Thirteenth International Con- ference on Learning Representations, 2025. 3, 8

    2025

  42. [49]

    Learning weakly supervised audio-visual violence detection in hyperbolic space, 2023

    Xiaogang Peng, Hao Wen, Yikai Luo, Xiao Zhou, Keyang Yu, Yigang Wang, and Zizhao Wu. Learning weakly supervised audio-visual violence detection in hyperbolic space, 2023. 3

    2023

  43. [50]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, p...

    2021

  44. [51]

    HOVER: Hyperbolic video-text re- trieval, 2024

    Ruiqi Shi, Jun Wen, Wei Ji, Menglin Yang, Difei Gao, and Roger Zimmermann. HOVER: Hyperbolic video-text re- trieval, 2024. 3

    2024

  45. [52]

    Spatial-temporal graphs for cross-modal text2video retrieval

    Xue Song, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Spatial-temporal graphs for cross-modal text2video retrieval. IEEE Transactions on Multimedia, 24:2914–2923, 2021. 2

    2021

  46. [53]

    Siamese learning with joint alignment and regression for weakly-supervised video paragraph grounding

    Chaolei Tan, Jianhuang Lai, Wei-Shi Zheng, and Jian-Fang Hu. Siamese learning with joint alignment and regression for weakly-supervised video paragraph grounding. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13569–13580, 2024. 2

    2024

  47. [54]

    Modeling uncertainty in composed image retrieval via probabilistic embeddings

    Haomiao Tang, Jinpeng Wang, Yuang Peng, GuangHao Meng, Ruisheng Luo, Bin Chen, Long Chen, Yaowei Wang, and Shu- Tao Xia. Modeling uncertainty in composed image retrieval via probabilistic embeddings. InProceedings of the 63rd Annual Meeting of the Association for Computational...

    2025

  48. [55]

    Missing target-relevant in- formation prediction with world model for accurate zero-shot composed image retrieval

    Yuanmin Tang, Jing Yu, Keke Gai, Jiamin Zhuang, Gang Xiong, Gaopeng Gou, and Qi Wu. Missing target-relevant in- formation prediction with world model for accurate zero-shot composed image retrieval. InProceedings of the Computer Vision and Pattern Recognition Conference, pages...

    2025

  49. [56]

    Poincar´e resnet

    Max Van Spengler, Erwin Berkhout, and Pascal Mettes. Poincar´e resnet. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 5419–5428,

  50. [57]

    Attention is all you need.Advances in neural information processing systems, 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.Advances in neural information processing systems, 30, 2017. 3

    2017

  51. [58]

    Hybrid contrastive quantization for efficient cross-view video retrieval

    Jinpeng Wang, Bin Chen, Dongliang Liao, Ziyun Zeng, Gongfu Li, Shu-Tao Xia, and Jin Xu. Hybrid contrastive quantization for efficient cross-view video retrieval. InPro- ceedings of the ACM Web Conference 2022, pages 3020–3030,

    2022

  52. [59]

    Hugs bring double benefits: Unsupervised cross-modal hash- ing with multi-granularity aligned transformers.International Journal of Computer Vision, 132(8):2765–2797, 2024

    Jinpeng Wang, Ziyun Zeng, Bin Chen, Yuting Wang, Dongliang Liao, Gongfu Li, Yiru Wang, and Shu-Tao Xia. Hugs bring double benefits: Unsupervised cross-modal hash- ing with multi-granularity aligned transformers.International Journal of Computer Vision, 132(8):2765–2797, 2024. 2

    2024

  53. [61]

    Gmmformer: Gaussian-mixture-model based transformer for efficient partially relevant video retrieval

    Yuting Wang, Jinpeng Wang, Bin Chen, Ziyun Zeng, and Shu-Tao Xia. Gmmformer: Gaussian-mixture-model based transformer for efficient partially relevant video retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence,

  54. [62]

    An empirical study of federated prompt learning for vision language model.arXiv preprint arXiv:2505.23024,

    Zhihao Wang, Wenke Huang, Tian Chen, Zekun Shi, Guancheng Wan, Yu Qiao, Bin Yang, Jian Wang, Bing Li, and Mang Ye. An empirical study of federated prompt learning for vision language model.arXiv preprint arXiv:2505.23024,

  55. [63]

    Cap4video: What can auxiliary captions do for text-video retrieval? InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10704–10713, 2023

    Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10704–10713, 2023. 6, 7

    2023

  56. [64]

    Exploiting instance-level relationships in weakly su- pervised text-to-video retrieval.ACM Trans

    Shukang Yin, Sirui Zhao, Hao Wang, Tong Xu, and Enhong Chen. Exploiting instance-level relationships in weakly su- pervised text-to-video retrieval.ACM Trans. Multim. Comput. Commun. Appl., 20(10):316:1–316:21, 2024. 2, 6, 7

    2024

  57. [65]

    Skin lesion recognition with class- hierarchy regularized hyperbolic embeddings

    Zhen Yu, Toan Nguyen, Yaniv Gal, Lie Ju, Shekhar S Chan- dra, Lei Zhang, Paul Bonnington, Victoria Mar, Zhiyong Wang, and Zongyuan Ge. Skin lesion recognition with class- hierarchy regularized hyperbolic embeddings. InInterna- tional conference on medical image computing and c...

    2022

  58. [66]

    A hierarchical multi-modal encoder for moment localization in video corpus.arXiv preprint arXiv:2011.09046, 2020

    Bowen Zhang, Hexiang Hu, Joonseok Lee, Ming Zhao, Sheide Chammas, Vihan Jain, Eugene Ie, and Fei Sha. A hierarchical multi-modal encoder for moment localization in video corpus.arXiv preprint arXiv:2011.09046, 2020. 6

    2011 arXiv

  59. [67]

    Video corpus moment retrieval with contrastive learning

    Hao Zhang, Aixin Sun, Wei Jing, Guoshun Nan, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Video corpus moment retrieval with contrastive learning. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, page...

  60. [68]

    Keyword-based di- verse image retrieval by semantics-aware contrastive learning and transformer

    Minyi Zhao, Jinpeng Wang, Dongliang Liao, Yiru Wang, Huanzhong Duan, and Shuigeng Zhou. Keyword-based di- verse image retrieval by semantics-aware contrastive learning and transformer. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in ...

    2023

Pith tools

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