Pith. sign in

REVIEW 5 major objections 4 minor 31 references

EdgeVidSum: Real-Time Personalized Video Summarization at the Edge

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read EdgeVidSum claims that thumbnail-container analysis lets a resource-limited edge device generate personalized fast-forward summaries of 90-minute videos in seconds, with all preference data staying on-device.

desk verdict A plausible demo whose central claims are unmeasured and whose headline numbers don't add up; not ready for research peer review. read the letter →

arxiv 2506.03171 v1 pith:7YYZ64JP submitted 2025-05-28 cs.CV cs.AI

Ghulam Mujtaba , Eun-Seok Ryu This is my paper
classification cs.CVcs.AI
keywords videosummarizationfast-forwardedgecomputingthumbnailcontainerspersonalizationtripletattentionJetsonNanolightweightCNN
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 tries to show that personalized fast-forward video summarization can be moved from cloud servers to low-power edge devices without losing the ability to find what a viewer wants. Its central move is to stop analyzing video frames and instead analyze thumbnail containers—tiny 160x90 images sampled throughout the video—with a lightweight 2D CNN, then narrow in on preferred moments by inspecting thumbnails at increasingly fine temporal granularity. On a Jetson Nano the authors report generating a summary of a 90-minute video in 2 to 5 seconds, with a 97.8% reduction in data processed compared with frame-by-frame methods. If this holds, viewers get personalized, temporally continuous recaps on their own devices, with preference data never leaving the device.

What carries the argument

The load-bearing object is the thumbnail container: a 160x90 pixel, roughly 4 KB image sampled at intervals throughout a video (6,734 containers per 112-minute football video versus about 202,036 frames), with intelligent selection algorithms claimed to preserve semantic integrity. This object shrinks the input to the neural network by orders of magnitude, making on-device inference feasible. The second mechanism is hierarchical temporal narrowing: the client first scores coarse thumbnails for preferred classes, then re-examines thumbnail containers inside candidate intervals at finer granularity until it can emit precise timestamps. The third is the Triplet Attention Module, three parallel branches that pool along height, width, and channel dimensions and reweight the feature maps; the authors credit it with a 4.2% accuracy gain over channel-only attention for 0.3M extra parameters.

What would settle it

Take a video with a known user-preferred event that is shorter than the thumbnail sampling interval or visible only at full resolution, run EdgeVidSum's thumbnail pipeline on a Jetson Nano, and check whether the generated summary includes the event; a miss would show that the semantic-integrity assumption fails at the claimed sampling density.

Watch

Extended reading notes

Core claim

EdgeVidSum's central claim is that thumbnail containers can replace full-frame analysis as the substrate for personalized video summarization without sacrificing semantic relevance. A classifier built on an EfficientNetV2-S backbone followed by a Triplet Attention Module predicts user-preferred content categories from 160x90 thumbnails; a hierarchical timestamp-generation step then refines coarse detections into segment-level boundaries; a client-side player plays preferred segments at normal speed and accelerates the rest. The authors state that this pipeline produces fast-forward summaries of 90-minute videos in 2 to 5 seconds on a Jetson Nano, processes 97.8% less data than frame-by-frame approaches, and keeps all preference inference local.

Load-bearing premise

The load-bearing premise is that a sparse set of 160x90 thumbnails, sampled about once per second, preserves every semantically relevant moment; the paper asserts this via thumbnail selection algorithms that maintain semantic integrity but does not test it.

Editorial extensions

If this is right

  • On a Jetson Nano-class device, personalized summaries of 90-minute videos can be generated in 2 to 5 seconds, making interactive on-device summarization feasible.
  • User preferences and viewing history stay local, since thumbnail classification and timestamp generation run on the edge; only thumbnails and requested video segments cross the network.
  • The 97.8% reduction in processed data means bandwidth and memory requirements drop by roughly two orders of magnitude relative to frame-by-frame analysis.
  • Hierarchical narrowing lets the system convert coarse preferred-content detections into precise playback timestamps, so the final video is a continuous fast-forward rather than a discontinuous keyframe montage.
  • Variable-speed playback preserves narrative continuity, addressing the fragmentation problem of keyframe and keyshot summaries.

