Pith. sign in

REVIEW 4 major objections 7 minor 74 references

Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Adaptive stacking filter makes video deraining work on real rain

desk verdict A credible incremental architecture paper with a genuinely useful benchmark, let down by an ill-defined semi-supervised loss that must be fixed before the real-world claims hold. read the letter →

arxiv 2505.16811 v1 pith:2LZORV7O submitted 2025-05-22 cs.CV

classification cs.CV
keywords videoderainingstate-spacemodeldynamicstackingfiltersemi-supervisedlearningtemporalmedianopticalflowrainstreakremovalrestoration
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

VDMamba, the method this paper presents, is a dual-branch spatio-temporal state-space model for video deraining that scans spatial and temporal features separately. The paper claims it reaches the highest PSNR and SSIM on RainSynLight25, RainSynComplex25, and NTURain, and that it is the only compared method that removes all rain streaks in real-world footage. The key technical idea is a dynamic stacking filter that learns, per pixel, to aggregate aligned neighboring frames in a way that approximates the mean, median, max, and min filters. To generalize to real rain without paired data, the model trains semi-supervised with a masked median stacking loss that uses temporal-median patches of warped frames as pseudo-clean targets. The paper also introduces RVDT, a 57-clip real-world benchmark with detection and tracking annotations, and reports that its derained videos improve downstream detection and tracking.

What carries the argument

The central mechanism is the dynamic stacking filter (DSF), a pixel-wise differentiable aggregation defined by Eq. (16), which uses softmax-weighted averaging with learnable scalar parameters a and b to interpolate among four statistical filters: mean, median, max, and min. A second load-bearing component is the masked median stacking loss (Eqs. 19-21), which generates pseudo-clean patches by taking the temporal median of warped frames and keeps only patches whose pixels are largely unchanged (threshold θ=80%, slack δ=0.1) to respect the sparsity prior of rain and reject moving objects. The dual-branch architecture itself, with spatial state-space layers (S3ML) and temporal state-space layers (TSML) in a U-Net, is what carries feature extraction and multi-frame fusion.

What would settle it

Train VDMamba with the median stacking loss disabled and compare PSNR/SSIM on a real-world rainy test set such as NTURain; the paper's claim predicts a clear drop, while a null result would show the semi-supervised component is not responsible for the real-world gains. A complementary check is to feed a real video with a large, slow-moving object and see whether the restored output erases it, which would show the mask is letting contaminated patches through.

Watch

Extended reading notes

Core claim

The central claim is that VDMamba outperforms existing video deraining methods on synthetic benchmarks and, more importantly, is the only method evaluated that successfully removes all rain streaks from real-world videos. The paper derives a dynamic stacking filter, expressed as a softmax-weighted combination of candidate pixels with learnable coefficients a and b, which in the limit reproduces the mean, median, max, and min filters; this filter is inserted into the temporal state-space layer to fuse aligned neighboring-frame features. For real-world generalization, the paper proposes a median stacking loss that builds pseudo-clean patches from the temporal median of optical-flow-warped frames, masked to keep only sub-patches where at least 80% of pixels are nearly unchanged relative to the central frame, thereby exploiting the sparsity of rain while excluding moving objects and occlusions. It then reports state-of-the-art PSNR/SSIM on three benchmarks and improvements in object detection (YOLO-v3, MEGA) and tracking (GTR) on the new RVDT benchmark.

Load-bearing premise

The pseudo-label mechanism assumes that the temporal median of optical-flow-warped frames, masked to keep only sub-patches where at least 80% of pixels are nearly unchanged, yields clean training targets that reliably exclude occlusions and moving objects.

