Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Meta-learning at test time lets full-frame stabilizers adapt to each video and beat classical state-of-the-art stability scores, even with a single adaptation pass — while keeping the output full-frame and giving users a stability dial.

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 →

Meta-learned test-time adaptation, focused on the shakiest parts of a video, improves the stability and quality of full-frame neural video stabilizers.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid, incremental extension of the authors' CVPR 2024 meta-learning stabilizer, with useful evaluation tools; the central claim survives but needs statistical and artifact rigor. the 3 major comments →

arxiv 2508.18859 v1 pith:NIDOAT5D submitted 2025-08-26 cs.CV

Harnessing Meta-Learning for Controllable Full-Frame Video Stabilization

classification cs.CV
keywords video stabilizationmeta-learningMAMLtest-time adaptationfull-frame video synthesisjerk localizationglobal optical flowvideo-language models
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 reading

This paper takes aim at a trade-off at the center of video stabilization: classical transform-based methods are stable and controllable but crop away visual content, while modern full-frame synthesis networks keep the whole frame but have fixed parameters that cannot cope with the diversity of real-world motion, and they give the user no control. The authors claim that a model-agnostic meta-learning (MAML) treatment resolves both problems: a stabilizer is meta-trained so that at test time it can adapt to the motion profile of each input video using only the unstable video itself. Their headline evidence is that a single adaptation pass lifts stability by up to ~8% absolute on the NUS benchmark, overtaking the longstanding leaders while keeping full-frame output. A jerk-localization module that directs adaptation to the most unstable segments, plus corner-and-center patch sampling, cuts the number of adaptation steps by about 96% relative to uniform adaptation — which, if correct, is what makes test-time adaptation practical for real, long videos.

Core claim

The paper's central claim is that pixel-level full-frame stabilizers underperform not because their architectures are wrong but because fixed parameters cannot cover the range of real-world motion profiles, and that this can be cured by making the model meta-learn how to adapt. The authors employ MAML with two loops: an inner loop adjusts the stabilizer's parameters on each test video with a self-supervised loss — the absolute mean of global optical flow between synthesized frames and a rigid-affine-aligned version of the same frames, combined with perceptual, gram-matrix, and contextual quality terms — and an outer loop trains the initialization using DeepStab ground-truth stable frames so

What carries the argument

Three components carry the argument. (1) The rigid-affine estimation network h_phi: a U-net-style encoder with a fully connected head, trained on randomly rigid-transformed images to regress rotation and translation from global optical flow; after training it aligns each frame of a shaky clip to the clip's first frame, producing the aligned sequence that stands in for ground truth. (2) The MAML inner/outer loop: the inner loop minimizes the absolute mean of global optical flow between synthesized and aligned frames plus perceptual, gram-matrix, and contextual quality losses; the outer loop, using DeepStab stable frames, penalizes deviation of the synthesized flow trajectory from the ground-t

Load-bearing premise

The load-bearing premise is that the rigid-affine estimation network — trained only on random synthetic transforms — aligns real shaky frames accurately enough for the flow-space adaptation loss and jerk-localized peak selection to be trustworthy; if it misreads camera motion under rolling shutter, parallax, low light, or large moving objects, the adaptation gradient is corrupted, the chosen segments are wrong, and the claimed gains shrink or reverse.

What would settle it

Take videos with strong rolling shutter and parallax (or synthesize them by mixing known camera paths with scene depth), compute the affine estimator's alignment error against an independent robust registration such as feature-based homographies, and correlate that error with the stability gain from one adaptation pass. If large alignment errors still yield the ~8% gain, the mechanism is sound; if the gain vanishes or reverses where the affine guide errs, the central claim fails. A cheaper variant: replace the affine network's estimates with ground-truth camera motion during adaptation — if ga

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

If this is right

  • Any existing full-frame pixel-synthesis stabilizer — DMBVS, DIFRINT, or future ones — can be wrapped in this adaptation loop with no architectural changes, gaining several absolute points of stability on standard benchmarks.
  • The user gains a dial: choosing how many segments (p) and how many gradient steps (M) to adapt trades stability against visual quality and runtime, a control mechanism classical methods had and regression models previously lacked.
  • Targeted adaptation on jerk peaks delivers roughly the stability of uniform adaptation with about 96% fewer adaptation steps, which is the difference between a research trick and something deployable on long consumer or professional videos.
  • Stabilization gains propagate downstream: adapted outputs improve object-persistence and temporal-IoU metrics (tracking feasibility) and raise caption-quality scores across four large video-language models.
  • Because adaptation needs only the unstable input, the method applies without paired ground truth at test time and generalizes across datasets (NUS, DOFVS, BiT) whose motion profiles differ from the meta-training set.