Reading between the lines

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

  • The thumbnail-container trick is general: any frame-based video-understanding task that can be posed as image classification could inherit the same speedup, at the price of the same sampling-resolution ceiling.
  • The reported 2-5 second processing time likely assumes thumbnails are already generated and stored on the video service provider; end-to-end latency would also include thumbnail transfer and segment requests, so the real-time claim concerns computation, not network round-trips.
  • If sampling is roughly one thumbnail per second, events shorter than that interval are invisible; an adaptive sampling rate tied to motion or audio salience would test and likely improve the semantic-preservation assumption.
  • The personalization ceiling is the classifier's label set; moving from categorical preferences to free-form natural-language queries would need a larger model and would likely exceed the Jetson Nano's budget.
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 / 4 minor

Summary. This manuscript proposes EdgeVidSum, a client-driven personalized fast-forward video summarization framework for edge devices. The system assumes that a video service provider supplies pre-generated 160x90 thumbnail containers; a lightweight EfficientNetV2 backbone with a Triplet Attention Module classifies thumbnails according to user preferences, and timestamps are generated by narrowing temporal periods at increasing granularity. The authors report real-time operation on a Jetson Nano, 2-5 second processing for 90-minute videos, a 97.8% reduction in data processing volume, a 98.2% memory reduction, and a 4.2% accuracy improvement from the attention module. The evaluation consists of an interactive demo; no datasets, baselines, benchmarks, ablations, or reproducibility details are provided.

Significance. If substantiated, the system would be practically relevant for on-device, privacy-preserving personalized video summarization, and the use of thumbnail containers is a sensible way to cut computational cost. The manuscript also references a broad related-work landscape and builds on established components, namely EfficientNetV2 and the Triplet Attention Module. However, the central claims are not currently supportable: the semantic-relevance assertion is never measured, the headline efficiency numbers are internally inconsistent, the only accuracy number lacks any experimental protocol, and the core sampling assumption is untested. The contribution is thus best described as a system sketch with a demo, and the quantitative claims cannot be verified from the paper.

major comments (5)
  1. [Abstract and Section 4] The claim that thumbnail containers reduce computational complexity 'without sacrificing semantic relevance' is never evaluated. Section 4 reports only an interactive demo and states processing times of 2 to 5 seconds for 90-minute videos without describing the measurement procedure, hardware conditions, error bars, or content complexity metric. No comparison is made to standard summarization benchmarks such as SumMe or TVSum, or to fast-forward baselines such as FFNet and SpeedNet, so the paper's main promise remains unverified.
  2. [Sections 3.1 and 3.3] The method assumes that roughly one 160x90 thumbnail per 1.2 seconds (202,036 frames versus 6,734 thumbnails) preserves every user-relevant event and permits segment-level timestamp precision. Section 3.3, step 3, says the client narrows temporal periods by analyzing thumbnail containers at increasing temporal granularity, but no finer-grained thumbnail layer is described or evaluated; if an event occurs between adjacent thumbnails or is unrecognizable at that resolution, the entire pipeline cannot recover it. The 'semantic integrity' claim is attributed to reference [11] and is not supported by any experiment, ablation, or dataset in this manuscript.
  3. [Section 3.1] The headline efficiency numbers are arithmetically inconsistent. A 30x reduction in thumbnail count corresponds to a 96.7% reduction, not the stated 97.8%; and 160x90 versus 1920x1080 corresponds to a 99.3% memory reduction, not the stated 98.2%. No derivation or measurement is given for either figure, which suggests the numbers are estimates rather than results of a controlled experiment.
  4. [Section 3.2.1] The only quantitative accuracy result, a 4.2% improvement from adding the Triplet Attention Module, is stated in a single sentence as coming from 'our experimental evaluation' with no dataset, training protocol, baseline definition, or statistical significance. Since this number is used to justify the core architecture, it is load-bearing and must be supported by a reproducible experimental section.
  5. [Section 3.3 and Figure 1] The claimed on-device efficiency depends on the video service provider having already generated thumbnail containers, a component that is not described, measured, or evaluated beyond a citation to the authors' prior work [11]. The preprocessing, storage, and transfer costs of thumbnail generation are therefore excluded from the reported savings; an end-to-end accounting should include these steps before claiming a reduction in total computational complexity.
minor comments (4)
  1. [Section 1, reference [24]] Reference [24] is cited as an example of keyshot-based summarization, but the cited paper is about automatic thumbnail selection from videos; please verify and correct the citation.
  2. [Section 3.2.1] The sentence comparing TAM to 'channel attention [6] and convolutional block attention modules [6]' cites the same reference twice; the CBAM paper should be cited correctly, and the comparison should name the datasets used.
  3. [Equations (1)-(2)] The notation for the convolution operation is inconsistent: equation (1) writes concatenated pooling, while the text says 'convolutional layer with 7x7 kernels' but equation (2) simply uses an undefined operator rho; please define all operators precisely.
  4. [Section 4] The demo description would be clearer if it stated which videos were used, how user preferences were selected, and how the generated summaries were verified to match user expectations.