Editorial extensions

If this is right

  • If the claims hold, deraining models can be trained on unlabeled real-world rainy videos via the median stacking loss, breaking the dependence on synthetic paired data.
  • The same network that estimates multi-frame optical flow also performs deraining, and the paper reports this single-forward-pass design reaches real-time speeds.
  • The dynamic stacking filter, by adaptively switching among mean, median, max and min behavior, may serve as a general multi-frame fusion module in other video restoration pipelines, such as denoising or deblurring.
  • The RVDT results indicate that deraining can act as an effective pre-processing step for object detection and tracking in rainy scenes, improving mAP and MOTA over direct inference on raw rain.

Reading between the lines

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

  • Inference: If the mask threshold θ correctly rejects moving objects, then the method's real-world performance should degrade gracefully as scene motion increases; a stress test with large, slow-moving foreground objects occupying most of a sub-patch would reveal whether the sparsity prior is doing the heavy lifting.
  • Inference: The differentiable softmax approximation of the median in Eq. (15) is a reusable primitive for any learning task that needs robust order statistics, not just deraining; it could serve as a differentiable median pooling layer in other architectures.
  • Inference: Because the semi-supervised loss depends on accurate optical-flow warping, the method likely inherits flow estimation failures under fast camera motion or large displacements; the degradation-free flow transfer reduces but does not remove this dependency, which the paper does not quantify.
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

4 major / 7 minor

Summary. The paper proposes VDMamba, a dual-branch spatio-temporal state-space model for video deraining. The spatial branch uses S3ML for single-frame features and the temporal branch uses TSML for multi-frame fusion, while a dynamic stacking filter (DSF) is introduced to aggregate neighboring pixels by adaptively approximating the mean, median, max, and min statistics. For real-world generalization, the authors propose a semi-supervised median stacking loss that generates pseudo-clean patches from aligned frames, and they introduce RVDT, a new benchmark for object detection and tracking in rainy videos. Experiments on RainSynLight25, RainSynComplex25, NTURain, and RVDT report state-of-the-art or competitive PSNR/SSIM and downstream detection/tracking metrics, along with qualitative comparisons on synthetic and real-world footage.

Significance. If the results are reproducible, the work would be a useful contribution to video deraining. The RVDT benchmark with detection and tracking annotations is a practical asset for evaluating deraining methods in downstream tasks, and the systematic comparison against nine baselines is a concrete strength. The DSF idea of learning a pixel-adaptive statistical filter is interesting and the ablations support the value of the proposed components. However, the central real-world generalization claim depends on a semi-supervised loss that is not fully specified as written, and there are mathematical presentation issues in the DSF derivation. The headline quantitative claims would also be more convincing with variance estimates, because several margins over the strongest baseline are small.

