Pith. sign in

REVIEW 4 major objections 6 minor 77 references

A multi-scale convolutional encoder plus capsule dynamic routing yields more precise video-text temporal localization than transformer baselines on ActivityNet Captions.

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 ↗

T0 review · grok-4.5

2026-07-11 08:54 UTC pith:BMU44ZV5

load-bearing objection Clean multi-scale + capsule combo with solid internal ablations on ActivityNet, but the 'surpassing strong baselines' claim rests only on a weak vanilla transformer and needs external numbers. the 4 major comments →

arxiv 2607.05093 v2 pith:BMU44ZV5 submitted 2026-07-06 cs.CV

Video-Text Temporal Localization via Multi-Scale Convolution and Dynamic Routing

Gengtian Shi , Jinze Yu , Chenhao Wu , Shaofei Wang , Eiji Fukuzawa , Junjie Tang , Hiroshi Onoda , Jiang Liu This is my paper
classification cs.CV
keywords video-text temporal localizationmulti-scale temporal convolutioncapsule dynamic routingmany-to-many alignmentActivityNet Captionstemporal sentence groundingcross-modal multi-task learning
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 paper sets out to show that video-text temporal localization fails when models treat all time scales the same and force simple one-to-one alignments. Real videos contain overlapping actions, disjoint segments for one query, and non-monotonic matches; global attention and fixed windows miss these. The authors therefore build an encoder that runs parallel 1-D convolutions of different kernel sizes to capture both instantaneous transitions and longer action arcs, then route the resulting features to text tokens through iterative capsule agreement. Jointly trained with boundary regression, contrastive alignment and capsule-diversity losses, the system reaches 42.9 percent Recall@0.5 and 41.1 percent mean IoU while staying linear in video length. A reader who cares about searchable instructional video, robot instruction following or interactive editing gains a concrete, efficient recipe for finer moment-level grounding.

Core claim

Hierarchical temporal modeling via multi-scale 1-D convolutions combined with capsule-based dynamic routing produces structured many-to-many video-text alignments that improve localization accuracy to 42.9 percent Recall@0.5 and 41.1 percent mean IoU on ActivityNet Captions, outperforming strong frozen-transformer baselines without end-to-end fine-tuning of large backbones.

What carries the argument

The multi-scale temporal convolutional encoder (parallel kernels of size 1, 3 and 5 with adaptive 1x1 fusion and residual connection) together with the capsule dynamic-routing module that iteratively updates coupling coefficients by agreement; the pair supplies both multi-granularity motion features and sparse, interpretable cross-modal assignments that feed a lightweight boundary head.

Load-bearing premise

The reported gains rest on the premise that internal ablations against a vanilla transformer baseline, rather than head-to-head numbers against recent published state-of-the-art methods on the identical split, sufficiently prove superiority for complex many-to-many alignments.

What would settle it

Retrain the full model and the strongest published competitors (for example 2D-TAN or LocVTP) on the exact same ActivityNet Captions train/val split with identical frame sampling and evaluation protocol; if the multi-scale-plus-capsule system no longer leads on R@0.5 and mIoU, the central claim fails.

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

If this is right

  • Long untrimmed videos can be localized in linear rather than quadratic time while still resolving both short transitions and extended actions.
  • Queries that map to multiple disjoint or overlapping segments become tractable without assuming monotonic order.
  • Frozen CLIP-style backbones plus a lightweight multi-task head suffice for competitive moment retrieval, lowering the cost of deployment.
  • Soft expectation decoding of start/end logits yields smoother boundaries than hard argmax, improving mean IoU on gradual transitions.
  • Capsule diversity regularization can be scheduled after initial training to avoid collapse while still specializing semantic parts.

Where Pith is reading between the lines

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

  • The same multi-scale-plus-routing stack should transfer to dense video captioning and video question answering where temporal boundaries are also ambiguous.
  • Because text-encoder quality produced larger gains than vision-backbone scaling, future work may profit more from stronger linguistic models than from larger visual transformers.
  • The O(T) design invites real-time streaming localization on edge devices once the routing iterations are further quantized or pruned.
  • If annotation noise is the dominant remaining error, replacing Gaussian-smoothed binary labels with learned uncertainty could raise high-IoU recall without new architecture.