Circularity Check

2 steps flagged · score 6.0 of 10

Central 'no semantic loss' claim rests on same-author citation [11]; the pipeline is prior work plus an external attention module, with no independent benchmark.

  1. self citation load bearing [Section 3.1 (Thumbnail-Based Approach), supported by Section 3.3 step 1]
    "Third, thumbnails provide higher quality representation through intelligent selection algorithms that maintain semantic integrity [11]."

    The paper's headline trade-off is 'significantly reduce computational complexity without sacrificing semantic relevance' (Abstract), yet the only evidence that thumbnails preserve semantics is this citation to [11], a prior paper by the same two authors that introduced thumbnail containers. No dataset, ground-truth comparison, or ablation in the present paper tests whether the 160x90, roughly one-thumbnail-per-second containers retain user-relevant events. Every downstream step (preference classification, timestamp narrowing, fast-forward boundaries) inherits that assumption, so the central claim reduces to trusting the authors' own previous framework rather than an independent result.

  2. ansatz smuggled in via citation [Section 3.3, System Architecture, step 1]
    "The VSP transmits pre-generated thumbnail containers representing the entire video at a very low bitrate [11]."

    The paper presents EdgeVidSum as a novel method, but the thumbnail-container representation, including the semantic-integrity assumption and the client-driven narrowing protocol, is imported from [11] (and [12]) without re-derivation or fresh validation. The only genuinely external piece, TAM [10], is an off-the-shelf module; the 'pre-generated thumbnail containers' that make the method work are taken as given by the same authors' earlier work. Thus the novelty claim is a restatement of the prior pipeline plus an attention swap, with the load-bearing ansatz supplied by self-citation.

full rationale

The paper's strongest assertion, that thumbnail containers cut computation 'without sacrificing semantic relevance', is not derived or measured in this manuscript. It is assigned to reference [11], a same-author paper that defined exactly this thumbnail-container approach, making the core premise a load-bearing self-citation rather than independent evidence. This matches the self_citation_load_bearing pattern and warrants a partial-circularity score. A second, overlapping issue is that the client-driven architecture is essentially the authors' prior LTC-SUM/thumbnail-container framework with an external Triplet Attention Module inserted; the 'novel' contribution is therefore largely a restatement of the authors' own published pipeline. I do not treat the 4.2% accuracy improvement as a circular step: it is an unsupported claim about an external module, not an equation that reduces to its inputs. Similarly, the inconsistent arithmetic (97.8% reduction vs. 30x fewer thumbnails, which is 96.7%; 98.2% memory reduction vs. 99.3% for 160x90 over 1920x1080) and the unvalidated 2-5 second processing time are correctness and consistency concerns, not circularity. Because the central 'semantic relevance' guarantee is carried entirely by a same-author citation and never independently tested, the score is 6 rather than 0-2; it is not 8-10 because the efficiency arithmetic and the external TAM component give the paper some independent content.