Where Pith is reading between the lines

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

  • The affine-guide trick is a generic no-ground-truth stabilization proxy: the same construction — regress a rigid transform, warp, and penalize flow residuals to the warp — could supervise other motion-sensitive restoration tasks such as deblurring, frame interpolation, or super-resolution that currently need paired data or synthetic degradations.
  • The jerk-localization module is effectively an unsupervised detector of camera-motion salience; it could be repurposed as a failure predictor (flagging segments where a stabilizer is most likely to distort) or as an editing heuristic telling a human which clips deserve attention.
  • The paper reports endpoints (p = 10 vs 100, M = 1 vs 5) but not the full stability-quality frontier; a testable extension is to make p and M per-segment decisions — adapting more on the highest jerk peaks and less elsewhere — which should approach full-adaptation quality at near-targeted cost.
  • The LLM-as-a-judge protocol, grounded in open-vocabulary detections, is demonstrated for video understanding; the same scaffold could score stability itself (e.g., judging smoothness across sampled frames), giving an end-to-end reproducible substitute for user studies without extra modeling.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper proposes a meta-learning framework for full-frame video stabilization. Using MAML, the authors meta-train two existing full-frame synthesis baselines (DMBVS and DIFRINT) so that at test time a self-supervised inner-loop loss adapts the model to each input video. The inner-loop loss combines a stability term based on rigid-affine alignment (Eqs. 5-6) with quality losses (Eq. 7); a jerk-localization module selects high-jerk segments and spatially informative crops for targeted adaptation. Evaluations on NUS, BiT, and DOFVS report stability, cropping, and distortion scores, claiming state-of-the-art stability on NUS with a single adaptation pass, and the paper also introduces two new metrics and an LLM-as-a-Judge protocol for downstream evaluation.

Significance. If the empirical claims hold, the work is significant: it demonstrates a model-agnostic method for improving full-frame pixel-synthesis stabilizers at test time without ground-truth stable data, and it addresses the limited controllability of such methods. The paper uses the standard external FFT stability metric for the central claim, so the main comparison is not circular. It also evaluates on multiple datasets and includes ablation studies. However, the significance is currently limited by the lack of validation of a load-bearing component (the affine estimation network), the absence of statistical rigor in the quantitative comparisons, and the reliance on two newly introduced metrics whose validation is only qualitative.

major comments (3)
  1. [Sec. III-B1, Eqs. (5)-(8); Sec. III-B2, Eqs. (12)-(13)] The test-time adaptation signal is entirely mediated by the affine estimation network h_phi, which is trained only on randomly generated rigid transforms (Eqs. 2-4). The paper never reports h_phi's alignment accuracy on the evaluation datasets, nor does it ablate the sensitivity of the inner-loop loss to this module. On real videos with rolling shutter, parallax, or low light, a single rigid transform cannot align frames; an inaccurate aligned frame ~I_t in Eq. (5) means the stability loss in Eq. (6) penalizes the model for not matching a moving reference, and the jerk peaks in Eqs. (12)-(13) can be misplaced. Since the central claim in Sec. I (up to ~8% absolute gain with a single adaptation pass) is carried by this loss, the missing quantification is load-bearing. Please report alignment error of h_phi on representative clips from NUS/BiT/DOFVS and include an ablation with an oracle or
  2. [Tables IV-XI, especially Tab. VII] All quantitative results are single point estimates with no error bars, confidence intervals, or significance tests. The headline SOTA claim rests on a 0.0036 stability difference (0.8786 vs. 0.8750) in Tab. VII, and Tab. IV shows a tie (0.8492) between DIFRINT TargetedAdapt(3)_10 and VanillaAdapt(1)_100 on the Parallax category. Without repeated runs or per-video paired statistics, these differences could be within run-to-run or video-to-video variance. Please provide standard deviations over independent runs, or per-video paired comparisons with significance tests (e.g., Wilcoxon signed-rank) for at least the main tables supporting the SOTA claim.
  3. [Sec. IV-B2 and Alg. 3] The targeted adaptation strategy relies on several unspecified hyperparameters: the peak-detection distance constraint, the number of tasks p, and the crop operator C in Fig. 6. Algorithm 3 line 10 also says 'Evaluate inner loss ∇θ′ Lin_T', which should be 'compute inner loss Lin_T and its gradient' before the update. The choice of these parameters affects the reported efficiency gain (e.g., 96% fewer steps in Tab. II), so the main text should specify the exact settings and, ideally, include a sensitivity analysis over the peak distance constraint and crop sizes.
minor comments (5)
  1. [Sec. IV-D3 and Tab. X] The two proposed metrics (Average Persistence and Temporal IoU) are introduced and used to claim SOTA results, but their validation is only qualitative ('trends consistent with established evaluation measures'), and implementation details are deferred to the supplementary material. Please provide quantitative correlation with standard metrics or an independent validation, and specify the detection/tracking protocol in the main text.
  2. [Sec. IV-D1] The text states that the adapted models show 'significant improvements across all metrics', but Table V shows a decrease in cropping for DMBVS under TargetedAdapt settings, and the table caption acknowledges a 'minor decrease in cropping'. Please reconcile this wording.
  3. [Eq. (6)] The definition of Lin_stability is ambiguous: it sums over t and also averages over N pixels. Please define the pixel summation explicitly, e.g., 1/N times a double sum over t and pixels.
  4. [Throughout] Typos: 'Stabililty' (Sec. IV-D), 'V anillaAdapt' (Sec. IV-B2), 'Avg Persistance' (Tab. X), 'theri corresponding captions' (Sec. IV-E1).
  5. [Sec. II and Sec. IV-D] DUT [14] is discussed in Related Works but is not included in the quantitative comparison. If it is omitted for a specific reason, please state it; otherwise consider adding it to the comparison tables.