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

4 major / 6 minor

Summary. The paper proposes a video–text temporal localization framework that combines a multi-scale 1D convolutional temporal encoder (parallel kernels, adaptive 1×1 fusion, residual connection) with a capsule-based dynamic routing module that iteratively refines frame–token coupling coefficients via agreement updates and squashing. Features come from frozen CLIP (and optional BERT/RoBERTa) encoders projected to a shared space; a lightweight MLP head predicts start/end boundaries under a multi-task loss (BCE localization with Gaussian-smoothed labels, InfoNCE alignment, and capsule diversity). On ActivityNet Captions the best reported configuration (ViT-B/32 + BERT-base) reaches 42.9% Recall@0.5 and 41.1% mIoU. Internal ablations (Tables 3–4) attribute gains to the multi-scale and capsule modules and claim synergy; the abstract and conclusion assert that the method surpasses strong transformer-based baselines while remaining computationally efficient (O(T) multi-scale path).

Significance. If the architectural claims hold under proper external comparison, the work offers a practical, frozen-backbone design that couples efficient hierarchical temporal modeling with structured routing as an alternative to pure attention for moment retrieval. The ablations cleanly isolate multi-scale convolution and capsule routing and show super-additive gains, which is a genuine experimental strength. Efficiency arguments (O(T) vs O(T²) attention) and multi-task supervision with a simple curriculum for diversity are useful for deployable systems. The contribution is incremental rather than foundational, but a well-validated multi-scale + routing recipe would be of interest to the video–language community provided the SOTA claim is substantiated.

major comments (4)
  1. [Main Results / Abstract / Conclusion] Abstract, Conclusion, and Main Results claim 42.9% R@0.5 / 41.1% mIoU while 'surpassing strong transformer-based baselines.' Tables 1–4 report only internal variants (CLIP backbones, text encoders, progressive addition of multi-scale and capsule modules). The ablation baseline is a weak 'Transformer only' model (R@0.5 = 0.311 / 0.3488). No head-to-head numbers appear against published ActivityNet Captions methods under the same split and protocol (e.g., VSLNet, 2D-TAN, DRN, LocVTP, or recent transformer moment-retrieval models discussed in Related Work). Without that comparison table the central superiority claim is not evidenced and cannot be assessed.
  2. [Experiments / Experimental Setup] Evaluation is restricted to ActivityNet Captions (37k/17k split). Standard temporal grounding benchmarks (Charades-STA, TACoS, or DiDeMo) are absent. Given that the paper’s second core claim is improved handling of complex many-to-many / non-monotonic alignments, single-dataset results leave open whether the multi-scale + routing design generalizes or is tuned to ActivityNet’s length and annotation style.
  3. [Capsule-Based Semantic Alignment / Ablation Study] The many-to-many / non-monotonic alignment advantage of capsule routing over softmax attention is asserted (Introduction, Methodology, Design Rationale) but not directly measured. There is no qualitative alignment visualization, no controlled non-monotonic subset analysis, and no comparison of alignment sparsity or multi-segment coverage against an attention baseline with matched capacity. Tables 3–4 show that adding capsules helps, but do not establish that the mechanism specifically solves the correspondence problem claimed in the abstract.
  4. [Multi-Scale Temporal Feature Encoding / Implementation Details] Methodology (Eqs. 3–6, Figure 3) specifies parallel kernels k ∈ {1,3,5}; Implementation Details state kernels {1,3,5,7} with 96-dimensional outputs. Which configuration produced Tables 1–4 is unclear. The same section mentions a Context-Aware Cross-Attention module in Figure 2 that is not formalized in the text. These inconsistencies make the reported system non-reproducible from the manuscript alone.
minor comments (6)
  1. [Temporal Localization Head] Soft expectation vs. argmax decoding (Eqs. 14–15) is discussed, but the tables do not state which decoder produced the reported R@K and mIoU numbers.
  2. [Unified Multi-Task Loss] Loss weights λ_loc=1.0, λ_align=0.5, λ_div=0.1 and the delayed activation of L_div after 10 epochs are given without sensitivity analysis; a short ablation or validation curve would strengthen the multi-task design claim.
  3. [Design Rationale and Efficiency Analysis] Efficiency is argued via asymptotic O(T) vs O(T²) only; no wall-clock, FLOPs, or parameter-count comparison against the transformer baseline appears.
  4. [Introduction / Figure 1] Figure 1 caption and body text refer to complex overlapping queries, but no quantitative breakdown of performance on multi-segment or overlapping queries is provided.
  5. [Feature Extraction / Capsule routing] Minor notation: V′/Q′ projections (Eqs. 1–2) use D then switch to 384 without restating dimensions consistently; capsule vote u_ij (Eq. 7) reuses index conventions that could be clarified.
  6. [Related Work] Related Work cites several 2023–2024 methods; ensure all arXiv-only works are consistently formatted and that any concurrent work on multi-scale temporal grounding is acknowledged.