major comments (4)
  1. [§3.3, Eq. (17) and Eq. (23)] The unpaired-data branch of the total loss is not computable as written. Eq. (17) defines Lspa = ||B_c − \tilde{B}_c||_1 with B_c denoting the ground-truth clean frame from Eq. (2), but Eq. (23) includes Lspa for unpaired real data, where B_c is unavailable. Additionally, the sentence after Eq. (17) states that “\tilde{B}_c and B_c denote the current frames restored by our dual-branch model and the spatial-branch-only model,” which overloads B_c with the ground-truth notation and makes Lrec and Lspa either identical (if B_c is ground truth) or notationally inconsistent (if B_c is a network output). Since the RVDT results and real-world visual comparisons depend on unpaired training, the training protocol needs an explicit correction: either omit Lspa for unpaired data, or redefine it against the masked median target, and state this clearly.
  2. [§3.2, Eq. (16)] The claimed reduction to the mean filter is not a well-defined limit. “mean(G), if a → 0, b → ∞” depends on the order of limits: taking a → 0 first gives the mean for any fixed b, while taking b → ∞ first can make the exponent a(1/N)∑|x_n − b x_j| diverge unless a b → 0. Eqs. (15) and (16) also show inconsistent normalization (1/N versus 1/n and summation ranges j=1..N versus j=1..n). Please correct the formula and specify the limiting order; otherwise the statement that DSF “adaptively approximates” the four statistical filters is not established.
  3. [§4.1, Tables 1 and 2] No repeated-run statistics are reported. The margins over RainMamba are 0.79 dB on RainSynLight25, 0.24 dB on RainSynComplex25, and 1.87 dB on NTURain; the downstream gains in Table 2 are also modest (e.g., MOTA 31.8 versus 29.5 and MOTP 21.2 versus 20.3). Without multiple seeds or a significance test, the claim of state-of-the-art performance is not statistically grounded. Please report mean ± std over at least three training runs, or a paired test, for the key comparisons and ablations.
  4. [§3.3, Eq. (20) and §4.4] The pseudo-label mask quality is load-bearing for the real-world claim, but it is only evaluated visually. The paper acknowledges in §3.3 that occlusions and moving objects can cause mismatches, and Fig. 8 shows the effect of θ for a few examples, yet there is no quantitative analysis of how many patches are accepted on real videos or how θ affects the RVDT detection/tracking numbers. Please report the acceptance statistics of the mask and the sensitivity of Table 2 to θ.
minor comments (7)
  1. [§3.1, Eq. (2)] The sentence “where In, Bn, and nt denote” appears to contain a typo; “nt” should be “Rn”.
  2. [§3.2] The statement “The proof is provided in the Supplements” for Theorem 1 is not verifiable because no supplementary material is included with the submission. Since this is a standard property of the median, either include the proof in an appendix or cite a standard reference.
  3. [§3.4] The paper says “Detailed statistics and examples are provided in the Supplements,” but the submission contains no supplementary file. Please include the supplement or move the benchmark statistics into the main text.
  4. [§3.3, Eq. (22)] In Ltem, the flow estimator F is applied to the pair (\tilde{B}_c, I_n), but Eq. (3) defines F as a pre-trained estimator for clean pairs. Clarify whether this is the same estimator and why it is expected to produce reliable flow when one input is restored and the other is rainy.
  5. [Figs. 5–7] Claims such as “only our method successfully removes all rain streaks” are based on visual inspection. Please add a quantitative real-world evaluation (e.g., a no-reference metric or the downstream task numbers) to support this wording.
  6. [§3.4 and Table 2] The paper does not state clearly whether RVDT is used for training, validation, or only evaluation. This information is needed to interpret the downstream results and the role of the semi-supervised loss.
  7. [§4.4, Table 3] The exact definitions of Strategy 1, Strategy 2, and Strategy 3 for optical flow estimation are described only briefly; please provide sufficient detail to make the ablations reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity: the dynamic stacking filter and losses are independently defined, with only a notation/specification flaw in Eq. 17-23 that is not a by-construction reduction.

full rationale

VDMamba's derivations do not exhibit circular reduction. The dynamic stacking filter (Eq. 16) is a softmax relaxation defined by the authors; its convergence to mean/median/max/min is an explicit limit computation in the paper itself, not an imported theorem or a fitted parameter. The semi-supervised median-stacking loss (Eqs. 18-21) builds pseudo-clean targets from the input's warped median and a fixed mask; this is self-supervision, not a prediction fitted to the evaluation metrics, and the reported PSNR/SSIM/downstream numbers are measured on external benchmarks after training, not used as training targets. The self-citations [45]-[49] occur only in background/related-work contexts and carry no load-bearing uniqueness claim. One genuine gap exists but is not circular: Eq. 17 defines Lspa with the same expression as Lrec, while the text identifies B_c as either ground truth (Eq. 2) or the spatial-branch output, and Eq. 23 includes Lspa in the unpaired branch; under the ground-truth reading the unpaired loss is not computable, and under the spatial-output reading the two losses are redundant. This is a specification/notation flaw, not a by-construction equivalence, so it does not raise the circularity score.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The central method depends on two learned filter parameters (a, b), several hand-chosen hyperparameters (theta, delta, lambda1, lambda2, patch count), and domain assumptions about additive rain, reliable optical flow, and the validity of median pseudo-labels. None of these are derived from first principles; they are standard engineering choices for a learning-based restoration system.