Circularity Check

0 steps flagged

No significant circularity: the central stability claim is verified with the external FFT benchmark, and the self-supervised adaptation loss is a proxy rather than the evaluation metric.

full rationale

The paper's headline claim (Sec. I, up to ~8% absolute stability gain on NUS) is measured with the standard external FFT trajectory metric defined in Eqs. (14)-(15), not with the proposed self-supervised losses. The test-time adaptation uses an inner-loop loss (Eqs. 6-8) that compares regressed frames to frames aligned by an affine-estimation network h_phi, which is independently trained on random rigid transforms (Eqs. 2-4) and is not fitted to the evaluation metric or to the test videos. Thus the main result does not reduce to the method's own objective. The introduced metrics (Average Persistence, Temporal IoU) are sanity-checked against conventional metrics and human/LLM judgments; that is consistency verification, not circular validation. The LLM-as-a-Judge study uses a small sample (12 pairs, 31 participants), but this is a limitation in statistical power, not a circularity. The paper does contain self-citations, notably 'Inspired by the findings of Ali et al. [6]' for using a reconstruction objective as a stabilization proxy and 'As an additional contribution to our prior work [9]' for the meta-learning framework, but these are not load-bearing for the central external benchmark result; the paper also provides its own ablations (Sec. IV-B and supplementary S-IV). The skepticism about h_phi's accuracy on real-world videos with rolling shutter, parallax, or low light is a legitimate correctness risk, but it is not a circularity because h_phi is not derived from the FFT metric or from the adapted model. Overall, the derivation chain is self-contained against external benchmarks, with only minor self-citation that does not force the conclusions.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical or conceptual entities; the affine network, jerk localization, and metrics are trained/defined constructs rather than postulated objects. The load-bearing content is the set of domain assumptions listed above.

free parameters (6)
  • Inner-loop loss weights = lambda_s : lambda_q = 10:1
    Empirically chosen in Sec. III-B1 to balance stability vs quality during adaptation; directly shapes the adapted model.
  • Outer-loop loss weights = lambda_s : lambda_q = 1:10
    Empirically chosen in Sec. III-B1 for meta-training; emphasizes quality because stable GT is available.
  • Number of local windows r = 5
    Set to T'=r=5 due to resource limitations (Sec. III-B1).
  • Adaptation steps and peak count = M=3, p=10 for targeted; M=1 or M=5, p=100 for vanilla
    User/experiment-selected budgets in Sec. IV-B; the control mechanism claim depends on these.
  • Peak detection distance constraint = 2k+1 frames
    Peaks are constrained so no two lie within a temporal window (Sec. III-B2).
  • Spatial crop operator C = diagonal patches (top-left, center, bottom-right)
    Heuristic 'empirically find' choice in Sec. III-B2; patch size not given in main text.
axioms (6)
  • domain assumption GlobalFlowNet [25] produces global optical flow that masks local object motion and fills boundary gaps
    Used to define both the affine training (Eq. 2) and the inner stability loss (Eq. 6).
  • domain assumption Affine estimation network h_phi trained on random rigid transforms generalizes to real shaky videos
    The entire self-supervised adaptation signal in Eq. (8) and the jerk localization in Eq. (13) depend on this alignment.
  • domain assumption VGG-16 perceptual and gram features are valid quality proxies
    Used in Eq. (7) to preserve content during adaptation.
  • domain assumption DeepStab stable videos are valid outer-loop targets despite perspective mismatch with unstable inputs
    Outer loop quality loss uses contextual loss to handle non-alignment (Eq. 10), but stability loss compares global flows (Eq. 9).
  • domain assumption FFT-based stability metric (Eq. 14-15) reflects perceived stability
    All headline stability scores use this metric; it measures low-frequency energy in feature trajectories, not human perception directly.
  • domain assumption LLM-as-a-Judge scores correlate with human ratings of caption quality
    Sec. IV-E validates on only 12 video-caption pairs before using the judge as the downstream evaluation tool.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Harnessing Meta-Learning for Controllable Full-Frame Video Stabilization." pith.science (2026). https://pith.science/paper/NIDOAT5D