Circularity Check

0 steps flagged

No circularity: empirical multi-scale + capsule architecture evaluated on public ActivityNet Captions split with conventional losses; metrics are not forced by construction or self-citation.

full rationale

The paper proposes an architecture (parallel 1D multi-scale convolutions fused with residual + LayerNorm, followed by capsule dynamic routing with standard agreement updates and squash, plus a two-layer MLP localization head) trained end-to-end with a multi-task loss that is the ordinary sum of BCE boundary regression (with Gaussian-smoothed targets), InfoNCE alignment, and a diversity regularizer on capsule vectors. All components are defined independently of the reported Recall@K / mIoU numbers; the numbers are obtained by training on the standard ActivityNet Captions train split and evaluating on the held-out validation split. Loss weights and the delayed activation of the diversity term are ordinary hyper-parameters chosen on validation, not quantities that algebraically determine the final metrics. There are no self-definitional equations, no fitted constants re-labeled as predictions, no uniqueness theorems imported from overlapping authors, and no load-bearing self-citations that close a derivation loop. The work is therefore self-contained against an external public benchmark; any weakness lies in missing external SOTA head-to-heads, which is an evidence gap rather than circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 2 invented entities

The central empirical claim rests on standard deep-learning practice plus a handful of hand-chosen architectural and training hyper-parameters. No new physical or mathematical axioms are introduced; the free parameters are ordinary model choices that affect the reported metrics.

free parameters (5)
  • loss weights λ_loc, λ_align, λ_div = 1.0 / 0.5 / 0.1
    Set to 1.0 / 0.5 / 0.1 'based on validation performance'; directly scales the multi-task objective that produces the final numbers.
  • routing iterations R = 3
    Fixed at 3; controls how many agreement updates the capsule module performs and therefore the final alignment matrix.
  • convolution kernel set = {1,3,5} or {1,3,5,7}
    Chosen as {1,3,5} (text) or {1,3,5,7} (implementation); determines the temporal receptive fields that feed the rest of the model.
  • diversity-loss activation epoch = after 10 epochs
    Diversity term is turned on only after epoch 10; a curriculum choice that affects capsule specialization and final accuracy.
  • Gaussian boundary smoothing σ = 0.1
    σ=0.1 applied to ground-truth start/end labels; softens supervision and influences localization loss.
axioms (3)
  • domain assumption Frozen CLIP visual features plus a linear projection suffice as input representation for fine-grained temporal localization.
    Stated in Feature Extraction; all subsequent modules operate on these frozen embeddings.
  • ad hoc to paper Capsule dynamic routing with squashing produces sparse, many-to-many alignments superior to softmax attention for non-monotonic correspondences.
    Core modeling claim of the Dynamic Routing section; not independently verified outside the paper's own ablations.
  • domain assumption ActivityNet Captions validation split is a faithful proxy for real-world temporal localization difficulty.
    Standard benchmark assumption used for all reported metrics.
invented entities (2)
  • multi-scale temporal convolutional encoder with adaptive 1x1 fusion for video-text localization no independent evidence
    purpose: Capture hierarchical motion patterns at multiple granularities while keeping O(T) cost.
    Architectural module assembled from known 1D convolutions; no independent theoretical existence outside this paper.
  • capsule-based dynamic routing alignment module for video frames and text tokens no independent evidence
    purpose: Iteratively refine many-to-many cross-modal correspondences via agreement updates.
    Adaptation of capsule routing to the video-text setting; evidence is only the paper's own ablation gains.

pith-pipeline@v1.1.0-grok45 · 14977 in / 2940 out tokens · 20133 ms · 2026-07-11T08:54:43.903007+00:00 · methodology

0 comments
read the original abstract

Video-text temporal localization requires precise alignment between natural language queries and corresponding video segments, a fundamental challenge in multimodal understanding. We present a novel framework that addresses two critical limitations of existing methods: inadequate modeling of hierarchical temporal structure and inability to handle complex many-to-many correspondences between modalities. Our approach introduces a multi-scale temporal convolutional encoder that captures motion patterns across different temporal granularities - from instantaneous frame transitions to extended action sequences. We further propose a capsule-based dynamic routing mechanism that iteratively refines segment-query associations through structured agreement updates, enabling flexible modeling of non-monotonic alignments. These components are unified through a multi-task learning objective that jointly optimizes temporal boundary regression, cross-modal semantic alignment, and capsule diversity. Extensive experiments on ActivityNet Captions demonstrate significant improvements, achieving 42.9% Recall@0.5 and 41.1% mean IoU, surpassing strong transformer-based baselines while maintaining computational efficiency. Our results validate that combining hierarchical temporal modeling with structured semantic routing provides an effective solution for fine-grained video-language understanding.