free parameters (6)
  • Temperature coefficient a in DSF = learned per-pixel parameter (matrix a in R^{HxW})
    Controls the sharpness of the softmax approximations to min/max/median in Eqs. 10-16. It is learned during training, not derived from first principles.
  • Gate coefficient b in DSF = learned per-pixel parameter (matrix b in R^{HxW})
    Gates the absolute deviations in Eq. 15 to steer the filter between mean, median, max, and min behavior. Learned, not derived.
  • Mask threshold theta = 80%
    The fraction of unchanged pixels that decides whether a sub-patch is adopted as pseudo-clean in Eq. 20. Chosen by hand; ablation shows 40% is too slack and 100% is too strict.
  • Slack delta = 0.1
    Pixel-value tolerance in Eq. 20 for deciding whether a pixel is unchanged. Chosen by hand in Sec. 4.1.
  • Loss weights lambda1, lambda2 = 0.1, 0.1
    Weights for the stacking loss and temporal loss in the total loss. Chosen by hand.
  • Number of patches P^2 for masking = P = 2 (implied by P^2 sub-patches; exact P not stated)
    The sub-patch grid size for the median stacking mask is used but the paper does not state P. It is a hand-chosen hyperparameter affecting pseudo-label quality.
assumptions (5)
  • domain assumption The rainy video is an additive model I_n = B_n + R_n with rain streaks as a sparse additive layer.
    Assumed in Sec. 2 and used to justify the sparsity-based median pseudo-label in Sec. 3.3. Real rain can include occlusion, motion blur, and complex photometric effects that violate strict additivity.
  • standard math The differentiable softmax approximations in Eqs. 10-15 converge to the true min/max/median as a goes to infinity.
    The convergence claim is standard for softmax-style approximations and is effectively correctness of a mathematical limit, not an empirical assumption.
  • domain assumption A pre-trained optical flow estimator F (LiteFlowNet3) on clean frames can be distilled to the multi-frame flow estimator f, and the distilled flow is reliable enough to warp rainy frames.
    Invoked in Sec. 3.1 for the flow transfer loss and in Fig. 3. If optical flow on rainy frames is too unreliable, both deraining alignment and the pseudo-label generation break.
  • domain assumption Using the model's own restored previous frames as additional input during training and inference helps rather than harms.
    Mentioned in Sec. 3.1 ('We also include the restored previous frames along with the rainy frames, though this detail is omitted for brevity'). This is an unstated design choice with no ablation.
  • domain assumption Sub-patches that are mostly unchanged between the median and central aligned frame are clean and well-aligned (Eq. 20 mask).
    Load-bearing assumption of the semi-supervised pseudo-label. Occasional moving objects, occlusions, or flow failure could produce unchanged patches that are not clean backgrounds.
invented entities (2)
  • Dynamic Stacking Filter (DSF)
    purpose: A per-pixel learned fusion of aligned neighboring frame features that adaptively approximates mean, median, max, and min statistics.
    The filter is a mathematical construction defined inside the paper (Eq. 16), with no external falsifiable handle beyond the benchmark results.
  • RVDT benchmark
    purpose: A new real-world rainy video dataset with 33,077 detection annotations and 408 tracking identities, used to evaluate downstream tasks.
    The benchmark is introduced by this paper; no independent external validation of the annotations is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining." pith.science (2026). https://pith.science/paper/2LZORV7O

@misc{pith2026250516811,
  author       = {Pith},
  title        = {Pith review of: Semi-Supervised State-Space Model with Dynamic Stacking Filter for Real-World Video Deraining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2LZORV7O}},
  note         = {Machine review of arXiv:2505.16811}
}
read the original abstract

