Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Group Relative Augmentation for Data Efficient Action Detection

T0 review · 3 major / 5 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 Frozen InternVideo2, adapted with LoRA and learned FiLM augmentations plus group-relative weighting, reaches 25.56% mAP on AVA and 24.6% on MOMA from ~15 examples per class, outperforming standard LoRA augmentation with one-third the…

desk verdict A plausible few-shot VLM-adaptation recipe, but the paper never shows how person boxes enter the pipeline, so the reported detection mAP is unverified; the group-weighting trick also measures zero in the key ablation. read the letter →

arxiv 2507.21353 v1 pith:5GOUI3VL submitted 2025-07-28 cs.CV cs.LG

Deep Anil Patel , Iain Melvin , Zachary Izzo , Martin Renqiang Min This is my paper
classification cs.CVcs.LG
keywords spatio-temporalactiondetectionvideo-languagemodelsfew-shotadaptationlow-rank(LoRA)feature-wiselinearmodulation(FiLM)internalfeatureaugmentationgroup-weightedlossdataefficiency
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 asks whether a large frozen video-language model can be steered toward fine-grained, multi-person action detection using only about 15 labeled examples per action class, without retraining the backbone. It argues yes: adapting InternVideo2's vision encoder with low-rank adapters (LoRA) and a small learnable module that generates N internal feature variations through feature-wise linear modulation (FiLM), then balancing those variations with a group-relative weighting loss, yields 25.56% mAP on AVA and 24.6% on MOMA. The method outperforms a LoRA-plus-standard-input-augmentation baseline and roughly matches a recent motion-coherent video augmentation method while using about a third of the trainable parameters. If true, this shows that feature-level augmentation inside a frozen VLM is a more data-efficient regularizer than input-level transformations for person-centric video understanding.

What carries the argument

The load-bearing machinery is the internal FiLM augmentation module: a trainable embedding table in $R^{{N x D_e}}$ feeds a small MLP that outputs per-channel scale and shift parameters for each of N augmentations; these modulate the feature map at a chosen intermediate layer of the frozen encoder, and both the original and augmented features pass through the remaining LoRA-adapted transformer blocks. A group-relative weighting term then computes, per sample, the binary cross-entropy distance of each augmented logit from the detached anchor logit, z-scores those distances across the N augmentations, and assigns a Gaussian weight so that near-average augmentations dominate the supervised loss. LoRA on the query and value projections from layer 30 onward supplies parameter-efficient adaptation, and the training objective combines anchor BCE, distillation to the anchor, weighted augmented BCE, and an entropy term.

What would settle it

Inspect the released code or reproduce the pipeline: if person bounding boxes are never localized or aligned to the features before computing the B x K logits, the AVA and MOMA mAP values are video-level action recognition scores rather than spatio-temporal detection results, and the comparison to prior box-based action detection methods is not supported.

Watch

Extended reading notes

Core claim

The central claim is that a data-efficient adaptation framework—frozen InternVideo2 backbone, LoRA applied only to the latter half of the vision encoder, a learnable FiLM-based augmentation layer inserted after an intermediate transformer block, and a group-weighted loss that up-weights augmentations whose prediction distance from the anchor is near the group mean—produces stronger multi-label action detection from roughly 15 examples per class than LoRA with standard input augmentation. On AVA v2.2 the authors report 25.56% mAP versus 24.30% for the baseline, and on MOMA they report 24.6% versus 23.7%, with 1.27M trainable parameters against 3.60M for the baseline. The paper attributes the gain to task-relevant diversity generated in feature space plus a weighting mechanism that filters out statistically atypical augmentations.

Load-bearing premise

The reported mAP numbers rest on the assumption that the evaluation actually measures spatio-temporal action detection—that person bounding boxes are extracted and condition the model's output—even though the method section describes only whole-video features producing per-class logits with no box-matching step.

Editorial extensions

If this is right

  • On AVA v2.2, the full method reaches 25.56% mAP versus 24.30% for LoRA with standard augmentation, using 1.27M trainable parameters instead of 3.60M.
  • On MOMA, it reaches 24.6% mAP versus 23.7% for the baseline and 24.3% for MCA, again with far fewer trainable parameters.
  • Ablations show each loss component contributes: distillation alone adds +0.93% and entropy alone +1.06% over the baseline on AVA, and the combination with group weighting gives the full +1.26%.
  • The optimal number of learned augmentations is dataset-dependent—N=8 for MOMA, N=10 for AVA—and too many or too few augmentations degrade mAP.

Reading between the lines

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

  • Because the learned augmentations are computed from the support set's own features, the same machinery could plausibly transfer to other frozen video encoders and to tasks like dense video captioning, though the paper only tests InternVideo2.
  • The group-relative weighting computes distances within the current batch; using a larger memory queue of augmentations could stabilize the z-score statistics in small-batch few-shot training, a testable extension.
  • If the evaluation protocol is confirmed to be box-free video-level classification, the method would still support a weaker claim about clip-level action recognition from few examples, not spatio-temporal action detection.
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

3 major / 5 minor

Summary. The paper proposes a few-shot action detection adaptation method for frozen video-language models. It combines LoRA applied to the latter half of the vision encoder with a learnable FiLM-based feature augmentation module that produces N augmented versions of intermediate features, plus a training objective with four terms: BCE on original predictions, distillation from detached original predictions, group-weighted BCE on augmented predictions, and entropy maximization. Experiments on AVA and MOMA with approximately 15 examples per class report mAP of 25.56% and 24.6%, claiming improvements over a LoRA-plus-standard-augmentation baseline and parameter efficiency relative to MCA. The paper includes ablations on the number of augmentations and on loss components, and it closes with a limitations section and an appendix describing the support-set sampling procedure.

Significance. If the reported results are verified, the paper would offer a practical parameter-efficient recipe for adapting video-language models to person-centric action detection from very few labels, with explicit loss definitions and a clear high-level design. The manuscript is generally well organized, the equations are concrete, and the authors include ablation studies and acknowledge limitations. However, the empirical claim currently rests on an unspecified action-detection evaluation protocol, and the direct evidence for the group-weighting contribution is weak: the group-weighted loss shows no improvement over the unweighted augmented loss in the ablation, and all numbers come from single runs without variance. The missing evaluation protocol is the most serious issue because it determines whether the reported numbers are action-detection results or clip-level action-recognition results.

major comments (3)
  1. [§3.2–§3.3, §4.1] The manuscript never specifies how the model produces spatio-temporal action detection predictions. Equations (3)–(8) and the surrounding text define z_orig ∈ R^{B×K} as logits from whole-batch video features f^(laug), with no person-proposal step, RoI alignment, box-conditioned feature pooling, or box-matching mechanism. Yet §4.1 reports 'standard mAP' on AVA v2.2 and MOMA, whose official protocols require per-person bounding-box predictions at keyframes (AVA mAP is frame-level AP over ground-truth person boxes). As written, the architecture can only produce clip-level multi-label logits and has no mechanism to assign actions to individuals. Please state exactly how person boxes are incorporated, how K is mapped to the 80/52 action classes, how multiple people per frame are handled, and how the reported mAP is computed. Without this, the numbers in Tables 1 and 2 are not interpretable as action detection and the comparison to prior detection methods is invalid.
  2. [§4.3, Table 4] The group-weighted loss, which is listed as a main contribution in the abstract and in §1, shows no direct effect in the ablations: adding LBCE-W yields 24.96% mAP versus 24.97% for LBCE without group weights. The only comparison that numerically favors group weighting is the full system (25.56% versus 25.26%), which is a single-run difference and could be noise. This does not substantiate the claim that group weighting 'promotes robust learning by prioritizing informative yet reasonable augmentations.' Please provide multi-seed results with variance and an analysis of when the learned weights differ from uniform, or substantially demote this component from a main contribution.
  3. [Tables 1–4 and Appendix §6] All reported results are single-run and no error bars or significance measures are given, despite the central claims relying on small margins (e.g., +0.9% on MOMA, +0.3% versus MCA). The appendix also states that K=15 is approximate and that exact counts 'may vary slightly' depending on label co-occurrence, but no actual counts per class or per split are reported. Given the few-shot setting and the data-efficiency claim, please report multiple seeds, exact support-set sizes, and the full evaluation settings (frame sampling, resolution, input preprocessing, and any proposal detector used at test time) so the reader can assess whether the observed differences are meaningful.
minor comments (5)
  1. [§4.3, Table 4] Clarify whether the rows labeled 'Ldistill only' and 'Lent only' are additions to the anchor LBCE loss or replacements; the current presentation is ambiguous.
  2. [Eq. (1)] State explicitly that the 2C-dimensional output of gψ is split into γ_i and β_i, and specify the broadcasting behavior in Eq. (2) over the batch and sequence dimensions.
  3. [§3.4, Eq. (13)] The weighting is described as a Gaussian 'centered at the mean distance,' but the argument is the z-score (di,b − µd,b)/σd,b; please clarify the intended behavior when σd,b is very small or zero.
  4. [§4.2] The phrase 'strong mAP performance' is not calibrated against published full-data results on AVA or MOMA; adding a few state-of-the-art reference numbers would help the reader understand the few-shot setting.
  5. [References] References [11] and [12] appear to describe the same work with slightly different titles and venues; please verify and deduplicate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical method study and no reported result reduces by construction to a fitted or self-cited input.

full rationale

The paper makes no derivation-from-principles claim and contains no equation that reduces a purported prediction to a fitted value. The group weights in Eq. 13 are computed from the model's own detached predictions, but this is an internal loss-reweighting mechanism for training (Eq. 7), not a way of producing the reported mAP numbers; the reported mAP is evaluated against held-out ground truth. The distillation loss in Eq. 5 uses the original logits as soft targets for augmented logits, but this is a regularizer and does not define the final evaluation metric. The only author-overlapping citation is Ref. [2] (Bao et al., WACV 2025), which includes two present authors, but it is cited merely as one of several prior efforts on zero-shot action detection and is not load-bearing for the architecture, losses, or evaluation. The paper's own limitation section concedes that hyper-parameters are dataset-dependent, which is a tuning concern, not circularity. The largest validity threat, that Section 3.3 defines B×K logits for whole-batch video features without an explicit person-proposal or RoI alignment step while Section 4.1 reports spatio-temporal mAP on AVA and MOMA, is a possible evaluation-protocol mismatch and a correctness risk; it is not a circular inference because the reported numbers are not derived by construction from the method's equations.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the unstated equivalence between video-level feature classification and person-level action detection, plus a set of manually chosen hyperparameters. No new physical entity or external benchmark is introduced, so the ledger is dominated by modeling assumptions and tunable constants.

free parameters (7)
  • N (number of augmentations) = 8 on MOMA, 10 on AVA
    Chosen by scanning N on MOMA (Table 3); AVA uses N=10 without an equivalent ablation shown.
  • s (Gaussian weighting sensitivity) = not reported
    Controls how quickly augmentation weights decay with z-score in Eq. 13; no value is given.
  • λdistill, λaug, λent = not reported
    Loss weights in Eq. 3 are set by hand and their values are never stated.
  • LoRA rank r = not reported
    Rank of the low-rank adapters in Sec. 3.1 is not specified.
  • laug (augmentation insertion layer) = not reported
    The FiLM module is inserted after an intermediate layer, but the actual layer index is not given.
  • llora (LoRA start layer) = layer 30 of 40
    LoRA is applied from layer 30 onward; this choice is stated but not ablated.
  • K (support examples per class) = 15 (approximate)
    Sampling target for the few-shot support set; exact counts vary with label co-occurrence per Sec. 6.
assumptions (3)
  • domain assumption Frozen InternVideo2 visual features carry enough person-centric signal that LoRA plus FiLM on later layers can bridge to action detection.
    Used throughout Sec. 3; never verified against alternatives such as full fine-tuning or other backbones.
  • domain assumption Predictions on global video-token features are a valid proxy for per-person action detection.
    The method outputs B×K logits without any described box-proposal or RoI extraction step (Sec. 3.3), but evaluates mAP on the spatio-temporal detection datasets AVA and MOMA.
  • ad hoc to paper BCE distance to the detached original prediction is a reliable measure of augmentation usefulness.
    The Gaussian group weighting in Eqs. 9-13 assumes that augmentations whose divergence is near the group mean are reasonable; this is a modeling choice with no independent evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Group Relative Augmentation for Data Efficient Action Detection." pith.science (2026). https://pith.science/paper/5GOUI3VL

@misc{pith2026250721353,
  author       = {Pith},
  title        = {Pith review of: Group Relative Augmentation for Data Efficient Action Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5GOUI3VL}},
  note         = {Machine review of arXiv:2507.21353}
}
read the original abstract

Adapting large Video-Language Models (VLMs) for action detection using only a few examples poses challenges like overfitting and the granularity mismatch between scene-level pre-training and required person-centric understanding. We propose an efficient adaptation strategy combining parameter-efficient tuning (LoRA) with a novel learnable internal feature augmentation. Applied within the frozen VLM backbone using FiLM, these augmentations generate diverse feature variations directly relevant to the task. Additionally, we introduce a group-weighted loss function that dynamically modulates the training contribution of each augmented sample based on its prediction divergence relative to the group average. This promotes robust learning by prioritizing informative yet reasonable augmentations. We demonstrate our method's effectiveness on complex multi-label, multi-person action detection datasets (AVA, MOMA), achieving strong mAP performance and showcasing significant data efficiency for adapting VLMs from limited examples.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 29 canonical work pages

  1. [1]

    Frozen in time: A joint video and image encoder for end-to-end retrieval

    Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728–1738, 2021

    work page 2021

  2. [2]

    Exploiting vlm localizability and semantics for open vocabulary action detection

    Wentao Bao, Kai Li, Yuxiao Chen, Deep Patel, Martin Renqiang Min, and Yu Kong. Exploiting vlm localizability and semantics for open vocabulary action detection. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages 8291–8301. IEEE, 2025

    work page 2025

  3. [3]

    Frozen feature augmentation for few-shot image classification

    Andreas Bär, Neil Houlsby, Mostafa Dehghani, and Manoj Kumar. Frozen feature augmentation for few-shot image classification. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 16046–16057, 2024

    work page 2024

  4. [4]

    Visualgpt: Data- efficient adaptation of pretrained language models for image captioning

    Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. Visualgpt: Data- efficient adaptation of pretrained language models for image captioning. In Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition , pages 18030–18040, 2022

    work page 2022

  5. [5]

    Adversarial Feature Augmentation and Normalization for Visual Recognition

    Tianlong Chen, Yu Cheng, Zhe Gan, Jianfeng Wang, Lijuan Wang, Zhangyang Wang, and Jingjing Liu. Adversarial feature augmentation and normalization for visual recog- nition. arXiv preprint arXiv:2103.12171, 2021

    work page Pith review arXiv 2021

  6. [6]

    Perceptionlm: Open-access data and models for detailed visual understanding

    Jang Hyun Cho, Andrea Madotto, Effrosyni Mavroudi, Triantafyllos Afouras, Tushar Nagarajan, Muhammad Maaz, Yale Song, Tengyu Ma, Shuming Hu, Suyog Jain, et al. Perceptionlm: Open-access data and models for detailed visual understanding. arXiv preprint arXiv:2504.13180, 2025

    arXiv 2025

  7. [7]

    Autoaugment: Learning augmentation strategies from data

    Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 113–123, 2019

    work page 2019

  8. [8]

    Clip-adapter: Better vision-language models with fea- ture adapters

    Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with fea- ture adapters. International Journal of Computer Vision, 132(2):581–595, 2024

    work page 2024

Show all 41 references
  1. [9]

    Ava: A video dataset of spatio-temporally localized atomic visual actions

    Chunhui Gu, Chen Sun, David A Ross, Carl V ondrick, Caroline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Proceedings of the IEEE confer...

    2018

  2. [10]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022

    2022

  3. [11]

    Interaction-aware prompting for zero-shot spatio-temporal action detection

    Wei-Jhe Huang, Jheng-Hsien Yeh, Min-Hung Chen, Gueter Josmy Faure, and Shang- Hong Lai. Interaction-aware prompting for zero-shot spatio-temporal action detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 284–293, 2023. PA TEL, ET AL.: ...

    2023

  4. [12]

    Interaction-aware prompting for zero-shot spatio-temporal action detection

    Wei-Jhe Huang, Jheng-Hsien Yeh, Min-Hung Chen, Gueter Josmy Faure, and Shang- Hong Lai. Interaction-aware prompting for zero-shot spatio-temporal action detection. In ICCV Workshop, pages 284–293, 2023

    2023

  5. [13]

    Spatio-temporal context prompting for zero-shot action detection

    Wei-Jhe Huang, Min-Hung Chen, and Shang-Hong Lai. Spatio-temporal context prompting for zero-shot action detection. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 9083–9092. IEEE, 2025

    2025

  6. [14]

    Scaling up visual and vision-language representation learning with noisy text supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904–4916....

    2021

  7. [15]

    Region-aware pretraining for open- vocabulary object detection with vision transformers

    Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open- vocabulary object detection with vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11144–11154, 2023

    2023

  8. [16]

    On feature normalization and data augmentation

    Boyi Li, Felix Wu, Ser-Nam Lim, Serge Belongie, and Kilian Q Weinberger. On feature normalization and data augmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12383–12392, 2021

    2021

  9. [17]

    Blip: Bootstrapping language- image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language- image pre-training for unified vision-language understanding and generation. In Inter- national conference on machine learning, pages 12888–12900. PMLR, 2022

    2022

  10. [18]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  11. [19]

    Learning object-language alignments for open-vocabulary object detection

    Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. arXiv preprint arXiv:2211.14843, 2022

    2022 arXiv

  12. [20]

    Univl: A unified video and language pre-training model for multimodal understanding and generation

    Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. Univl: A unified video and language pre-training model for multimodal understanding and generation. arXiv preprint arXiv:2002.06353, 2020

    2002 arXiv

  13. [21]

    Moma: Multi-object multi-actor activity parsing

    Zelun Luo, Wanze Xie, Siddharth Kapoor, Yiyun Liang, Michael Cooper, Juan Carlos Niebles, Ehsan Adeli, and Fei-Fei Li. Moma: Multi-object multi-actor activity parsing. Advances in neural information processing systems, 34:17939–17955, 2021

    2021

  14. [22]

    Film: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  15. [23]

    Learning transferable visual models from natural language supervision

    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 supervision. In International confer- ence on machine learning, ...

    2021

  16. [24]

    Videomae: Masked autoen- coders are data-efficient learners for self-supervised video pre-training

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoen- coders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35:10078–10093, 2022. 12 PA TEL, ET AL.: GROUP RELA TIVE AUGMENTA TION

    2022

  17. [25]

    Internvideo2: Scaling foundation models for multimodal video understanding

    Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Zun Wang, Yansong Shi, et al. Internvideo2: Scaling foundation models for multimodal video understanding. In European Conference on Computer Vision, pages 396–416. Springer, 2024

    2024

  18. [26]

    Feature adaptation with clip for few-shot classification

    Guangxing Wu, Junxi Chen, Wentao Zhang, and Ruixuan Wang. Feature adaptation with clip for few-shot classification. In Proceedings of the 5th ACM International Conference on Multimedia in Asia, pages 1–7, 2023

    2023

  19. [27]

    Cora: Adapting clip for open- vocabulary detection with region prompting and anchor pre-matching

    Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open- vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7031– 7040, 2023

    2023

  20. [28]

    Videoclip: Contrastive pre-training for zero-shot video-text understanding

    Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Flo- rian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021

    2021 arXiv

  21. [29]

    Videococa: Video-text modeling with zero-shot transfer from contrastive captioners

    Shen Yan, Tao Zhu, Zirui Wang, Yuan Cao, Mi Zhang, Soham Ghosh, Yonghui Wu, and Jiahui Yu. Videococa: Video-text modeling with zero-shot transfer from contrastive captioners. arXiv preprint arXiv:2212.04979, 2022

    2022 arXiv

  22. [30]

    Vid2seq: Large-scale pretraining of a visual language model for dense video captioning

    Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision ...

    2023

  23. [31]

    Image data augmentation for deep learning: A survey

    Suorong Yang, Weikang Xiao, Mengchen Zhang, Suhan Guo, Jian Zhao, and Fu- rao Shen. Image data augmentation for deep learning: A survey. arXiv preprint arXiv:2204.08610, 2022

    2022 arXiv

  24. [32]

    Textmania: Enriching visual feature by text-driven manifold augmentation

    Moon Ye-Bin, Jisoo Kim, Hongyeob Kim, Kilho Son, and Tae-Hyun Oh. Textmania: Enriching visual feature by text-driven manifold augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2526–2537, 2023

    2023

  25. [33]

    Coca: Contrastive captioners are image-text foundation models

    Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022

    2022 arXiv

  26. [34]

    Cutmix: Regularization strategy to train strong classifiers with local- izable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with local- izable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023–6032, 2019

    2019

  27. [35]

    Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation

    Yuhang Zang, Chen Huang, and Chen Change Loy. Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3457–3466, 2021

    2021

  28. [36]

    Open- vocabulary object detection using captions

    Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open- vocabulary object detection using captions. In Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, pages 14393–14402, 2021. PA TEL, ET AL.: GROUP RELA TIVE AUGMENTA TION 13

    2021

  29. [37]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international con- ference on computer vision, pages 11975–11986, 2023

    2023

  30. [38]

    Tip-adapter: Training-free clip-adapter for better vision- language modeling

    Renrui Zhang, Rongyao Fang, Wei Zhang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision- language modeling. arXiv preprint arXiv:2111.03930, 2021

    2021 arXiv

  31. [39]

    Don’t judge by the look: Towards motion coherent video representation.arXiv preprint arXiv:2403.09506, 2024

    Yitian Zhang, Yue Bai, Huan Wang, Yizhou Wang, and Yun Fu. Don’t judge by the look: Towards motion coherent video representation.arXiv preprint arXiv:2403.09506, 2024

    2024 arXiv

  32. [40]

    Regionclip: Region- based language-image pretraining

    Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liu- nian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region- based language-image pretraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognitio...

    2022

  33. [41]

    Not all features matter: Enhancing few-shot clip with adaptive prior refine- ment

    Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refine- ment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2605–2615, 2023. 6 Multi L...

    2023

Pith tools

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