@misc{pith2026250818859,
  author       = {Pith},
  title        = {Pith review of: Harnessing Meta-Learning for Controllable Full-Frame Video Stabilization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NIDOAT5D}},
  note         = {Machine review of arXiv:2508.18859}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Video stabilization remains a fundamental problem in computer vision, particularly pixel-level synthesis solutions for video stabilization, which synthesize full-frame outputs, add to the complexity of this task. These methods aim to enhance stability while synthesizing full-frame videos, but the inherent diversity in motion profiles and visual content present in each video sequence makes robust generalization with fixed parameters difficult. To address this, we present a novel method that improves pixel-level synthesis video stabilization methods by rapidly adapting models to each input video at test time. The proposed approach takes advantage of low-level visual cues available during inference to improve both the stability and visual quality of the output. Notably, the proposed rapid adaptation achieves significant performance gains even with a single adaptation pass. We further propose a jerk localization module and a targeted adaptation strategy, which focuses the adaptation on high-jerk segments for maximizing stability with fewer adaptation steps. The proposed methodology enables modern stabilizers to overcome the longstanding SOTA approaches while maintaining the full frame nature of the modern methods, while offering users with control mechanisms akin to classical approaches. Extensive experiments on diverse real-world datasets demonstrate the versatility of the proposed method. Our approach consistently improves the performance of various full-frame synthesis models in both qualitative and quantitative terms, including results on downstream applications.

Figures

Figures reproduced from arXiv: 2508.18859 by Dongjin Kim, Eun Woo Im, Haonan Luo, Muhammad Kashif Ali, Tae Hyun Kim, Tianrui Li, Vivek Gupta.

Figure 2
Figure 2. Figure 2: Comparison of conventional and global optical flow. Optical flow estimated from RAFT [56] (bottom left) and Global Optical Flow [25] (bottom right). Global optical flow, in addition to masking locally different motions by dynamic objects or abrupt depth changes, also fills in the gaps near the frame boundaries introduced due to coarse alignment and only provides global motion, which supports coarse alignme… view at source ↗
Figure 3
Figure 3. Figure 3: Affine estimation network. Overview of architecture and the training pipeline of the proposed affine estimation network. The input image I is transformed to I ′ with a random rigid affine A, I ′ is used for both global flow estimation and as the target for the pixel loss Lpixel to train this network. The network learns to regress rigid affine parameters Aˆ, containing rotation and translation, from the est… view at source ↗
Figure 4
Figure 4. Figure 4: Affine alignment. This affine alignment strategy is analogous to the classical stabilization strategies which estimate and smooth transforms to stabilize videos. Please note that these frames are not neighboring frames and were selected to highlight the crops near the image boundaries in aligned frames V˜ . where Aˆ denotes the estimated rigid-affine transform, hϕ is the proposed affine estimation network,… view at source ↗
Figure 5
Figure 5. Figure 5: Jerk Localization. Instantaneous jerk intensity of a video from the NUS dataset is presented. Red dots mark high-magnitude, non-overlapping tasks selected for guiding the proposed targeted adaptation. For this illustration we use p = 10 and k = 2. (a) Original frame (blue) with cropping areas (red) (b) Pure Rotation (c) Pure Transition [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Spatially Aware Patch Sampling Strategy. Spatially aware crop selection strategy via structured crop operator C used for targeted optimization. (a) The original frame (blue) and adaptation patches selected by C at the corners and the center (red). (b) Under pure rotation, corner regions undergo the most significant displacement, making them highly informative for stability supervision, while the center pat… view at source ↗
Figure 7
Figure 7. Figure 7: Finetuning vs. meta-adaptation. The dashed lines represent the stability scores achieved by the meta-adapted models, while the solid lines denote the results obtained through naive finetuning. Naive finetuning takes more iterations for a mediocre improvement in the stability score. Whereas, the proposed meta-adaptation algorithm achieves a significant performance gain even with a single adaptation pass ove… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative Results. This figure presents the qualitative comparison of the meta-trained, baseline models, and classical SOTA methods. The proposed test-time adaptation strategy not only improves the stability but also enhances the visual quality of the stabilized videos (Best viewed on a computer screen with zoom). videos. Recent approaches, despite their intricate designs and efficacies, fall short in co… view at source ↗
Figure 9
Figure 9. Figure 9: Comparison Between LLM and Human Evaluations This figure presents the average scores by human subjects and the proposed LLM-as-a￾judge pipeline. Dashed horizontal lines indicate the mean scores across all categories present in the NUS dataset. The proposed method demonstrates consistent scoring trends with human evaluations; however, a noticeable disparity appears in categories involving a large number of … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Concrete Jungle: Towards Concreteness Paved Contrastive Negative Mining for Compositional Understanding

    cs.LG 2026-04 unverdicted novelty 5.0

    Using lexical concreteness to guide contrastive negative mining and a new margin-based Cement loss, the Slipform framework reaches state-of-the-art on compositional benchmarks for vision-language models.

Reference graph

Works this paper leans on

77 extracted references · 71 canonical work pages · cited by 1 Pith paper · 6 internal anchors

  1. [1]

    Auto-directed video stabilization with robust l1 optimal camera paths,

    M. Grundmann, V . Kwatra, and I. Essa, “Auto-directed video stabilization with robust l1 optimal camera paths,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2011

  2. [2]

    Bundled camera paths for video stabilization,

    S. Liu, L. Yuan, P. Tan, and J. Sun, “Bundled camera paths for video stabilization,” ACM TOG, 2013

  3. [3]

    Learning video stabilization using optical flow,

    J. Yu and R. Ramamoorthi, “Learning video stabilization using optical flow,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2020

  4. [4]

    Out-of-boundary view synthesis towards full-frame video stabilization,

    Y . Xu, J. Zhang, and D. Tao, “Out-of-boundary view synthesis towards full-frame video stabilization,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2021

  5. [5]

    Self-supervised real-time video stabilization,

    J. Choi, J. Park, and I. S. Kweon, “Self-supervised real-time video stabilization,” arXiv preprint arXiv:2111.05980 , 2021

  6. [6]

    Deep Motion Blind Video Stabilization

    M. K. Ali, S. Yu, and T. H. Kim, “Deep motion blind video stabilization,” arXiv preprint arXiv:2011.09697 , 2020

  7. [7]

    Deep iterative frame interpolation for full-frame video stabilization,

    J. Choi and I. S. Kweon, “Deep iterative frame interpolation for full-frame video stabilization,” ACM TOG, 2020

  8. [8]

    Hybrid neural fusion for full-frame video stabilization,

    Y .-L. Liu, W.-S. Lai, M.-H. Yang, Y .-Y . Chuang, and J.-B. Huang, “Hybrid neural fusion for full-frame video stabilization,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2021

  9. [9]

    Harnessing meta-learning for improving full-frame video stabilization,

    M. K. Ali, E. W. Im, D. Kim, and T. H. Kim, “Harnessing meta-learning for improving full-frame video stabilization,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2024, pp. 12 605–12 614

  10. [10]

    Full-reference stability assessment of digital video stabilization based on riemannian metric,

    L. Zhang, Q.-Z. Zheng, H.-K. Liu, and H. Huang, “Full-reference stability assessment of digital video stabilization based on riemannian metric,” IEEE Transactions on Image Processing (TIP) , vol. 27, no. 12, pp. 6051–6063, 2018

  11. [11]

    Deep online fused video stabilization,

    Z. Shi, F. Shi, W.-S. Lai, C.-K. Liang, and Y . Liang, “Deep online fused video stabilization,” in Winter Conference on Applications of Computer Vision (WACV), 2022, pp. 1250–1258

  12. [12]

    Video stabilization with a depth camera,

    S. Liu, Y . Wang, L. Yuan, J. Bu, P. Tan, and J. Sun, “Video stabilization with a depth camera,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2012

  13. [13]

    Digital video stabilization and rolling shutter correction using gyroscopes,

    A. Karpenko, D. Jacobs, J. Baek, and M. Levoy, “Digital video stabilization and rolling shutter correction using gyroscopes,” CSTR, 2011

  14. [14]

    Dut: Learning video stabilization by simply watching unstable videos,

    Y . Xu, J. Zhang, S. J. Maybank, and D. Tao, “Dut: Learning video stabilization by simply watching unstable videos,” IEEE Transactions on Image Processing (TIP) , 2022

  15. [15]

    Content-preserving warps for 3d video stabilization,

    F. Liu, M. Gleicher, H. Jin, and A. Agarwala, “Content-preserving warps for 3d video stabilization,” ACM Transactions on Graphics (SIGGRAPH) , 2009

  16. [16]

    Light field video stabilization,

    B. M. Smith, L. Zhang, H. Jin, and A. Agarwala, “Light field video stabilization,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2009

  17. [17]

    Plane-based content preserving warps for video stabilization,

    Z. Zhou, H. Jin, and Y . Ma, “Plane-based content preserving warps for video stabilization,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2013

  18. [18]

    Non-metric image-based rendering for video stabilization,

    C. Buehler, M. Bosse, and L. McMillan, “Non-metric image-based rendering for video stabilization,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2001

  19. [19]

    Full-frame video stabilization with motion inpainting,

    Y . Matsushita, E. Ofek, W. Ge, X. Tang, and H.-Y . Shum, “Full-frame video stabilization with motion inpainting,” IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) , 2006

  20. [20]

    Subspace video stabilization,

    F. Liu, M. Gleicher, J. Wang, H. Jin, and A. Agarwala, “Subspace video stabilization,” ACM TOG, 2011

  21. [21]

    Video stabilization using epipolar geometry,

    A. Goldstein and R. Fattal, “Video stabilization using epipolar geometry,” ACM TOG, 2012

  22. [22]

    Video stabilization using robust feature trajectories,

    K.-Y . Lee, Y .-Y . Chuang, B.-Y . Chen, and M. Ouhyoung, “Video stabilization using robust feature trajectories,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2009

  23. [23]

    Spatially and temporally optimized video stabilization,

    Y .-S. Wang, F. Liu, P.-S. Hsu, and T.-Y . Lee, “Spatially and temporally optimized video stabilization,” IEEE transactions on visualization and computer graphics , 2013

  24. [24]

    Robust video stabilization by optimization in cnn weight space,

    J. Yu and R. Ramamoorthi, “Robust video stabilization by optimization in cnn weight space,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2019

  25. [25]

    Globalflownet: Video stabilization using deep distilled global motion estimates,

    J. Geo, D. Jain, and A. Rajwade, “Globalflownet: Video stabilization using deep distilled global motion estimates,” in Winter Conference on Applications of Computer Vision (WACV) , 2023

  26. [26]

    Fast adaptation to super- resolution networks via meta-learning,

    S. Park, J. Yoo, D. Cho, J. Kim, and T. H. Kim, “Fast adaptation to super- resolution networks via meta-learning,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2020

  27. [27]

    Ada-vsr: Adaptive video super-resolution with meta-learning,

    A. Gupta, P. Jonnalagedda, B. Bhanu, and A. K. Roy-Chowdhury, “Ada-vsr: Adaptive video super-resolution with meta-learning,” in ACM International Conference on Multimedia (MM) , 2021

  28. [28]

    Dynavsr: Dynamic adaptive blind video super-resolution,

    S. Lee, M. Choi, and K. M. Lee, “Dynavsr: Dynamic adaptive blind video super-resolution,” in Winter Conference on Applications of Computer Vision (WACV), 2021

  29. [29]

    Deep meta learning for real-time target-aware visual tracking,

    J. Choi, J. Kwon, and K. M. Lee, “Deep meta learning for real-time target-aware visual tracking,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2019

  30. [30]

    Meta-learning deep visual words for fast video object segmentation,

    H. S. Behl, M. Naja, A. Arnab, and P. H. Torr, “Meta-learning deep visual words for fast video object segmentation,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2020

  31. [31]

    Minet: Meta-learning instance identifiers for video object detection,

    J. Deng, Y . Pan, T. Yao, W. Zhou, H. Li, and T. Mei, “Minet: Meta-learning instance identifiers for video object detection,” IEEE Transactions on Image Processing (TIP) , 2021

  32. [32]

    Camera distortion-aware 3d human pose estimation in video with optimization-based meta-learning,

    H. Cho, Y . Cho, J. Yu, and J. Kim, “Camera distortion-aware 3d human pose estimation in video with optimization-based meta-learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2021

  33. [33]

    Bilevel Fast Scene Adaptation for Low-Light Image Enhancement

    L. Ma, D. Jin, N. An, J. Liu, X. Fan, and R. Liu, “Bilevel fast scene adaptation for low-light image enhancement,” arXiv preprint arXiv:2306.01343, 2023

  34. [34]

    Scene-adaptive video frame interpolation via meta-learning,

    M. Choi, J. Choi, S. Baik, T. H. Kim, and K. M. Lee, “Scene-adaptive video frame interpolation via meta-learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2020

  35. [35]

    Siamese neural networks for one-shot image recognition,

    G. Koch, R. Zemel, R. Salakhutdinov et al. , “Siamese neural networks for one-shot image recognition,” in ICML deep learning workshop , vol. 2, no. 1. Lille, 2015, pp. 1–30

  36. [36]

    Prototypical networks for few-shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few-shot learning,” Advances in Neural Information Processing Systems (NIPS) , vol. 30, 2017

  37. [37]

    Matching networks for one shot learning,

    O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstra et al. , “Matching networks for one shot learning,” Advances in Neural Information Processing Systems (NIPS) , vol. 29, 2016

  38. [38]

    Meta networks,

    T. Munkhdalai and H. Yu, “Meta networks,” in International Conference on Machine Learning (ICML) . PMLR, 2017, pp. 2554–2563

  39. [39]

    Rapid adaptation with conditionally shifted neurons,

    T. Munkhdalai, X. Yuan, S. Mehri, and A. Trischler, “Rapid adaptation with conditionally shifted neurons,” in International Conference on Machine Learning (ICML) . PMLR, 2018, pp. 3664–3673

  40. [40]

    Meta-learning with memory-augmented neural networks,

    A. Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. Lillicrap, “Meta-learning with memory-augmented neural networks,” in Interna- tional Conference on Machine Learning (ICML) . PMLR, 2016, pp. 1842–1850

  41. [41]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” in International Conference on Machine Learning (ICML) , 2017

  42. [42]

    Meta-learning-based incremental few-shot object detection,

    M. Cheng, H. Wang, and Y . Long, “Meta-learning-based incremental few-shot object detection,” IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) , 2021

  43. [43]

    Embodied one-shot video recognition: Learning from actions of a virtual embodied agent,

    Y . Fu, C. Wang, Y . Fu, Y .-X. Wang, C. Bai, X. Xue, and Y .-G. Jiang, “Embodied one-shot video recognition: Learning from actions of a virtual embodied agent,” in ACM International Conference on Multimedia (MM) , 2019

  44. [44]

    Few-shot personality-specific image captioning via meta-learning,

    M. Hosseinzadeh and Y . Wang, “Few-shot personality-specific image captioning via meta-learning,” in Conference on Robots and Vision , 2023

  45. [45]

    MetaPix: Few-Shot Video Retargeting

    J. Lee, D. Ramanan, and R. Girdhar, “Metapix: Few-shot video retargeting,” arXiv preprint arXiv:1910.04742 , 2019

  46. [46]

    Self-supervised video representation learning with meta-contrastive network,

    Y . Lin, X. Guo, and Y . Lu, “Self-supervised video representation learning with meta-contrastive network,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2021

  47. [47]

    A deep meta-learning neural network for single image rain removal,

    Y . Lu, J. Cai, H. Zheng, and Y . Zeng, “A deep meta-learning neural network for single image rain removal,” in International Congress on Image and Signal Processing, BioMedical Engineering and Informatics , 2020

  48. [48]

    Video Deblurring by Fitting to Test Data

    X. Ren, Z. Qian, and Q. Chen, “Video deblurring by fitting to test data,” arXiv preprint arXiv:2012.05228 , 2020

  49. [49]

    Meta-learning based siamese network with channel-wise self-attention for visual tracking,

    R. Wang, B. Kang, and W.-P. Zhu, “Meta-learning based siamese network with channel-wise self-attention for visual tracking,” in International Conference on Image, Video and Signal Processing , 2021

  50. [50]

    Tracking by instance detection: A meta-learning approach,

    G. Wang, C. Luo, X. Sun, Z. Xiong, and W. Zeng, “Tracking by instance detection: A meta-learning approach,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2020

  51. [51]

    Toward human perception-centric video thumbnail generation,

    T. Yang, F. Wang, J. Lin, Z. Qi, Y . Wu, J. Xu, Y . Shan, and C. Chen, “Toward human perception-centric video thumbnail generation,” in ACM International Conference on Multimedia (MM) , 2023. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 16

  52. [52]

    One to many: Adaptive instrument segmentation via meta learning and dynamic online adaptation in robotic surgical video,

    Z. Zhao, Y . Jin, B. Lu, C.-F. Ng, Q. Dou, Y .-H. Liu, and P.-A. Heng, “One to many: Adaptive instrument segmentation via meta learning and dynamic online adaptation in robotic surgical video,” in IEEE International Conference on Robotics and Automation (ICRA) , 2021

  53. [53]

    L 2 c–learning to learn to compress,

    N. Zou, H. Zhang, F. Cricri, H. R. Tavakoli, J. Lainema, M. Hannuksela, E. Aksu, and E. Rahtu, “L 2 c–learning to learn to compress,” in IEEE 22nd International Workshop on Multimedia Signal Processing , 2020

  54. [54]

    Meta-learning for adaptation of deep optical flow networks,

    C. Min, T. H. Kim, and J. Lim, “Meta-learning for adaptation of deep optical flow networks,” in Winter Conference on Applications of Computer Vision (WACV), 2023

  55. [55]

    Deep online video stabilization with multi-grid warping transformation learning,

    M. Wang, G.-Y . Yang, J.-K. Lin, S.-H. Zhang, A. Shamir, S.-P. Lu, and S.-M. Hu, “Deep online video stabilization with multi-grid warping transformation learning,” IEEE Transactions on Image Processing (TIP) , 2018

  56. [56]

    Raft: Recurrent all-pairs field transforms for optical flow,

    Z. Teed and J. Deng, “Raft: Recurrent all-pairs field transforms for optical flow,” in Proceedings of the European Conference on Computer Vision (ECCV), 2020

  57. [57]

    Task Agnostic Restoration of Natural Video Dynamics

    M. K. Ali, D. Kim, and T. H. Kim, “Learning task agnostic temporal consistency correction,” arXiv preprint arXiv:2206.03753 , 2022

  58. [58]

    Perceptual losses for real-time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” in European conference on computer vision. Springer, 2016, pp. 694–711

  59. [59]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2009

  60. [60]

    The contextual loss for image transformation with non-aligned data,

    R. Mechrez, I. Talmi, and L. Zelnik-Manor, “The contextual loss for image transformation with non-aligned data,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2018

  61. [61]

    Multiple testing of local maxima for detection of peaks in random fields,

    D. Cheng and A. Schwartzman, “Multiple testing of local maxima for detection of peaks in random fields,” Annals of statistics , vol. 45, no. 2, p. 529, 2019

  62. [62]

    Minimum latency deep online video stabilization,

    Z. Zhang, Z. Liu, P. Tan, B. Zeng, and S. Liu, “Minimum latency deep online video stabilization,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2023

  63. [63]

    Video stabilization: A comprehensive survey,

    Y . Wang, Q. Huang, C. Jiang, J. Liu, M. Shang, and Z. Miao, “Video stabilization: A comprehensive survey,” Neurocomputing, 2022

