Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

BVINet: Unlocking Blind Video Inpainting with Zero Annotations

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read A blind video inpainting network that locates and fills corrupted regions without mask annotations.

desk verdict A plausible and well-executed blind video inpainting system with strong in-distribution results, but the 'first' claim needs qualification and the evaluation does not yet cover seamlessly composited corruption. read the letter →

arxiv 2502.01181 v1 pith:7FBMSA4H submitted 2025-02-03 cs.CV

Zhiliang Wu , Kerui Chen , Kun Li , Hehe Fan , Yi Yang This is my paper · ORCID
classification cs.CV
keywords blindvideoinpaintingmaskpredictionsemanticdiscontinuitytemporalconsistencywaveletsparsetransformerlosscompletionbulletremoval
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

This paper tries to remove the mask-annotation requirement from video inpainting by defining a blind setting in which a network learns the mapping from a corrupted video straight to a completed one. It claims that a mask prediction network can locate corrupted regions by detecting semantic-discontinuous areas and refining them with temporal consistency, while a video completion network fills those regions using only valid context. A consistency loss ties the two subnetworks together by requiring the binarized difference between input and output to match both the predicted and ground-truth masks. If the claim holds, the practical cost of video inpainting drops substantially, and the reported numbers put blind performance close to non-blind baselines on YouTube-VOS and DAVIS.

What carries the argument

The load-bearing mechanism is the consistency loss that couples mask prediction with video completion. It turns the completion network's own change into a self-supervised mask target, so the two subnetworks constrain each other instead of being trained independently. The completion network is a wavelet sparse transformer that splits query, key, and value features into low- and high-frequency components with a discrete wavelet transform, runs dense and sparse self-attention branches on the clean low frequencies, masks attention inside predicted corrupted regions, and recombines the branches with an inverse discrete wavelet transform. The mask network uses wavelet downsampling for noise robustness and a sequence-to-sequence transformer to refine per-frame predictions over time.

What would settle it

Construct corrupted videos whose damaged regions are filled with semantically coherent, boundary-smoothed natural texture matched to the scene and kept temporally stable, then run BVINet and measure the IoU of its predicted masks; if the IoU collapses toward zero while non-blind inpainting with the true masks still yields high PSNR, the central blind-mapping claim is falsified.

Watch

Extended reading notes

Core claim

The paper proposes BVINet, an end-to-end network for blind video inpainting, and reports that it performs comparably to non-blind methods that receive ground-truth masks. The central claim is that the mapping can be learned without any corrupted-region annotation at inference: MPNet predicts binary masks from single frames via semantic-discontinuity detection, then refines them over time; VCNet completes the video using the predicted masks, and the two are jointly trained. The consistency loss in Eq. (8), $L_c = \| m_i^l - B(\hat{y}_i - x_i) \|_1 + \| m_i - B(\hat{y}_i - x_i) \|_1$, encodes the identity that the binarized residual between completed and corrupted frames should be the mask of what was changed. On the tested splits the blind model reaches PSNR 30.528 on YouTube-VOS and 30.961 on DAVIS, above all blind baselines and close to the best non-blind methods.

Load-bearing premise

The method assumes every corrupted region is visually detectable as a semantic discontinuity within a single frame, so seamlessly blended or temporally persistent corruption leaves the mask predictor with no signal to separate it from valid content.

Editorial extensions

If this is right

  • Mask annotation can be dropped from the video inpainting pipeline; a single model both finds and fills corruption at inference.
  • Blind performance on YouTube-VOS and DAVIS is comparable to non-blind methods, with the blind model beating all blind baselines by a large margin.
  • The model generalizes to corruption patterns unseen in training, such as Gaussian noise and constant-color fills, suggesting it detects visual inconsistency rather than memorizing dataset patterns.
  • Real-world tasks such as bullet removal can be automated without per-frame mask annotation.
  • The completion network is also efficient, with the lowest FLOPs and inference time among the compared transformer-based inpainting networks.