Figures

Figures reproduced from arXiv: 2607.05093 by Chenhao Wu, Eiji Fukuzawa, Gengtian Shi, Hiroshi Onoda, Jiang Liu, Jinze Yu, Junjie Tang, Shaofei Wang.

Figure 1
Figure 1. Figure 1: Temporal localization challenges in real-world videos. Complex queries like ”People sail on inflatable boats down [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture overview for video–text temporal localization. The Multi-Scale Video Encoder processes frame se [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Multi-scale temporal encoder architecture. Parallel [PITH_FULL_IMAGE:figures/full_fig_p004_3.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

77 extracted references · 1 canonical work pages

  1. [1]

    ACM Transactions on Multimedia Computing, Communications and Applications , volume=

    A survey on temporal sentence grounding in videos , author=. ACM Transactions on Multimedia Computing, Communications and Applications , volume=. 2023 , publisher=

    2023

  2. [2]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Temporal sentence grounding in videos: A survey and future directions , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2023 , publisher=

    2023

  3. [3]

    Proceedings of the IEEE international conference on computer vision , pages=

    Weakly-supervised alignment of video with text , author=. Proceedings of the IEEE international conference on computer vision , pages=

  4. [5]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  5. [6]

    International conference on machine learning , pages=

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  6. [7]

    arXiv preprint arXiv:2209.06430 , year=

    Clip-vip: Adapting pre-trained image-text model to video-language representation alignment , author=. arXiv preprint arXiv:2209.06430 , year=

    Pith/arXiv arXiv

  7. [8]

    European conference on computer Vision , pages=

    Connectionist temporal modeling for weakly supervised action labeling , author=. European conference on computer Vision , pages=. 2016 , organization=

    2016

  8. [9]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Msr-vtt: A large video description dataset for bridging video and language , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  9. [10]

    Proceedings of the ieee conference on computer vision and pattern recognition , pages=

    Activitynet: A large-scale video benchmark for human activity understanding , author=. Proceedings of the ieee conference on computer vision and pattern recognition , pages=

  10. [11]

    Proceedings of the 30th ACM international conference on multimedia , pages=

    X-clip: End-to-end multi-grained contrastive learning for video-text retrieval , author=. Proceedings of the 30th ACM international conference on multimedia , pages=

  11. [12]

    Advances in neural information processing systems , volume=

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence , author=. Advances in neural information processing systems , volume=

  12. [13]

    arXiv preprint arXiv:2109.08039 , year=

    A Survey on Temporal Sentence Grounding in Videos , author=. arXiv preprint arXiv:2109.08039 , year=

    Pith/arXiv arXiv

  13. [14]

    Neurocomputing , year=

    Aligning the Background Penalty and the Video-Level Joint Representation for Weakly Supervised Temporal Sentence Grounding , author=. Neurocomputing , year=

  14. [15]

    NeurIPS , volume=

    Semantic Conditioned Dynamic Modulation for Temporal Sentence Grounding in Videos , author=. NeurIPS , volume=. 2019 , url=

    2019

  15. [16]

    EMNLP , year=

    Span-based Question Answering for Temporal Localization in Video , author=. EMNLP , year=

  16. [17]

    AAAI , volume=

    Learning 2D Temporal Adjacent Networks for Moment Localization with Natural Language , author=. AAAI , volume=

  17. [18]

    arXiv preprint arXiv:2312.14055 , year=

    A Strong Baseline for Temporal Video-Text Alignment , author=. arXiv preprint arXiv:2312.14055 , year=

    Pith/arXiv arXiv

  18. [19]

    ICCV , pages=

    VideoBERT: A Joint Model for Video and Language Representation Learning , author=. ICCV , pages=. 2019 , url=

    2019

  19. [20]

    arXiv preprint arXiv:2104.08722 , year=

    ClipBERT: What is it like to do things with words? , author=. arXiv preprint arXiv:2104.08722 , year=

    Pith/arXiv arXiv

  20. [21]

    arXiv preprint arXiv:2104.12478 , year=

    VIOLET: Vision-Language Transformer for Video Scene Understanding , author=. arXiv preprint arXiv:2104.12478 , year=