Significant progress has been made in video restoration under rainy conditions over the past decade, largely propelled by advancements in deep learning. Nevertheless, existing methods that depend on paired data struggle to generalize effectively to real-world scenarios, primarily due to the disparity between synthetic and authentic rain effects. To address these limitations, we propose a dual-branch spatio-temporal state-space model to enhance rain streak removal in video sequences. Specifically, we design spatial and temporal state-space model layers to extract spatial features and incorporate temporal dependencies across frames, respectively. To improve multi-frame feature fusion, we derive a dynamic stacking filter, which adaptively approximates statistical filters for superior pixel-wise feature refinement. Moreover, we develop a median stacking loss to enable semi-supervised learning by generating pseudo-clean patches based on the sparsity prior of rain. To further explore the capacity of deraining models in supporting other vision-based tasks in rainy environments, we introduce a novel real-world benchmark focused on object detection and tracking in rainy conditions. Our method is extensively evaluated across multiple benchmarks containing numerous synthetic and real-world rainy videos, consistently demonstrating its superiority in quantitative metrics, visual quality, efficiency, and its utility for downstream tasks.

Figures

Figures reproduced from arXiv: 2505.16811 by the authors.

Figure 1
Figure 1. The functionality of various stacking filters in video [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The comparison of performance, time and space com [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Given two adjacent rainy frames (3a & 3d), rain degrada￾tion leads to errors in optical flow estimation and subsequent frame alignment (3b & 3c). In contrast, our degradation-free multi-frame estimation pipeline facilitates more accurate warping. In this manner, a single model forward pass generates mul￾tiple optical flows simultaneously, effectively reducing run￾time. Furthermore, this approach helps mitigate the d… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The architecture of our proposed VDMamba for video deraining, consisting of spatial state-space model layers (S3ML) for [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: The qualitative comparison among the existing deraining methods on synthetic datasets [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The qualitative comparison among the existing deraining methods on real-world rainy videos from [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Rainy-scene object detection and tracking results on RVDT after video deraining. Three results are obtained using YOLO [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: The qualitative ablation of the stacking loss. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 60 canonical work pages

  1. [1]

    A review on deep convolu- tional neural networks

    Neena Aloysius and M Geetha. A review on deep convolu- tional neural networks. In ICCSP, 2017. 4

  2. [2]

    Spatio-temporal frequency analysis for removing rain and snow from videos

    Peter Barnum, Takeo Kanade, and Srinivasa G Narasimhan. Spatio-temporal frequency analysis for removing rain and snow from videos. In PACV, 2007. 2

  3. [3]

    Analysis of rain and snow in frequency space.IJCV, 2010

    Peter Barnum, Srinivasa Narasimhan, and Takeo Kanade. Analysis of rain and snow in frequency space.IJCV, 2010. 2

  4. [4]

    Evaluating multiple object tracking performance: the clear mot metrics

    Keni Bernardin and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. JIVP,

  5. [5]

    Bossu, N

    J. Bossu, N. Hauti `ere, and J.-P. Tarel. Rain or snow detection in image sequences through use of a histogram of orientation of streaks. IJCV, 2011. 2

  6. [6]

    Vi- sual depth guided color image rain streaks removal using sparse coding

    Duan-Yu Chen, Chien-Cheng Chen, and Li-Wei Kang. Vi- sual depth guided color image rain streaks removal using sparse coding. IEEE TCSVT, 2014. 2

  7. [7]

    Robust video content alignment and compensation for rain removal in a cnn framework

    Jie Chen, Cheen-Hau Tan, Junhui Hou, Lap-Pui Chau, and He Li. Robust video content alignment and compensation for rain removal in a cnn framework. In CVPR, 2018. 1, 3, 6, 7

  8. [8]

    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 CVPR, 2023. 1, 2, 6, 7, 8

Show all 74 references
  1. [9]

    Mem- ory enhanced global-local aggregation for video object de- tection

    Yihong Chen, Yue Cao, Han Hu, and Liwei Wang. Mem- ory enhanced global-local aggregation for video object de- tection. In CVPR, 2020. 6, 8

  2. [10]

    A generalized low- rank appearance model for spatio-temporally correlated rain streaks

    Yi-Lei Chen and Chiou-Ting Hsu. A generalized low- rank appearance model for spatio-temporally correlated rain streaks. In ICCV, 2013. 2

  3. [11]

    Sigmoid- weighted linear units for neural network function approxi- mation in reinforcement learning

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid- weighted linear units for neural network function approxi- mation in reinforcement learning. Neural Netw., 2018. 4

  4. [12]

    Clearing the skies: A deep network ar- chitecture for single-image rain removal

    Xueyang Fu, Jiabin Huang, Xinghao Ding, Yinghao Liao, and John Paisley. Clearing the skies: A deep network ar- chitecture for single-image rain removal. IEEE TIP, 2017. 2

  5. [13]

    Removing rain from single images via a deep detail network

    Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In CVPR, 2017. 2

  6. [14]

    Garg and S

    K. Garg and S. Nayar. Detection and removal of rain from videos. In CVPR, 2004. 2

  7. [15]

    Kshitiz Garg and Shree K. Nayar. When does a camera see rain? In ICCV, 2005

  8. [16]

    Kshitiz Garg and Shree K. Nayar. Vision and rain. IJCV,

  9. [17]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3

  10. [18]

    Efficiently modeling long sequences with structured state spaces

    Albert Gu, Karan Goel, and Christopher R ´e. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3

  11. [19]

    Mambair: A simple baseline for image restoration with state-space model

    Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. arXiv preprint arXiv:2402.15648, 2024. 3

  12. [20]

    Vehicle detection and tracking in ad- verse weather using a deep learning framework

    Mahmoud Hassaballah, Mourad A Kenk, Khan Muhammad, and Shervin Minaee. Vehicle detection and tracking in ad- verse weather using a deep learning framework. IEEE TITS,

  13. [21]

    Mambavision: A hy- brid mamba-transformer vision backbone

    Ali Hatamizadeh and Jan Kautz. Mambavision: A hy- brid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024. 3

  14. [22]

    Hinton, Oriol Vinyals, and Jeffrey Dean

    Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531, 2015. 3, 4

  15. [23]

    Object detection under rainy conditions for autonomous vehicles: A review of state- of-the-art and emerging techniques

    Mazin Hnewa and Hayder Radha. Object detection under rainy conditions for autonomous vehicles: A review of state- of-the-art and emerging techniques. IEEE SPM, 2020. 1

  16. [24]

    Liteflownet3: Resolv- ing correspondence ambiguity for more accurate optical flow estimation

    Tak-Wai Hui and Chen Change Loy. Liteflownet3: Resolv- ing correspondence ambiguity for more accurate optical flow estimation. In ECCV, 2020. 6

  17. [25]

    Multi-scale progressive fusion network for single image deraining

    Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In CVPR, 2020. 2

  18. [26]

    Dawn: Direction-aware attention wavelet network for image deraining

    Kui Jiang, Wenxuan Liu, Zheng Wang, Xian Zhong, Junjun Jiang, and Chia-Wen Lin. Dawn: Direction-aware attention wavelet network for image deraining. In ACM MM, 2023. 2

  19. [27]

    A novel tensor-based video rain streaks removal approach via utilizing discriminatively in- trinsic priors

    Tai-Xiang Jiang, Ting-Zhu Huang, Xi-Le Zhao, Liang-Jian Deng, and Yao Wang. A novel tensor-based video rain streaks removal approach via utilizing discriminatively in- trinsic priors. In CVPR, 2017. 1, 2

  20. [28]

    Fastderain: A novel video rain streak removal method using directional gradient priors

    Tai-Xiang Jiang, Ting-Zhu Huang, Xi-Le Zhao, Liang-Jian Deng, and Yao Wang. Fastderain: A novel video rain streak removal method using directional gradient priors. IEEE TIP,

  21. [29]