Reading between the lines

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

  • Because mask prediction relies on single-frame semantic discontinuity, the method would likely fail on seamlessly composited or semantically coherent corruption; testing with boundary-free blended edits would reveal the limit.
  • The second term of the consistency loss uses ground-truth masks during training, so 'zero annotations' holds at inference; a fully annotation-free variant that replaces the ground-truth term with the model's own mask would be the pure form of the claim.
  • The per-frame-first design means corruption missed in short-term prediction cannot be recovered by temporal refinement, only cleaned; very large, slow, or textureless corrupted regions are the natural stress test.
  • If the approach holds, it suggests a general recipe: pair any region detector with any completer and couple them through a change-consistency loss, which could extend to other restoration tasks where damage is locally visible.
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 / 6 minor

Summary. The paper introduces BVINet, an end-to-end network for blind video inpainting that jointly predicts corruption masks and completes corrupted videos, without requiring masks at inference time. The method comprises a mask prediction network (MPNet) with short-term and long-term refinement modules, and a video completion network (VCNet) based on a wavelet sparse transformer. A consistency loss is proposed to couple mask prediction and completion. The authors construct a synthetic dataset by filling free-form strokes with natural image patches and applying iterative Gaussian smoothing to avoid edge priors, plus 1,250 real bullet-removal clips. Experiments on YouTube-VOS and DAVIS report that BVINet outperforms six blind baselines and achieves performance comparable to several non-blind methods, and qualitative results on bullet removal are shown. The paper also reports ablations on MPNet components, sparse attention, and the consistency loss.

Significance. If the claims hold, BVINet would be the first deep blind video inpainting method, with potential practical value for removing scratches, watermarks, and bullets without manual mask annotation at inference. The paper contributes a new task formulation, a dataset, and a reasonably thorough experimental comparison. The method's central idea—coupling mask prediction with completion in an end-to-end trainable framework—is sensible and the reported quantitative gains over blind baselines are substantial. However, the title's 'Zero Annotations' claim is not supported by the training procedure, which uses ground-truth masks for supervision, and the evaluation does not stress-test the method on truly unseen or seamlessly composited corruption, so the general 'blind video inpainting' claim is currently over-scoped.

major comments (3)
  1. [Sec. 3.3, Eq. (8), Sec. 3.4, Title] The consistency loss in Eq. (8) explicitly uses the ground-truth mask m_i, and the total loss in Eq. (9) includes the mask prediction loss Lm (described as 'mask prediction loss' in Sec. 3.4), which requires per-frame binary mask annotations during training. The title 'Zero Annotations' and the abstract's claim that the method 'eliminates the need of corrupted region annotations' are therefore misleading: BVINet is annotation-free only at inference. The manuscript should clearly state that training requires mask labels (e.g., on synthetic data) and rephrase the contribution as 'mask-free at inference' rather than 'zero annotations'.
  2. [Sec. 4.1, Sec. 4.3, Table 2] The test sets are corrupted 'using a similar manner to the training set' (Sec. 4.1), with the same free-form strokes, natural-patch fills, and iterative Gaussian smoothing. Under this protocol, the MPNet reaches an IOU of 0.9017 (Table 2), but this measures the network's ability to recognize the specific corruption distribution used in training, not a general ability to localize arbitrary unseen corruptions. The central blind-inpainting claim would be better supported by evaluating on held-out corruption models—for example, seamlessly composited patches without boundary smoothing, semi-transparent overlays, or temporally stable corruptions—and by reporting mask localization accuracy on those cases. Without such experiments, the method's applicability to real-world corruptions (e.g., watermarks that blend into textured regions) remains unsubstantiated.
  3. [Sec. 4.4, Fig. 4] The real-world bullet removal evaluation is qualitative only, with no quantitative metric such as detection accuracy, false-positive rate, or comparison against a mask-based method on the same real clips. Since the paper motivates the task by practical applications (Sec. 1), this evidence is too limited to support the stated generalization claim. Adding even a small quantitative benchmark on real data (e.g., manually annotated bullet regions scored by IOU, or a user study) would materially strengthen the claim.
minor comments (6)
  1. [Sec. 4.2, Table 1] The six 'blind video inpainting baselines' are not independent blind methods; they are the authors' MPNet combined with existing completion networks (MPNet+CPVINet, MPNet+STTN, etc.). This should be stated explicitly, and the contribution of MPNet versus the completion backbone should be disentangled in the analysis.
  2. [Table 1, Sec. 4.3] The comparison with non-blind methods is informative but the gap should be discussed more carefully: on YouTube-VOS, the best non-blind method (WaveFormer) achieves PSNR 33.264, 2.7 dB higher than BVINet's 30.528, while BVINet is closer to E2FGVI (30.064). The claim 'comparable performance to non-blind methods' is fair only against a subset of baselines, and this nuance should be acknowledged.
  3. [Sec. 3.3, Eqs. (6)-(8)] The notation is confusing because m_i appears both as the ground-truth mask and in the binarization relationship (Eq. (6)), while m_i^l is the predicted mask. Please define both clearly and perhaps use different symbols (e.g., m_i^gt) to avoid ambiguity.
  4. [Sec. 3.4] The loss hyperparameters (lambda_m=3, lambda_v=5, lambda_c=0.02) are stated to be determined by grid search, but no search range or sensitivity analysis is provided. A brief sensitivity table or a statement that results are stable across a range of values would increase confidence.
  5. [References] Several references contain formatting errors; for example, reference [11] has garbled author text ('Zhang Haotian, Mai Long, Wang Hailin, JinZha ando wen, and Ning Xu; John Collomosse'). The reference list should be cleaned before publication.
  6. [General] The paper reports single-run results without error bars or significance tests. Since the claimed improvements over blind baselines are large, this is not a block issue, but reporting variance across multiple seeds or a paired test would strengthen the empirical claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central claims are supported by external benchmark evaluation and independent ablations; the consistency loss is a training objective, not a derived prediction.