    Pith/arXiv arXiv

  21. [22]

    arXiv preprint arXiv:2111.09703 , year=

    CLIP4Clip: An Empirical Study of CLIP for End to End Video Clip Retrieval , author=. arXiv preprint arXiv:2111.09703 , year=

    Pith/arXiv arXiv

  22. [23]

    ECCV , pages=

    LocVTP: Localization-aware Video–Text Pre‑training , author=. ECCV , pages=. 2022 , url=

    2022

  23. [24]

    CVPR , pages=

    VT-TWINS: Video-Text Pretraining by Weakly Supervised Temporal Alignment , author=. CVPR , pages=. 2022 , url=

    2022

  24. [25]

    CVPR , year=

    Temporal Alignment Networks for Instructional Video Understanding , author=. CVPR , year=

  25. [26]

    NeurIPS , volume=

    Mean Teachers are Better Role Models: Weight-averaged Consistency Targets Improve Semi-Supervised Deep Learning Results , author=. NeurIPS , volume=. 2017 , url=

    2017

  26. [27]

    ACL , year=

    Teacher–Student Pseudo-labeling for Video-Language Retrieval , author=. ACL , year=

  27. [28]

    CVPR , year=

    Generative Pseudo-Labeling for Vision-Language Matching , author=. CVPR , year=

  28. [29]

    CVPR , year=

    End-to-End Learning of Visual Representations from Uncurated Instructional Videos , author=. CVPR , year=

  29. [30]

    ACM MM , year=

    Few-Shot Video Moment Retrieval with Prototype Learning , author=. ACM MM , year=

  30. [31]

    ECCV , year=

    Few-Shot Temporal Sentence Grounding via Meta-Learning , author=. ECCV , year=

  31. [32]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) , year=

    TALL: Temporal Activity Localization via Language Query , author=. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , year=

  32. [33]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) , year=

    Localizing Moments in Video with Natural Language , author=. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , year=

  33. [34]

    Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

    Temporally Grounding Natural Sentence in Video , author=. Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=. doi:10.18653/v1/D18-1015 , url=

    work page doi:10.18653/v1/d18-1015

  34. [35]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Learning 2d temporal adjacent networks for moment localization with natural language , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  35. [36]

    arXiv preprint arXiv:2005.00200 , year=

    Hero: Hierarchical encoder for video+ language omni-representation pre-training , author=. arXiv preprint arXiv:2005.00200 , year=

    Pith/arXiv arXiv 2005

  36. [37]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Howto100m: Learning a text-video embedding by watching hundred million narrated video clips , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  37. [38]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Less is more: Clipbert for video-and-language learning via sparse sampling , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  38. [39]

    2020 International joint conference on neural networks (IJCNN) , pages=

    Pseudo-labeling and confirmation bias in deep semi-supervised learning , author=. 2020 International joint conference on neural networks (IJCNN) , pages=. 2020 , organization=

    2020

  39. [40]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    CoMatch: Semi-supervised Learning with Contrastive Graph Regularization , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , pages=. 2021 , url=

    2021

  40. [41]

    arXiv preprint arXiv:2106.00609 , year=

    Robust Mutual Learning for Semi-supervised Semantic Segmentation , author=. arXiv preprint arXiv:2106.00609 , year=

    Pith/arXiv arXiv

  41. [42]

    arXiv preprint arXiv:2002.06353 , year=

    Univl: A unified video and language pre-training model for multimodal understanding and generation , author=. arXiv preprint arXiv:2002.06353 , year=

    Pith/arXiv arXiv 2002

  42. [43]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Temporal alignment networks for long-term video , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  43. [44]

    arXiv preprint arXiv:2104.08860 , year=

    Clip4clip: An empirical study of clip for end to end video clip retrieval , author=. arXiv preprint arXiv:2104.08860 , year=

    Pith/arXiv arXiv

  44. [45]

    AAAI , year=

    Rethinking weakly supervised temporal sentence grounding with reference , author=. AAAI , year=

  45. [46]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Tree-structured policy based progressive reinforcement learning for temporally language grounded video segment retrieval , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  46. [47]

    Proceedings of the 2018 ACM Multimedia Conference on Multimedia Conference , year=

    Attentive moment retrieval in videos , author=. Proceedings of the 2018 ACM Multimedia Conference on Multimedia Conference , year=

    2018

  47. [48]

    CVPR , year=

    Dense regression network for video grounding , author=. CVPR , year=

  48. [49]