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

The central claim depends on three unvalidated assumptions: thumbnails preserve event-level semantic content, user preferences map to a fixed set of image-detectable classes, and the video service provider has pre-generated thumbnail containers. No free parameters are fitted in the paper; the listed items are design choices that are unablated.

free parameters (3)
  • thumbnail resolution (160x90)
    Chosen resolution; no ablation shows why 160x90 preserves semantic relevance or how it affects detection of small objects or events.
  • thumbnail sampling rate = approximately 1 per second for the football example
    The paper does not specify the exact sampling rule, yet it determines whether brief events are captured at all.
  • number of preference categories
    The classifier outputs a fixed set of activity categories, but the mapping from user preferences to these categories is unspecified and unvalidated.
assumptions (3)
  • domain assumption Thumbnails preserve the semantic content of the video sufficiently to identify all user-relevant segments.
    Section 3.1 claims thumbnail selection maintains semantic integrity via [11], but no evidence is provided in this paper.
  • domain assumption User preferences are representable as a fixed set of activity classes detectable from single images.
    Section 3.3 step 2 maps preferences to CNN categories; this is assumed, not validated.
  • ad hoc to paper The video service provider has pre-generated thumbnail containers for all videos.
    Section 3.3 step 1 assumes the VSP hosts a thumbnail container database; this is a deployment assumption, not a general capability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EdgeVidSum: Real-Time Personalized Video Summarization at the Edge." pith.science (2026). https://pith.science/paper/7YYZ64JP

@misc{pith2026250603171,
  author       = {Pith},
  title        = {Pith review of: EdgeVidSum: Real-Time Personalized Video Summarization at the Edge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7YYZ64JP}},
  note         = {Machine review of arXiv:2506.03171}
}
read the original abstract

EdgeVidSum is a lightweight method that generates personalized, fast-forward summaries of long-form videos directly on edge devices. The proposed approach enables real-time video summarization while safeguarding user privacy through local data processing using innovative thumbnail-based techniques and efficient neural architectures. Unlike conventional methods that process entire videos frame by frame, the proposed method uses thumbnail containers to significantly reduce computational complexity without sacrificing semantic relevance. The framework employs a hierarchical analysis approach, where a lightweight 2D CNN model identifies user-preferred content from thumbnails and generates timestamps to create fast-forward summaries. Our interactive demo highlights the system's ability to create tailored video summaries for long-form videos, such as movies, sports events, and TV shows, based on individual user preferences. The entire computation occurs seamlessly on resource-constrained devices like Jetson Nano, demonstrating how EdgeVidSum addresses the critical challenges of computational efficiency, personalization, and privacy in modern video consumption environments.

Figures

Figures reproduced from arXiv: 2506.03171 by the authors.

Figure 1
Figure 1. System architecture of EdgeVidSum showing the flow [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Complete architecture of our Thumbnail Container Analyzer. The RGB input image is processed by the EfficientNetV2-S [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Detailed architecture of the Triplet Attention Mod [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Screenshot of EdgeVidSum demo interface showing [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 30 canonical work pages

  1. [11]

    Client-driven person- alized trailer framework using thumbnail containers

    Ghulam Mujtaba and Eun-Seok Ryu. Client-driven person- alized trailer framework using thumbnail containers. IEEE Access, 8:60417–60427, 2020. 3, 4

    work page 2020

  2. [1]

    Security in cloud computing: Opportunities and challenges

    Mazhar Ali, Samee U Khan, and Athanasios V Vasilakos. Security in cloud computing: Opportunities and challenges. Information sciences, 305:357–383, 2015. 2

    work page 2015

  3. [2]

    Speednet: Learning the speediness in videos

    Sagie Benaim, Ariel Ephrat, Oran Lang, Inbar Mosseri, William T Freeman, Michael Rubinstein, Michal Irani, and Tali Dekel. Speednet: Learning the speediness in videos. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 9922–9931, 2020. 2

    work page 2020

  4. [3]

    A multimodal hy- perlapse method based on video and songs’ emotion align- ment

    Diognei de Matos, Washington Ramos, Michel Silva, Luiz Romanhol, and Erickson R Nascimento. A multimodal hy- perlapse method based on video and songs’ emotion align- ment. Pattern Recognition Letters, 166:174–181, 2023. 2

    work page 2023

  5. [4]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255, 2009. 3

    work page 2009

  6. [5]

    Video summarization by learning submodular mixtures of objec- tives

    Michael Gygli, Helmut Grabner, and Luc Van Gool. Video summarization by learning submodular mixtures of objec- tives. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3090–3098, 2015. 2

    work page 2015

  7. [6]

    Squeeze-and-excitation net- works

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation net- works. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 7132–7141,

  8. [7]

    Huang and H

    C. Huang and H. Wang. A novel key-frames selection frame- work for comprehensive video summarization. IEEE Trans- actions on Circuits and Systems for Video Technology, 30(2): 577–589, 2020. 1

    work page 2020

Show all 31 references
  1. [8]

    Ffnet: Video fast-forwarding via reinforcement learning

    Shuyue Lan, Rameswar Panda, Qi Zhu, and Amit K Roy- Chowdhury. Ffnet: Video fast-forwarding via reinforcement learning. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition , pages 6771–6780,

  2. [9]

    A generic framework of user attention model and its application in video summarization

    Yu-Fei Ma, Xian-Sheng Hua, Lie Lu, and Hong-Jiang Zhang. A generic framework of user attention model and its application in video summarization. IEEE transactions on multimedia, 7(5):907–919, 2005. 2

    2005

  3. [10]

    Rotate to attend: Convolutional triplet attention module

    Diganta Misra, Trikay Nalamada, Ajay Uppili Aransani- palai, and Qibin Hou. Rotate to attend: Convolutional triplet attention module. In Proceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision (WACV), pages 3139–3148, 2021. 3

    2021

  4. [12]

    Ltc- sum: Lightweight client-driven personalized video summa- rization framework using 2d cnn

    Ghulam Mujtaba, Adeel Malik, and Eun-Seok Ryu. Ltc- sum: Lightweight client-driven personalized video summa- rization framework using 2d cnn. IEEE access, 10:103041– 103055, 2022. 1

    2022

  5. [13]

    Frc-gif: Frame 5 ranking-based personalized artistic media generation method for resource constrained devices

    Ghulam Mujtaba, Sunder Ali Khowaja, Muhammad Aslam Jarwar, Jaehyuk Choi, and Eun-Seok Ryu. Frc-gif: Frame 5 ranking-based personalized artistic media generation method for resource constrained devices. IEEE Transactions on Big Data, 10(4):343–355, 2023. 2

    2023

  6. [14]

    Per- sonalized semantic fast-forward videos for next generation streaming platforms, 2025

    Ghulam Mujtaba, Eun Seok Ryu, and Reda Harb. Per- sonalized semantic fast-forward videos for next generation streaming platforms, 2025. US Patent App. 18/942,257. 1, 2

    2025

  7. [15]

    Generating personalized summaries of day long egocentric videos

    Pravin Nagar, Anuj Rathore, CV Jawahar, and Chetan Arora. Generating personalized summaries of day long egocentric videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):6832–6845, 2021. 2

    2021

  8. [16]

    Video summarization and scene detection by graph modeling.IEEE Transactions on circuits and systems for video technology , 15(2):296–305, 2005

    Chong-Wah Ngo, Yu-Fei Ma, and Hong-Jiang Zhang. Video summarization and scene detection by graph modeling.IEEE Transactions on circuits and systems for video technology , 15(2):296–305, 2005. 2

    2005

  9. [17]

    Clip-based similarity measure for query-dependent clip retrieval and video sum- marization

    Yuxin Peng and Chong-Wah Ngo. Clip-based similarity measure for query-dependent clip retrieval and video sum- marization. IEEE Transactions on Circuits and Systems for Video Technology, 16(5):612–627, 2006. 2

    2006

  10. [18]

    Enhancing video summarization via vision-language embed- ding

    Bryan A Plummer, Matthew Brown, and Svetlana Lazebnik. Enhancing video summarization via vision-language embed- ding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5781–5789, 2017. 2

    2017

  11. [19]