full rationale

The paper proposes a two-network system (MPNet for mask prediction, VCNet for completion) trained end-to-end with mask, completion, and consistency losses (Eq. 9). The consistency loss (Eq. 8) enforces agreement between the predicted mask and binarized frame difference during training; this is a regularizer, and the reported results are evaluated with external metrics (PSNR, SSIM, LPIPS, Ewarp) on YouTube-VOS and DAVIS against a ground-truth completed video, so the headline numbers do not reduce to the training objective. The 'zero annotation' claim refers to inference: masks are synthesized automatically with free-form strokes and iterative Gaussian smoothing (Sec. 4.1), and at test time only corrupted video is fed to the model. While [48] (WaveFormer) is prior work by the same authors and motivates the wavelet transformer design in Sec. 3.2.2, the article does not rely on it as a theorem or as evidence for the blind-inpainting claim; WaveFormer is also included as an external baseline and outperformed by the proposed VCNet. The scope limitation that the method targets visually detectable semantic-discontinuous corruption (Sec. 1 and Eq. 1) and that the synthetic test data use the same corruption model as training is a generalization/correctness concern, not a circularity in the derivation. No equation in the paper is equivalent by construction to a fitted parameter that is later called a prediction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on four domain assumptions: corrupted regions are visually discontinuous, temporal neighbors share mask positions, synthetic free-form corruption with smoothed edges is representative of real corruption, and in-distribution synthetic testing is a valid proxy for real-world blind inpainting. The only explicitly hand-set free parameters reported are the three grid-searched loss weights; many architecture hyperparameters are deferred to supplementary material.

free parameters (3)
  • lambda_m = 3
    Mask prediction loss weight in Eq. (9), tuned by grid search (Sec. 3.4).
  • lambda_v = 5
    Video completion loss weight in Eq. (9), tuned by grid search.
  • lambda_c = 0.02
    Consistency loss weight in Eq. (9), tuned by grid search; it is small relative to the other weights.
assumptions (4)
  • domain assumption Corrupted regions are detectable as semantic-discontinuous regions within a frame.
    Short-term prediction module (Sec. 3.2.1) assumes appearance inconsistency is sufficient to localize corruption; this fails for seamless compositing.
  • domain assumption Corrupted region positions are similar across adjacent frames.
    Long-term refinement uses temporal neighbors because 'the positions of corrupted regions in adjacent frames are similar [10,13]' (Sec. 3.2.1); slowly moving or appearing/disappearing corruptions violate this.
  • domain assumption Free-form strokes plus iterative Gaussian smoothing produce realistic corruptions without learnable border or content priors.
    Dataset construction (Sec. 4.1) relies on this to make mask prediction learn semantics instead of dataset artifacts.
  • domain assumption Synthetic test corruptions generated similarly to training are a valid proxy for real-world blind inpainting.
    Testing dataset (Sec. 4.1) evaluates in-distribution; real-world bullet removal is qualitative only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BVINet: Unlocking Blind Video Inpainting with Zero Annotations." pith.science (2026). https://pith.science/paper/7FBMSA4H

@misc{pith2026250201181,
  author       = {Pith},
  title        = {Pith review of: BVINet: Unlocking Blind Video Inpainting with Zero Annotations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7FBMSA4H}},
  note         = {Machine review of arXiv:2502.01181}
}
read the original abstract

Video inpainting aims to fill in corrupted regions of the video with plausible contents. Existing methods generally assume that the locations of corrupted regions are known, focusing primarily on the "how to inpaint". This reliance necessitates manual annotation of the corrupted regions using binary masks to indicate "whereto inpaint". However, the annotation of these masks is labor-intensive and expensive, limiting the practicality of current methods. In this paper, we expect to relax this assumption by defining a new blind video inpainting setting, enabling the networks to learn the mapping from corrupted video to inpainted result directly, eliminating the need of corrupted region annotations. Specifically, we propose an end-to-end blind video inpainting network (BVINet) to address both "where to inpaint" and "how to inpaint" simultaneously. On the one hand, BVINet can predict the masks of corrupted regions by detecting semantic-discontinuous regions of the frame and utilizing temporal consistency prior of the video. On the other hand, the predicted masks are incorporated into the BVINet, allowing it to capture valid context information from uncorrupted regions to fill in corrupted ones. Besides, we introduce a consistency loss to regularize the training parameters of BVINet. In this way, mask prediction and video completion mutually constrain each other, thereby maximizing the overall performance of the trained model. Furthermore, we customize a dataset consisting of synthetic corrupted videos, real-world corrupted videos, and their corresponding completed videos. This dataset serves as a valuable resource for advancing blind video inpainting research. Extensive experimental results demonstrate the effectiveness and superiority of our method.

Figures

Figures reproduced from arXiv: 2502.01181 by the authors.

Figure 1
Figure 1. Fig.(a) shows the general pipeline of existing non-blind [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overview of the proposed blind video inpainting framework. Our framework are composed of a mask prediction network (MPNet) and a video completion network (VCNet). The former aims to predict the masks of corrupted regions by detecting semantic￾discontinuous regions of the frame and utilizing temporal consistency prior of the video, while the latter perceive valid context information from uncorrupted regions using… view at source ↗
Figure 5
Figure 5. Example of corrupted regions segmentation. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative results compared with OGNet [ [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 7
Figure 7. Figure 7: (a) Visual evaluations with random masks filled with different content (first four examples). (b) Visual comparison of different [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Exp2VLA: Enabling Vision-Language-Action for Drone Navigation from Expert Demonstrations

    cs.RO 2026-07 conditional novelty 4.0 of 10

    Expert demonstrations distilled into LeRobot-format data let fine-tuned mid-scale VLAs navigate multi-object drone scenes from language, with partial zero-shot color-shape generalization in simulation and SITL.

Reference graph

Works this paper leans on

61 extracted references · 59 canonical work pages · cited by 1 Pith paper

  1. [16]

    Deep blind video decaptioning by temporal aggre- gation and recurrence

    Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Deep blind video decaptioning by temporal aggre- gation and recurrence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4263–4272, 2019. 2, 6

    work page 2019

  2. [1]

    Restoration of analog videos using swin-unet

    Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Restoration of analog videos using swin-unet. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM) , pages 6985–6987,

    work page

  3. [2]

    Devit: Deformed vision transformers in video inpaint- ing

    Jiayin Cai, Changlin Li, Xin Tao, Chun Yuan, and Yu-Wing Tai. Devit: Deformed vision transformers in video inpaint- ing. In Proceedings of the 30th ACM International Confer- ence on Multimedia (ACMMM), pages 779–789, 2022. 1, 2, 3, 4

    work page 2022

  4. [3]

    Blind inpainting using the fully convolutional neural network

    Nian Cai, Zhenghang Su, Zhineng Lin, Han Wang, Zhijing Yang, and Bingo Wing-Kuen Ling. Blind inpainting using the fully convolutional neural network. The Visual Com- puter, 33:249–261, 2017. 3

    work page 2017

  5. [4]

    Learning prior feature and attention enhanced image inpainting

    Chenjie Cao, Qiaole Dong, and Yanwei Fu. Learning prior feature and attention enhanced image inpainting. In Pro- ceedings of the European Conference on Computer Vision (ECCV), pages 306–322, 2022. 3

    work page 2022

  6. [5]

    Free-form video inpainting with 3d gated convolution and temporal patchgan

    Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9066–9075, 2019. 2, 6

    work page 2019

  7. [6]

    Learn- ing a sparse transformer network for effective image derain- ing

    Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learn- ing a sparse transformer network for effective image derain- ing. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 5896– 5905, 2023. 5

    work page 2023

  8. [7]

    Infusion: Internal diffusion for video inpaint- ing

    Nicolas Cherel, Andr ´es Almansa, Yann Gousseau, and Alas- dair Newson. Infusion: Internal diffusion for video inpaint- ing. arXiv preprint arXiv:2311.01090, 2023. 1

    arXiv 2023

Show all 61 references
  1. [8]

    Incremental transformer structure enhanced image inpainting with mask- ing positional encoding

    Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with mask- ing positional encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11358–11368, 2022. 3

    2022

  2. [9]

    Flow-edge guided video completion

    Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 713–729, 2020. 2, 6

    2020

  3. [10]

    Pyramid constrained self- attention network for fast video salient object detection

    Yuchao Gu, Lijuan Wang, Ziqin Wang, Yun Liu, Ming- Ming Cheng, and Shao-Ping Lu. Pyramid constrained self- attention network for fast video salient object detection. In Proceedings of the AAAI conference on artificial intelligence (AAAI), pages 10869–10876, 2020. 4

    2020

  4. [11]

    An internal learning ap- proach to video inpainting

    Zhang Haotian, Mai Long, Wang Hailin, JinZha ando wen, and Ning Xu; John Collomosse. An internal learning ap- proach to video inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2720–2729, 2019. 7

    2019

  5. [12]

    Temporally coherent completion of dynamic video

    Jiabin Huang, Sing Bing Kang, Narendra Ahuja, and Jo- hannes Kopf. Temporally coherent completion of dynamic video. ACM Transactions on Grapics (TOG), 35(6):196.1– 196.11, 2016. 6

    2016

  6. [13]

    Progressively nor- malized self-attention network for video polyp segmentation

    Ge-Peng Ji, Yu-Cheng Chou, Deng-Ping Fan, Geng Chen, Huazhu Fu, Debesh Jha, and Ling Shao. Progressively nor- malized self-attention network for video polyp segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI) , pages 142–

  7. [14]

    Error compensation framework for flow-guided video inpainting

    Jaeyeon Kang, Seoung Wug Oh, and Seon Joo Kim. Error compensation framework for flow-guided video inpainting. In Proceedings of the European Conference on Computer Vi- sion (ECCV), pages 375–390, 2022. 2, 6

    2022

  8. [15]

    Occlusion-aware video object inpainting

    Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Occlusion-aware video object inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14468–14478, 2021. 2

    2021

  9. [17]

    Recurrent temporal aggregation framework for deep video inpainting

    Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Recurrent temporal aggregation framework for deep video inpainting. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 42(5):1038–1052, 2020. 6

    2020

  10. [18]

    Learning blind video temporal consistency

    Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European Con- ference on Computer Vision (ECCV), pages 179–195, 2018. 7

    2018

  11. [19]

    Copy-and-paste networks for deep video inpainting

    Sungho Lee, Seoung Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4413–4421, 2019. 2, 6, 7

    2019

  12. [20]

    Gong, Jianzhong Qi, Rui Zhang, Dacheng Tao, and R

    Ang Li, Shanshan Zhao, Xingjun Ma, M. Gong, Jianzhong Qi, Rui Zhang, Dacheng Tao, and R. Kotagiri. Short-term and long-term context aggregation network for video inpaint- ing. In Proceedings of the European Conference on Com- puter Vision (ECCV), pages 728–743, 2020. 2, 6

    2020

  13. [21]

    Decontam- ination transformer for blind image inpainting

    Chun-Yi Li, Yen-Yu Lin, and Wei-Chen Chiu. Decontam- ination transformer for blind image inpainting. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 1–5. IEEE,

    2023

  14. [22]

    Wavelet integrated cnns for noise-robust image classification

    Qiufu Li, Linlin Shen, Sheng Guo, and Zhihui Lai. Wavelet integrated cnns for noise-robust image classification. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7243–7252, 2020. 4

    2020

  15. [23]

    Wavec- net: Wavelet integrated cnns to suppress aliasing effect for noise-robust image classification

    Qiufu Li, Linlin Shen, Sheng Guo, and Zhihui Lai. Wavec- net: Wavelet integrated cnns to suppress aliasing effect for noise-robust image classification. IEEE Transactions on Im- age Processing (TIP), 30:7074–7089, 2021. 4

    2021

  16. [24]

    Misf: Multi-level interactive siamese filter- ing for high-fidelity image inpainting

    Xiaoguang Li, Qing Guo, Di Lin, Ping Li, Wei Feng, and Song Wang. Misf: Multi-level interactive siamese filter- ing for high-fidelity image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1869–1878, 2022. 3

    2022

  17. [25]

    Semid: Blind image inpainting with semantic inconsistency detection

    Xin Li, Zhikuan Wang, Chenglizhao Chen, Chunfeng Tao, Yuanbo Qiu, Junde Liu, and Baile Sun. Semid: Blind image inpainting with semantic inconsistency detection. Tsinghua Science and Technology, 29(4):1053–1068, 2024. 3 9

    2024

  18. [26]

    Towards an end-to-end framework for flow-guided video inpainting

    Zhen Li, Cheng-Ze Lu, Jianhua Qin, Chun-Le Guo, and Ming-Ming Cheng. Towards an end-to-end framework for flow-guided video inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17562–17571, 2022. 2, 6, 7, 8

    2022

  19. [27]

    Reduce information loss in transformers for pluralistic image inpainting

    Qiankun Liu, Zhentao Tan, Dongdong Chen, Qi Chu, Xiyang Dai, Yinpeng Chen, Mengchen Liu, Lu Yuan, and Nenghai Yu. Reduce information loss in transformers for pluralistic image inpainting. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (C...

    2022

  20. [28]

    Fuseformer: Fusing fine-grained information in transformers for video inpainting

    Rui Liu, Hanming Deng, Yangyi Huang, Xiaoyu Shi, Lewei Lu, Wenxiu Sun, Xiaogang Wang, Jifeng Dai, and Hong- sheng Li. Fuseformer: Fusing fine-grained information in transformers for video inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (I...

    2021

  21. [29]

    Temporal group fusion network for deep video inpainting.IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) , 32 (6):3539–3551, 2022

    Ruixin Liu, Bairong Li, and Yuesheng Zhu. Temporal group fusion network for deep video inpainting.IEEE Transactions on Circuits and Systems for Video Technology (TCSVT) , 32 (6):3539–3551, 2022. 2

    2022

  22. [30]

    Deep blind image inpainting

    Yang Liu, Jinshan Pan, and Zhixun Su. Deep blind image inpainting. arXiv preprint arXiv:1712.09078, 2017. 3

    2017 arXiv

  23. [31]

    A theory for multiresolution signal de- composition: the wavelet representation

    Stephane G Mallat. A theory for multiresolution signal de- composition: the wavelet representation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) , 11 (4):674–693, 1989. 5

    1989

  24. [32]

    Perazzi, J

    F. Perazzi, J. Pont-Tuset, B. Mcwilliams, L. Van Gool, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 724–732, 2016. 6

    2016

  25. [33]

    Blind image inpainting via omni-dimensional gated attention and wavelet queries

    Shruti S Phutke, Ashutosh Kulkarni, Santosh Kumar Vip- parthi, and Subrahmanyam Murala. Blind image inpainting via omni-dimensional gated attention and wavelet queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops , page...

    2023

  26. [34]

    Dlformer: Discrete latent transformer for video inpainting

    Jingjing Ren, Qingqing Zheng, Yuanyuan Zhao, Xuemiao Xu, and Chen Li. Dlformer: Discrete latent transformer for video inpainting. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , pages 3511–3520, 2022. 1, 2, 3, 4

    2022

  27. [35]

    Blind im- age inpainting with sparse directional filter dictionaries for lightweight cnns

    Jenny Schmalfuss, Erik Scheurer, Heng Zhao, Nikolaos Karantzas, Andr ´es Bruhn, and Demetrio Labate. Blind im- age inpainting with sparse directional filter dictionaries for lightweight cnns. Journal of Mathematical Imaging and Vi- sion (JMIV), pages 1–17, 2022. 2

    2022

  28. [36]

    Onion-peel networks for deep video completion

    Oh Seoung, Wug, Lee Sungho, Lee Joon-Young, and Kim Seon, Joo. Onion-peel networks for deep video completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4402–4411, 2019. 2, 7

    2019

  29. [37]

    Transinpaint: Transformer-based image inpainting with context adaptation

    Pourya Shamsolmoali, Masoumeh Zareapoor, and Eric Granger. Transinpaint: Transformer-based image inpainting with context adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 849–858, 2023. 3

    2023

  30. [38]

    Spatial-temporal residual aggrega- tion for high resolution video inpainting

    Vishnu Sanjay Ramiya Srinivasan, Rui Ma, Qiang Tang, Zili Yi, and Zhan Xu. Spatial-temporal residual aggrega- tion for high resolution video inpainting. arXiv preprint arXiv:2111.03574, 2021. 2

    2021 arXiv

  31. [39]

    Self-prior guided pixel adversar- ial networks for blind image inpainting

    Juan Wang, Chunfeng Yuan, Bing Li, Ying Deng, Weiming Hu, and Stephen Maybank. Self-prior guided pixel adversar- ial networks for blind image inpainting. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) , 45 (10):12377–12393, 2023. 2, 3

    2023

  32. [40]

    Text- video completion networks with motion compensation and attention aggregation

    Jianan Wang, Zhiliang Wu, Hanyu Xuan, and Yan Yan. Text- video completion networks with motion compensation and attention aggregation. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP) , pages 2990–2994, 2024. 2

    2024

  33. [41]

    Image inpainting via generative multi-column convolu- tional neural networks

    Yi Wang, Xin Tao, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Image inpainting via generative multi-column convolu- tional neural networks. In Proceedings of the Advances in neural information processing systems (NIPS), 2018. 2, 6

    2018

  34. [42]

    Vcnet: A robust approach to blind image inpainting

    Yi Wang, Ying-Cong Chen, Xin Tao, and Jiaya Jia. Vcnet: A robust approach to blind image inpainting. InProceedings of the European Conference on Computer Vision (ECCV) , pages 752–768, 2020. 2, 3

    2020

  35. [43]

    Dapc-net: Deformable alignment and pyramid context completion networks for video inpainting

    Zhiliang Wu, Kang Zhang, Hanyu Xuan, Jian Yang, and Yan Yan. Dapc-net: Deformable alignment and pyramid context completion networks for video inpainting. IEEE Signal Pro- cessing Letters (SPL), 28:1145–1149, 2021. 7

    2021

  36. [44]

    Deep stereo video inpainting

    Zhiliang Wu, Changchang Sun, Hanyu Xuan, and Yan Yan. Deep stereo video inpainting. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , pages 5693–5702,

  37. [45]

    Divide-and-conquer completion network for video inpainting

    Zhiliang Wu, Changchang Sun, Hanyu Xuan, Kang Zhang, and Yan Yan. Divide-and-conquer completion network for video inpainting. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 33(6):2753–2766, 2023. 2, 7

    2023

  38. [46]

    Semi-supervised video inpaint- ing with cycle consistency constraints

    Zhiliang Wu, Hanyu Xuan, Changchang Sun, Weili Guan, Kang Zhang, and Yan Yan. Semi-supervised video inpaint- ing with cycle consistency constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22586–22595, 2023. 1

    2023

  39. [47]

    Flow-guided deformable alignment network with self-supervision for video inpainting

    Zhiliang Wu, Kang Zhang, Changchang Sun, Hanyu Xuan, and Yan Yan. Flow-guided deformable alignment network with self-supervision for video inpainting. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP) , pages 1–5, 2023. 2

    2023

  40. [48]

    Waveformer: Wavelet transformer for noise- robust video inpainting

    Zhiliang Wu, Changchang Sun, Hanyu Xuan, Gaowen Liu, and Yan Yan. Waveformer: Wavelet transformer for noise- robust video inpainting. In Proc. AAAI Conf. Artif. Intell. (AAAI), pages 6180–6188, 2024. 5, 6, 7, 8

    2024

  41. [49]

    Youtube-vos: Sequence-to-sequence video object segmentation

    Ning Xu, Linjie Yang, Yuchen Fan, Jianchao Yang, Dingcheng Yue, Yuchen Liang, Brian Price, Scott Cohen, and Thomas Huang. Youtube-vos: Sequence-to-sequence video object segmentation. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 603–619,

  42. [50]

    Deep flow-guided video inpainting

    Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3723–3732, 2019. 2, 6 10

    2019

  43. [51]

    Deficiency- aware masked transformer for video inpainting

    Yongsheng Yu, Heng Fan, and Libo Zhang. Deficiency- aware masked transformer for video inpainting. arXiv preprint arXiv:2307.08629, 2023. 1

    2023 arXiv

  44. [52]

    Learning joint spatial-temporal transformations for video inpainting

    Yanhong Zeng, Jianlong Fu, and Hongyang Chao. Learning joint spatial-temporal transformations for video inpainting. In Proceedings of the European Conference on Computer Vi- sion (ECCV), pages 3723–3732, 2020. 1, 2, 3, 4, 6, 7, 8

    2020

  45. [53]

    Inertia-guided flow completion and style fusion for video inpainting

    Kaidong Zhang, Jingjing Fu, and Dong Liu. Inertia-guided flow completion and style fusion for video inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR) , pages 5982–5991,

  46. [54]

    Flow-guided transformer for video inpainting

    Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. InProceedings of the Euro- pean Conference on Computer Vision (ECCV), pages 74–90,

  47. [55]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586–595, 2018. 7

    2018

  48. [56]

    Pfta-net: Pro- gressive feature alignment and temporal attention fusion net- works for video inpainting

    Yanni Zhang, Zhiliang Wu, and Yan Yan. Pfta-net: Pro- gressive feature alignment and temporal attention fusion net- works for video inpainting. In Proc. IEEE Int. Conf. Image Process. (ICIP), pages 191–195, 2023. 2

    2023

  49. [57]

    Transcnn-hae: Transformer-cnn hybrid autoencoder for blind image inpainting

    Haoru Zhao, Zhaorui Gu, Bing Zheng, and Haiyong Zheng. Transcnn-hae: Transformer-cnn hybrid autoencoder for blind image inpainting. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM) , pages 6813–6821, 2022. 3

    2022

  50. [58]

    Ciri: Curricular inactivation for residue- aware one-shot video inpainting

    Weiying Zheng, Cheng Xu, Xuemiao Xu, Wenxi Liu, and Shengfeng He. Ciri: Curricular inactivation for residue- aware one-shot video inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13012–13022, 2023. 1

    2023

  51. [59]

    Propainter: Improving propagation and transformer for video inpainting

    Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. Propainter: Improving propagation and transformer for video inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10477–10486, 2023. 6

    2023

  52. [60]

    Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration

    Shihao Zhou, Duosheng Chen, Jinshan Pan, Jinglei Shi, and Jufeng Yang. Adapt or perish: Adaptive sparse transformer with attentive feature refinement for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR) , pages 2952–2963,

  53. [61]

    Pro- gressive temporal feature alignment network for video in- painting

    Xueyan Zou, Linjie Yang, Ding Liu, and Yong Jae Lee. Pro- gressive temporal feature alignment network for video in- painting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 16448–16457, 2021. 2 11

    2021

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.