    Transactions of the Association for Computational Linguistics , year=

    Grounding action descriptions in videos , author=. Transactions of the Association for Computational Linguistics , year=

  49. [50]

    ECCV , year=

    Hollywood in homes: Crowdsourcing data collection for activity understanding , author=. ECCV , year=

  50. [51]

    CVPR , year=

    Graph structured network for image-text matching , author=. CVPR , year=

  51. [52]

    NeurIPS , year=

    Unsupervised data augmentation for consistency training , author=. NeurIPS , year=

  52. [53]

    ICLR , year=

    Remixmatch: Semi-supervised learning with distribution matching and augmentation anchoring , author=. ICLR , year=

  53. [54]

    CVPR , year=

    Guided collaborative training for pixel-wise semi-supervised learning , author=. CVPR , year=

  54. [55]

    NeurIPS , year=

    Self-paced contrastive learning with hybrid memory for domain adaptive object re-id , author=. NeurIPS , year=

  55. [56]

    2023 , note=

    Video-LLaVA: Learning United Visual Representation for Large Language Models , author=. 2023 , note=

    2023

  56. [57]

    ICCV , year=

    Hierarchical Annotation for Procedural Video Understanding , author=. ICCV , year=

  57. [58]

    ICCV , year=

    Video-ChatGPT: Towards Detailed Video Understanding via Large Vision-Language Models , author=. ICCV , year=

  58. [59]

    CVPR , year=

    Quo vadis, action recognition? A new model and the kinetics dataset , author=. CVPR , year=

  59. [60]

    ICCV , year=

    Slowfast networks for video recognition , author=. ICCV , year=

  60. [61]

    ICML , year=

    Is Space-Time Attention All You Need for Video Understanding? , author=. ICML , year=

  61. [62]

    ICCV , year=

    BLIVA: Bridging Language and Vision with Attribution , author=. ICCV , year=

  62. [63]

    arXiv preprint arXiv:2311.17040 , year=

    InstructVid: Towards General Video Understanding via Multimodal Instruction Tuning , author=. arXiv preprint arXiv:2311.17040 , year=

    Pith/arXiv arXiv

  63. [64]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Towards automatic learning of procedures from web instructional videos , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  64. [65]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Coin: A large-scale dataset for comprehensive instructional video analysis , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  65. [66]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Video-text representation learning via differentiable weak temporal alignment , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  66. [67]

    Proceedings of the AAAI Conference on Artificial Intelligence , number=

    To find where you talk: Temporal sentence localization in video with attention based location regression , author=. Proceedings of the AAAI Conference on Artificial Intelligence , number=

  67. [68]

    International Conference on Computer Vision (ICCV) , year=

    Dense-Captioning Events in Videos , author=. International Conference on Computer Vision (ICCV) , year=

  68. [69]

    Advances in Neural Information Processing Systems (NeurIPS) , pages=

    Attention Is All You Need , author=. Advances in Neural Information Processing Systems (NeurIPS) , pages=

  69. [70]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Video-Text Prompting for Weakly Supervised Spatio-Temporal Video Grounding , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  70. [71]

    European Conference on Computer Vision , pages=

    Training-free video temporal grounding using large-scale pre-trained models , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  71. [72]

    arXiv preprint arXiv:2410.05643 , year=

    Trace: Temporal grounding video llm via causal event modeling , author=. arXiv preprint arXiv:2410.05643 , year=

    Pith/arXiv arXiv

  72. [73]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages=

    Let me finish my sentence: Video temporal grounding with holistic text understanding , author=. Proceedings of the 32nd ACM International Conference on Multimedia , pages=

  73. [74]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    Video-Text Prompting for Weakly Supervised Spatio-Temporal Video Grounding , author =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    2024

  74. [75]

    European Conference on Computer Vision (ECCV) Workshops , year =

    Training-Free Video Temporal Grounding using Large-scale Pre-trained Models , author =. European Conference on Computer Vision (ECCV) Workshops , year =

  75. [76]

    Proceedings of the 2024 International Conference on Multimedia Retrieval (ICMR) , pages =

    A Parallel Transformer Framework for Video Moment Retrieval , author =. Proceedings of the 2024 International Conference on Multimedia Retrieval (ICMR) , pages =. 2024 , doi =

    2024

  76. [77]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    On the Consistency of Video Large Language Models in Temporal Comprehension , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  77. [78]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    DeCafNet: Delegate and Conquer for Efficient Temporal Grounding in Long Videos , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =