Pith. sign in

REVIEW 3 major objections 5 minor 37 references

DIVA-VQA: Detecting Inter-frame Variations in UGC Video Quality

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read DIVA-VQA claims that the largest inter-frame pixel differences mark the regions where UGC video quality is decided, and that a dual-branch network over frames, residuals, and aligned fragments turns this into top-ranked no-reference quality

desk verdict A credible fragmentation-based NR-VQA model whose headline SOTA claim depends on fine-tuning on each target dataset while baselines are zero-shot; the paper's own no-fine-tune rows show the gap largely disappears. read the letter →

arxiv 2508.10605 v1 pith:2HEXIAKS submitted 2025-08-14 eess.IV cs.CVcs.MM

classification eess.IVcs.CVcs.MM
keywords NR-VQAvideoqualityassessmentuser-generatedcontentinter-frameresidualspatchfragmentationSlowFastSwinTransformerspatio-temporalfeatures
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 aims to show that a no-reference video quality model need not look at entire frames: by ranking non-overlapping patches by their summed absolute difference between consecutive frames and keeping only the top patches, a video can be condensed into quality-sensitive fragments. A dual-branch extractor—SlowFast for motion and SwinT for spatial structure—reads the resized frame, the fragmented residual, and the position-aligned fragmented frame together. On UGC benchmarks (LSVQ, CVD2014, KoNViD-1k, LIVE-VQC, YouTube-UGC) the authors report average rank correlations of 0.898 (large) and 0.886 (base) after fine-tuning, with the base model the fastest among compared no-reference methods. The point of caring: reliable no-reference quality scoring at low runtime is what enables large-scale monitoring of user-uploaded video.

What carries the argument

Patch-difference fragmentation with residual alignment. The paper computes residual $R = |F_{\text{cur}} - F_{\text{pre}}|$, sums absolute differences inside each $p \times p$ patch to get $D_p$, and keeps the top $T = s^2/p^2$ patches by $D_p$; the same patch coordinates are then used to cut both the residual map and the current frame, producing a fragmented residual and an aligned fragmented frame. This object is the load-bearing mechanism: it localizes attention where inter-frame change is largest, links motion to spatial artifacts, and reduces the amount of video the network must process. The dual-branch extractor (SlowFast for temporal dynamics, SwinT for spatial structure) and the MLP

What would settle it

Take a UGC clip with a busy moving foreground and a static, heavily compressed/blocky background; run DIVA-VQA's patch ranking and compare its selected patches with a human quality map. If the model's correlation holds up when it is forced to use only the lowest-difference patches, or if re-ranking patches by local blur or blockiness instead of residual magnitude improves correlation, then the fragmentation premise is not doing the work claimed.

Watch

Extended reading notes

Core claim

The paper's central claim is that inter-frame variation, measured at patch granularity, is the right currency for no-reference UGC video quality. For each consecutive frame pair the method forms residual $R = |F_{\text{cur}} - F_{\text{pre}}|$, divides it into $p \times p$ patches, scores each patch by summed absolute pixel difference $D_p$, and keeps the top $T = s^2/p^2$ patches. It trains a model on three aligned inputs per chunk: the resized frame, the fragmented residual, and the fragmented frame occupying the same patch positions. The paper argues that this alignment stops the model from spreading attention over redundant static regions and points it at motion- and compression-active a

Load-bearing premise

The load-bearing premise is that the patches with the largest absolute pixel differences between consecutive frames are the patches where perceptual quality is most visibly damaged; if motion dominates those residuals while compression artifacts sit in low-difference static regions, the selection discards exactly the evidence the quality score needs.

Editorial extensions

If this is right

  • If correct, a no-reference quality model can reach state-of-the-art accuracy while processing only a small aligned subset of each frame, not the whole frame.
  • Fragmentation by inter-frame residual gives a content-adaptive way to keep high-resolution details without resizing or cropping away local artifacts.
  • The two configurations offer a practical speed/accuracy trade-off: DIVA-VQA-L tops average correlation, DIVA-VQA-B is the fastest.
  • The LSVQ-trained model transfers to other UGC datasets, and fine-tuning on target datasets adds further gains.
  • The same fragment-selection rule is directly testable for streaming and monitoring settings where chunk-level quality scores are needed.

Reading between the lines

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

  • A testable extension this reader infers: because top-T selection is hard and non-differentiable, replacing it with a differentiable soft top-k patch weighting might let the network learn which residual magnitudes matter, possibly improving generalization on videos where motion and quality-critical artifacts are not co-located.
  • The residual-hotspot rule could be read as a crude visual-attention prior; comparing DIVA-VQA's selected patches against eye-tracking or saliency maps on UGC clips would clarify whether it works by finding what viewers watch or by finding compression artifacts.
  • If the premise holds, per-chunk deployment is plausible: the method already operates on short chunks, so the same fragments could drive adaptive bitrate decisions in streaming.
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 DIVA-VQA, a no-reference video quality assessment model for user-generated content. The method extracts, for each frame, a resized frame, a fragmented residual between consecutive frames, and a fragmented frame aligned with the residual, using a patch-difference fragmentation strategy (Section 2.1). A dual-branch extractor combines SlowFast 3D motion features and Swin Transformer 2D spatial features, followed by an MLP regressor trained with a MAE plus rank loss. Experiments are reported on CVD2014, KoNViD-1k, LIVE-VQC, YouTube-UGC, and LSVQ test sets (Tables 1–2), with ablations on patch size and frame sampling (Tables 3–4) and runtime comparisons (Figure 2). The abstract and conclusion claim the method ranks among the top 2 in average rank correlation (DIVA-VQA-L: 0.898 SRCC) and outperforms existing NR-VQA methods on average at low runtime complexity.

Significance. If the central performance claim were supported by a like-for-like comparison, the paper would make a useful contribution to NR-VQA: the fragmentation idea is simple, the code is released, and the runtime results are promising. The use of public benchmarks, multiple datasets, and ablations is a strength. However, the headline claim of state-of-the-art accuracy is built on a comparison in which the proposed model is fine-tuned on each target dataset while the named baselines are evaluated zero-shot. The paper's own Table 1 shows that without fine-tuning, DIVA-VQA falls below several baselines. The significance of the empirical contribution therefore hinges on an unequal evaluation protocol, which is a correctness issue for the main claim rather than a presentation concern.

major comments (3)
  1. [§3.1, Table 1; Abstract; Conclusion] The central claim of top-2 average SRCC (DIVA-VQA-L: 0.898) is based on the 'w/ fine-tune' rows, where the model is fine-tuned for 200 epochs on 80% of each target dataset, while all compared baselines (FAST-VQA, DOVER, ReLaX-VQA, SAMA) are evaluated as LSVQ-pretrained zero-shot models. The 'wo/ fine-tune' rows show DIVA-VQA-L at 0.862 SRCC on KoNViD-1k and 0.750 on YouTube-UGC, versus DOVER at 0.909 and 0.890. This is not a like-for-like comparison and does not support the conclusion that the method 'outperforms existing NR-VQA methods on average.' The authors must either restrict the SOTA claim to the fine-tuned setting with equally fine-tuned baselines, or clearly label the zero-shot comparison as the primary one.
  2. [§3.2, Table 1] The text states 'Our proposed model exhibits the best performance across all datasets,' but Table 1 contradicts this. On KoNViD-1k, DOVER achieves SRCC 0.909 vs. DIVA-VQA-L (w/ fine-tune) 0.905; on YouTube-UGC, DOVER achieves 0.890 and SAMA 0.881 vs. DIVA-VQA-L 0.881. Even within the fine-tuned rows, DIVA is not first on every dataset. The claim should be corrected to report where the method is actually first, second, etc., and should not overstate the results.
  3. [§2.1, Eqs. (1)–(3)] The fragmentation strategy selects the top-T patches by absolute pixel difference between consecutive frames, on the premise that these are the regions where compression artifacts and perceptual quality loss are most visible. This premise is not tested. For videos with large foreground motion, the largest residuals may correspond to moving objects rather than artifact-rich static regions, and subtle compression artifacts in flat areas may have small residuals and be discarded. If this assumption fails, the claimed advantage over frame-level or random-fragment methods disappears. An ablation or analysis that directly validates the quality-sensitivity of the selected fragments is needed to support the design choice.
minor comments (5)
  1. [Throughout] The model name is repeatedly typeset as 'DIV A-VQA' with a spurious space (e.g., in the header, abstract, and table entries). This should be fixed to 'DIVA-VQA.'
  2. [§2.2, Eq. (4)] The notation 'Fresized(Ci)' mixes the symbol for a frame sequence and a chunk; clarify whether the resized component is a set of resized frames for the whole chunk or a single representative. This would improve reproducibility.
  3. [§3.1] The paper states that each experiment was repeated 21 times and the median was reported, but no variance or significance measures are given for the main results. Reporting median-only makes it hard to assess whether differences from baselines (e.g., 0.898 vs 0.886) are meaningful.
  4. [§3.3, Tables 3–4] The ablation studies are conducted on KoNViD-1k only. It would be helpful to state whether the patch-size and sampling-rate conclusions also hold on at least one other dataset, especially YouTube-UGC where residual-based motion cues may behave differently.
  5. [§3.2, Table 1] KRCC is mentioned as an evaluation metric in Section 3.1 but is not reported in Table 1 for the main comparison. If KRCC is available, reporting it would make the comparison more complete.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is an empirical comparison and no derivation step reduces to its inputs.

full rationale

The paper's derivation chain is empirical rather than definitional. Equations (1)–(3) define the residual R = |Fcur − Fpre| and the top-T patch selection Dp, which are inputs to the feature extractors, not quantities that encode the target MOS or the final SRCC. The dual-branch SlowFast/SwinT features and the MLP regressor are standard supervised components, and the quality predictions are evaluated on held-out test splits after fine-tuning on each dataset's training partition. There is no equation in which the target metric is defined in terms of the fragmentation parameters, and no fitted parameter is renamed as a prediction. The only self-citation is the explicitly disclosed inspiration from ReLaX-VQA ('We propose a patch difference fragmentation strategy inspired by ReLaX-VQA [19]'), but the fragmentation mechanism is justified in the text by its own rationale and is not used as a proof of correctness or as a uniqueness theorem. The abstract's top-2 claim is based on the 'w/ fine-tune' rows of Table 1, which is a legitimate intra-dataset protocol even though the comparison may be unequal relative to zero-shot baselines; that is a protocol-fairness concern, not circularity. Because no circular step can be exhibited with a specific reduction to the paper's own inputs, the appropriate score is 0.

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

The performance claim depends on several hand-chosen hyperparameters: loss weights, patch size, and input resolution, all selected via ablations or manual tuning. The chunking length Lc is used in Section 2.2 but not reported. No new physical or conceptual entities are introduced.

free parameters (3)
  • loss weights (maew, rankw) = 0.6, 1.0
    Section 2.3 and 3.1; chosen by hand to balance MAE and rank loss, no sensitivity analysis reported.
  • patch size p = 16
    Selected via ablation on KoNViD-1k (Table 3); affects the number of selected patches T in Eq. (3).
  • input resolution = 224x224
    Ablation showed 224 outperforms 384 for the model; this is a hand-picked operating point.
assumptions (3)
  • domain assumption The inter-frame residual R = |Fcur - Fpre| is a reliable proxy for perceptual quality changes, so selecting patches with largest Dp concentrates on quality-sensitive regions.
    Invoked in Section 2.1, Eq. (1)-(3); no perceptual evidence is provided beyond correlation with benchmark results.
  • domain assumption Features from SlowFast pretrained on Kinetics-400 and SwinT pretrained on ImageNet-22k contain the spatial and temporal information needed for quality prediction.
    Section 2.2 and 3.1; the transferability is assumed, and the paper relies on these pretrained weights rather than training from scratch.
  • ad hoc to paper The MAERank loss with weights 0.6 and 1.0 drives the regressor to a good optimum.
    Section 2.3 and 3.1; the weights are hand-picked for this paper and no sensitivity analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DIVA-VQA: Detecting Inter-frame Variations in UGC Video Quality." pith.science (2026). https://pith.science/paper/2HEXIAKS

@misc{pith2026250810605,
  author       = {Pith},
  title        = {Pith review of: DIVA-VQA: Detecting Inter-frame Variations in UGC Video Quality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2HEXIAKS}},
  note         = {Machine review of arXiv:2508.10605}
}
read the original abstract

The rapid growth of user-generated (video) content (UGC) has driven increased demand for research on no-reference (NR) perceptual video quality assessment (VQA). NR-VQA is a key component for large-scale video quality monitoring in social media and streaming applications where a pristine reference is not available. This paper proposes a novel NR-VQA model based on spatio-temporal fragmentation driven by inter-frame variations. By leveraging these inter-frame differences, the model progressively analyses quality-sensitive regions at multiple levels: frames, patches, and fragmented frames. It integrates frames, fragmented residuals, and fragmented frames aligned with residuals to effectively capture global and local information. The model extracts both 2D and 3D features in order to characterize these spatio-temporal variations. Experiments conducted on five UGC datasets and against state-of-the-art models ranked our proposed method among the top 2 in terms of average rank correlation (DIVA-VQA-L: 0.898 and DIVA-VQA-B: 0.886). The improved performance is offered at a low runtime complexity, with DIVA-VQA-B ranked top and DIVA-VQA-L third on average compared to the fastest existing NR-VQA method. Code and models are publicly available at: https://github.com/xinyiW915/DIVA-VQA.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 33 canonical work pages

  1. [1]

    INTRODUCTION Video Quality Assessment (VQA) is a critical component for optimiz- ing user experience on sharing platforms such as YouTube, Instagram, and TikTok, which attract billions of daily video views [ 1]. User- generated (video) content (UGC) is typically captured and encoded on consumer devices (smartphones or consumer-grade cameras) and is subseq...

  2. [2]

    Slow Pathway

    PROPOSED METHOD 2.1. Fragmentation using Patch Differences We propose a patch difference fragmentation method inspired by ReLaX-VQA [19], to track motion changes in the video, and employ a patch alignment strategy to ensure the consistency of the positions between the extracted fragmented frames and fragmented residuals. This method divides the video at t...

  3. [3]

    EXPERIMENTS 3.1. Evaluation setup Training & Benchmark Datasets: We conducted intra-dataset performance evaluations on four state-of-the-art in-the-wild VQA datasets: CVD2014 [ 29], KoNViD-1k [30], LIVE-VQC [ 31], and YouTube-UGC [2]. Our model was built on the large-scale LSVQ dataset [4], comprising 38,793 videos, for feature extraction and training. Th...

  4. [4]

    CONCLUSION AND FUTURE WORK In this paper, we propose a novel NR-VQA model that addresses the challenges posed by complex spatiotemporal distortions encountered in UGC videos. Our method introduces ranked patch differences be- tween consecutive video frames to generate fragmented video chunks, effectively improving the alignment between video frames and fr...

  5. [5]

    Tiktok by the numbers,

    Omnicore, “Tiktok by the numbers,” [Online], 2023, Available: https://www.omnicoreagency. com/tiktok-statistics/

  6. [6]

    Youtube ugc dataset for video compression research,

    Yilin Wang, Sasi Inguva, and Balu Adsumilli, “Youtube ugc dataset for video compression research,” in IEEE MMSP, 2019, pp. 1–5

  7. [7]

    Toward a practical perceptual video quality metric,

    Zhi Li, Anne Aaron, Ioannis Katsavounidis, Anush Moorthy, and Megha Manohara, “Toward a practical perceptual video quality metric,” The Netflix Tech Blog, vol. 6, no. 2, 2016

  8. [8]

    Patch-vq:’patching up’the video quality problem,

    Zhenqiang Ying, Maniratnam Mandal, Deepti Ghadiyaram, and Alan Bovik, “Patch-vq:’patching up’the video quality problem,” in IEEE/CVF CVPR, 2021, pp. 14019–14029

Show all 37 references
  1. [9]

    Mean squared error: Love it or leave it? a new look at signal fidelity measures,

    Zhou Wang and Alan C Bovik, “Mean squared error: Love it or leave it? a new look at signal fidelity measures,” IEEE signal processing magazine, vol. 26, no. 1, pp. 98–117, 2009

  2. [10]

    Ugc-vqa: Benchmarking blind video quality assessment for user generated content,

    Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik, “Ugc-vqa: Benchmarking blind video quality assessment for user generated content,” IEEE TIP, vol. 30, pp. 4449–4464, 2021

  3. [11]

    Blindly assess quality of in-the-wild videos via quality-aware pre-training and motion perception,

    Bowen Li, Weixia Zhang, Meng Tian, Guangtao Zhai, and Xianpei Wang, “Blindly assess quality of in-the-wild videos via quality-aware pre-training and motion perception,” IEEE TCSVT, vol. 32, no. 9, pp. 5944–5958, 2022

  4. [12]

    Exploring video quality assessment on user generated contents from aesthetic and technical perspectives,

    Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jingwen Hou, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin, “Exploring video quality assessment on user generated contents from aesthetic and technical perspectives,” in IEEE/CVF CVPR, 2023, pp. 20144–20154

  5. [13]

    Neigh- bourhood representative sampling for efficient end-to-end video quality assessment,

    Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, Jinwei Gu, and Weisi Lin, “Neigh- bourhood representative sampling for efficient end-to-end video quality assessment,” IEEE T-PAMI, vol. 45, no. 12, pp. 15185– 15202, 2023

  6. [14]

    Blind/referenceless image spatial quality evaluator,

    Anish Mittal, Anush K Moorthy, and Alan C Bovik, “Blind/referenceless image spatial quality evaluator,” in Record of the 45th Asilomar Conference on Signals, Systems and Com- puters. IEEE, 2011, pp. 723–727

  7. [15]

    Blind prediction of natural video quality,

    Michele A Saad, Alan C Bovik, and Christophe Charrier, “Blind prediction of natural video quality,” IEEE TIP, vol. 23, no. 3, pp. 1352–1365, 2014

  8. [16]

    Two-level approach for no-reference consumer video quality assessment,

    Jari Korhonen, “Two-level approach for no-reference consumer video quality assessment,” IEEE TIP, vol. 28, no. 12, pp. 5923– 5938, 2019

  9. [17]

    Quality as- sessment of in-the-wild videos,

    Dingquan Li, Tingting Jiang, and Ming Jiang, “Quality as- sessment of in-the-wild videos,” in ACM MM ’27, 2019, pp. 2351–2359

  10. [18]

    Rapique: Rapid and accurate video quality prediction of user generated content,

    Zhengzhong Tu, Xiangxu Yu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik, “Rapique: Rapid and accurate video quality prediction of user generated content,” IEEE OJSP, vol. 2, pp. 425–440, 2021

  11. [19]

    End-to- end blind quality assessment of compressed videos using deep neural networks.,

    Wentao Liu, Zhengfang Duanmu, and Zhou Wang, “End-to- end blind quality assessment of compressed videos using deep neural networks.,” in ACM Multimedia, 2018, pp. 546–554

  12. [20]

    A deep learning based no-reference quality assessment model for ugc videos,

    Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai, “A deep learning based no-reference quality assessment model for ugc videos,” in ACM MM ’30, 2022, pp. 856–865

  13. [21]

    Discovqa: Temporal distortion-content transformers for video quality assessment,

    Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, and Weisi Lin, “Discovqa: Temporal distortion-content transformers for video quality assessment,” IEEE TCSVT, 2023

  14. [22]

    Zoom-vqa: Patches, frames and clips integration for video quality assess- ment,

    Kai Zhao, Kun Yuan, Ming Sun, and Xing Wen, “Zoom-vqa: Patches, frames and clips integration for video quality assess- ment,” in IEEE/CVF CVPR, 2023, pp. 1302–1310

  15. [23]

    Relax-vqa: Residual fragment and layer stack extraction for enhancing video quality assessment,

    Xinyi Wang, Angeliki Katsenou, and David Bull, “Relax-vqa: Residual fragment and layer stack extraction for enhancing video quality assessment,” arXiv preprint:2407.11496, 2024

  16. [24]

    Temporal hystere- sis model of time varying subjective video quality,

    Kalpana Seshadrinathan and Alan C Bovik, “Temporal hystere- sis model of time varying subjective video quality,” in IEEE ICASSP, 2011, pp. 1153–1156

  17. [25]

    Long short-term convolutional transformer for no-reference video quality assessment,

    Junyong You, “Long short-term convolutional transformer for no-reference video quality assessment,” in ACM MM ’29, 2021, pp. 2112–2120

  18. [26]

    Conviqt: Contrastive video quality estimator,

    Pavan C Madhusudana, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and Alan C Bovik, “Conviqt: Contrastive video quality estimator,” IEEE TIP, vol. 32, pp. 5138–5152, 2023

  19. [27]

    Scaling and masking: A new paradigm of data sampling for image and video quality assessment,

    Yongxu Liu, Yinghui Quan, Guoyao Xiao, Aobo Li, and Jinjian Wu, “Scaling and masking: A new paradigm of data sampling for image and video quality assessment,” in AAAI Conference on Artificial Intelligence, 2024, vol. 38, pp. 3792–3801

  20. [28]

    Slowfast networks for video recognition,

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaim- ing He, “Slowfast networks for video recognition,” in IEEE/CVF CVPR, 2019, pp. 6202–6211

  21. [29]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in IEEE/CVF CVPR, 2021, pp. 10012–10022

  22. [30]

    Sgdr: Stochastic gradient descent with warm restarts,

    Ilya Loshchilov and Frank Hutter, “Sgdr: Stochastic gradient descent with warm restarts,” arXiv preprint:1608.03983, 2016

  23. [31]

    Averaging weights leads to wider optima and better generalization,

    Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson, “Averaging weights leads to wider optima and better generalization,” arXiv preprint:1803.05407, 2018

  24. [32]

    A strong baseline for image and video quality assessment,

    Shaoguo Wen and Junle Wang, “A strong baseline for image and video quality assessment,” arXiv preprint:2111.07104, 2021

  25. [33]

    Cvd2014—a database for evaluating no-reference video quality assessment algorithms,

    Mikko Nuutinen, Toni Virtanen, Mikko Vaahteranoksa, Tero Vuori, Pirkko Oittinen, and Jukka Häkkinen, “Cvd2014—a database for evaluating no-reference video quality assessment algorithms,” IEEE TIP, vol. 25, no. 7, pp. 3073–3086, 2016

  26. [34]

    The konstanz natural video database (konvid-1k),

    Vlad Hosu, Franz Hahn, Mohsen Jenadeleh, Hanhe Lin, Hui Men, Tamás Szirányi, Shujun Li, and Dietmar Saupe, “The konstanz natural video database (konvid-1k),” inQoMEX. IEEE, 2017, pp. 1–6

  27. [35]

    Large-scale study of perceptual video quality,

    Zeina Sinno and Alan Conrad Bovik, “Large-scale study of perceptual video quality,”IEEE TIP, vol. 28, no. 2, pp. 612–627, 2018

  28. [36]

    The kinetics human action video dataset,

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al., “The kinetics human action video dataset,” arXiv preprint:1705.06950, 2017

  29. [37]

    Ima- genet classification with deep convolutional neural networks,

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton, “Ima- genet classification with deep convolutional neural networks,” Advances in Neural Information Processing Systems, vol. 25, 2012

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.