    Category-specific video summarization

    Danila Potapov, Matthijs Douze, Zaid Harchaoui, and Cordelia Schmid. Category-specific video summarization. In Computer Vision–ECCV 2014: 13th European Confer- ence, Zurich, Switzerland, September 6-12, 2014, Proceed- ings, Part VI 13, pages 540–555. Springer, 2014. 2

    2014

  12. [20]

    Personalizing fast-forward videos based on visual and textual features from social net- work

    Washington Ramos, Michel Silva, Edson Araujo, Alan Neves, and Erickson Nascimento. Personalizing fast-forward videos based on visual and textual features from social net- work. In Proceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pages 3271–3280,

  13. [21]

    Query- focused video summarization: Dataset, evaluation, and a memory network based approach

    Aidean Sharghi, Jacob S Laurel, and Boqing Gong. Query- focused video summarization: Dataset, evaluation, and a memory network based approach. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 4788–4797, 2017. 2

    2017

  14. [22]

    A sparse sampling-based framework for semantic fast-forward of first-person videos.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 43(4): 1438–1444, 2020

    Michel Silva, Washington Ramos, Mario Campos, and Er- ickson R Nascimento. A sparse sampling-based framework for semantic fast-forward of first-person videos.IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 43(4): 1438–1444, 2020. 2

    2020

  15. [23]

    Tvsum: Summarizing web videos using titles

    Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. Tvsum: Summarizing web videos using titles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5179–5187, 2015. 2

    2015

  16. [24]

    To click or not to click: Automatic selection of beautiful thumbnails from videos

    Yale Song, Miriam Redi, Jordi Vallmitjana, and Alejandro Jaimes. To click or not to click: Automatic selection of beautiful thumbnails from videos. In Proceedings of the 25th ACM International on Conference on Information and Knowledge Management , page 659–668, New York, NY ,...

    2016

  17. [25]

    Efficientnetv2: Smaller mod- els and faster training

    Mingxing Tan and Quoc Le. Efficientnetv2: Smaller mod- els and faster training. In Proceedings of the 38th Inter- national Conference on Machine Learning , pages 10096– 10106. PMLR, 2021. 3

    2021

  18. [26]

    Per- sonalized egocentric video summarization of cultural tour on user preferences input

    Patrizia Varini, Giuseppe Serra, and Rita Cucchiara. Per- sonalized egocentric video summarization of cultural tour on user preferences input. IEEE Transactions on Multimedia , 19(12):2832–2845, 2017. 2

    2017

  19. [27]

    Future of tv distribution: A report prepared by the university of exeter with the university of leeds, mtm, 3 reasons and real wireless

    J Vernon, C Johnson, J Evans, R Collier, R Ellis, A Sumanasena, G Mills, A Kumar, T Jacks, B Moullier, et al. Future of tv distribution: A report prepared by the university of exeter with the university of leeds, mtm, 3 reasons and real wireless. 2024. 1

    2024

  20. [28]

    Video personalization in resource-constrained multimedia environ- ments

    Yong Wei, Suchendra M Bhandarkar, and Kang Li. Video personalization in resource-constrained multimedia environ- ments. In Proceedings of the 15th ACM international con- ference on Multimedia, pages 902–911, 2007. 2

    2007

  21. [29]

    Gaze-enabled egocentric video summarization via constrained submodular maximiza- tion

    Jia Xu, Lopamudra Mukherjee, Yin Li, Jamieson Warner, James M Rehg, and Vikas Singh. Gaze-enabled egocentric video summarization via constrained submodular maximiza- tion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2235–2244, 2015. 2

    2015

  22. [30]

    Personalized video adaptation based on video content analysis

    Min Xu, Jesse S Jin, and Suhuai Luo. Personalized video adaptation based on video content analysis. In Proceedings of the 9th International Workshop on Multimedia Data Min- ing: held in conjunction with the ACM SIGKDD 2008, pages 26–35, 2008. 2

    2008

  23. [31]

    En- coded semantic tree for automatic user profiling applied to personalized video summarization

    Yifang Yin, Roshan Thapliya, and Roger Zimmermann. En- coded semantic tree for automatic user profiling applied to personalized video summarization. IEEE Transactions on Circuits and Systems for Video Technology, 28(1):181–192,

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.