  64. [64]

    Video-llava: Learning united visual representation by alignment before projection,

    B. Lin, Y . Ye, B. Zhu, J. Cui, M. Ning, P. Jin, and L. Yuan, “Video-llava: Learning united visual representation by alignment before projection,” arXiv preprint arXiv:2311.10122 , 2023

  65. [65]

    Sharegpt4video: Improving video understanding and generation with better captions,

    L. Chen, X. Wei, J. Li, X. Dong, P. Zhang, Y . Zang, Z. Chen, H. Duan, Z. Tang, L. Yuan et al., “Sharegpt4video: Improving video understanding and generation with better captions,” Advances in Neural Information Processing Systems (NIPS) , vol. 37, pp. 19 472–19 495, 2024

  66. [66]

    Llava-next: A strong zero-shot video understanding model,

    Y . Zhang, B. Li, h. Liu, Y . j. Lee, L. Gui, D. Fu, J. Feng, Z. Liu, and C. Li, “Llava-next: A strong zero-shot video understanding model,” April 2024. [Online]. Available: https://llava-vl.github.io/blog/ 2024-04-30-llava-next-video/

  67. [67]

    Videollama 3: Frontier multimodal foundation models for image and video understanding,

    B. Zhang, K. Li, Z. Cheng, Z. Hu, Y . Yuan, G. Chen, S. Leng, Y . Jiang, H. Zhang, X. Li et al. , “Videollama 3: Frontier multimodal foundation models for image and video understanding,” arXiv preprint arXiv:2501.13106, 2025

  68. [68]

    Deep variational bayesian modeling of haze degradation process,

    E. W. Im, J. Shin, S. Baik, and T. H. Kim, “Deep variational bayesian modeling of haze degradation process,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , 2023, pp. 895–904

  69. [69]

    VidCapBench: A Comprehensive Benchmark of Video Captioning for Controllable Text-to-Video Generation

    X. Chen, Y . Zhang, C. Rao, Y . Guan, J. Liu, F. Zhang, C. Song, Q. Liu, D. Zhang, and T. Tan, “Vidcapbench: A comprehensive benchmark of video captioning for controllable text-to-video generation,” arXiv preprint arXiv:2502.12782, 2025

  70. [70]

    Vidhalluc: Evaluating temporal hallucina- tions in multimodal large language models for video understanding,

    C. Li, E. W. Im, and P. Fazli, “Vidhalluc: Evaluating temporal hallucina- tions in multimodal large language models for video understanding,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2025, pp. 13 723–13 733

  71. [71]

    Scaling open-vocabulary object detection,

    M. Minderer, A. Gritsenko, and N. Houlsby, “Scaling open-vocabulary object detection,” Advances in Neural Information Processing Systems (NIPS), vol. 36, pp. 72 983–73 007, 2023

  72. [72]

    Semantic understanding of scenes through the ade20k dataset,

    B. Zhou, H. Zhao, X. Puig, T. Xiao, S. Fidler, A. Barriuso, and A. Torralba, “Semantic understanding of scenes through the ade20k dataset,” International Journal of Computer Vision (IJCV) , vol. 127, pp. 302–321, 2019

  73. [73]

    Blink: Multimodal large language models can see but not perceive,

    X. Fu, Y . Hu, B. Li, Y . Feng, H. Wang, X. Lin, D. Roth, N. A. Smith, W.-C. Ma, and R. Krishna, “Blink: Multimodal large language models can see but not perceive,” in Proceedings of the European Conference on Computer Vision (ECCV) . Springer, 2024, pp. 148–166

  74. [74]

    Judging llm-as-a-judge with mt-bench and chatbot arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. Xing et al. , “Judging llm-as-a-judge with mt-bench and chatbot arena,” Advances in Neural Information Processing Systems (NIPS), vol. 36, pp. 46 595–46 623, 2023

  75. [75]

    Aligning with human judgement: The role of pairwise preference in large language model evaluators,

    Y . Liu, H. Zhou, Z. Guo, E. Shareghi, I. Vuli´c, A. Korhonen, and N. Col- lier, “Aligning with human judgement: The role of pairwise preference in large language model evaluators,” arXiv preprint arXiv:2403.16950 , 2024

  76. [76]

    Generative judge for evaluating alignment,

    J. Li, S. Sun, W. Yuan, R.-Z. Fan, H. Zhao, and P. Liu, “Generative judge for evaluating alignment,” arXiv preprint arXiv:2310.05470 , 2023

  77. [77]

    Judgelm: Fine-tuned large language models are scalable judges,

    L. Zhu, X. Wang, and X. Wang, “Judgelm: Fine-tuned large language models are scalable judges,” arXiv preprint arXiv:2310.17631 , 2023. Muhammad Kashif Ali received his B.Eng degree in electrical engineering from National University of Sciences and Technology (NUST), Pakistan, and an MS-PhD degree in Computer Science from Hanyang University, Seoul, Korea, ...

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.