    Automatic single-image-based rain streaks removal via image decom- position

    Li-Wei Kang, Chia-Wen Lin, and Yu-Hsiang Fu. Automatic single-image-based rain streaks removal via image decom- position. IEEE TIP, 2012. 2

  22. [30]

    Single-image deraining using an adaptive nonlocal means filter

    Jin-Hwan Kim, Chul Lee, Jae-Young Sim, and Chang-Su Kim. Single-image deraining using an adaptive nonlocal means filter. In ICIP, 2013. 2

  23. [31]

    Video deraining and desnowing using temporal correlation and low-rank matrix completion

    Jin-Hwan Kim, Jae-Young Sim, and Chang-Su Kim. Video deraining and desnowing using temporal correlation and low-rank matrix completion. IEEE TIP, 2015. 2

  24. [32]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6

  25. [33]

    Fouriermamba: Fourier learning integration with state space models for image deraining

    Dong Li, Yidi Liu, Xueyang Fu, Senyan Xu, and Zheng- Jun Zha. Fouriermamba: Fourier learning integration with state space models for image deraining. arXiv preprint arXiv:2405.19450, 2024. 3

  26. [34]

    Video rain streak removal by multi- scale convolutional sparse coding

    Minghan Li, Qi Xie, Qian Zhao, Wei Wei, Shuhang Gu, Jing Tao, and Deyu Meng. Video rain streak removal by multi- scale convolutional sparse coding. In CVPR, 2018. 2

  27. [35]

    Jamba: A hybrid transformer-mamba language model

    Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024. 3

  28. [36]

    Erase or fill? deep joint recurrent rain removal and recon- struction in videos

    Jiaying Liu, Wenhan Yang, Shuai Yang, and Zongming Guo. Erase or fill? deep joint recurrent rain removal and recon- struction in videos. In CVPR, 2018. 3, 6, 7

  29. [37]

    D3r-net: Dynamic routing residue recurrent network for video rain removal

    Jiaying Liu, Wenhan Yang, Shuai Yang, and Zongming Guo. D3r-net: Dynamic routing residue recurrent network for video rain removal. IEEE TIP, 2019. 3

  30. [38]

    Removing rain from a single image via discriminative sparse coding

    Yu Luo, Yong Xu, and Hui Ji. Removing rain from a single image via discriminative sparse coding. In ICCV, 2015. 2

  31. [39]

    Yolov3: An incremental improvement

    Joseph Redmon. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767, 2018. 6, 8

  32. [40]

    Video desnowing and deraining based on matrix decomposition

    Weihong Ren, Jiandong Tian, Zhi Han, Antoni Chan, and Yandong Tang. Video desnowing and deraining based on matrix decomposition. In CVPR, 2017. 2

  33. [41]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3, 8

  34. [42]

    Varun Santhaseelan and Vijayan K. Asari. Utilizing local phase information to remove rain from video. IJCV, 2015. 2

  35. [43]

    Prior-based domain adaptive object detec- tion for hazy and rainy conditions

    Vishwanath A Sindagi, Poojan Oza, Rajeev Yasarla, and Vishal M Patel. Prior-based domain adaptive object detec- tion for hazy and rainy conditions. In ECCV, 2020. 1

  36. [44]

    Simplified state space layers for sequence modeling

    Jimmy TH Smith, Andrew Warrington, and Scott W Linder- man. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 3

  37. [45]

    Rethinking image restoration for object detection

    Shangquan Sun, Wenqi Ren, Tao Wang, and Xiaochun Cao. Rethinking image restoration for object detection. NeurIPS,

  38. [46]

    Event-aware video derain- ing via multi-patch progressive learning

    Shangquan Sun, Wenqi Ren, Jingzhi Li, Kaihao Zhang, Meiyu Liang, and Xiaochun Cao. Event-aware video derain- ing via multi-patch progressive learning. IEEE TIP, 2023. 1, 2, 3

  39. [47]

    Restoring images in adverse weather condi- tions via histogram transformer

    Shangquan Sun, Wenqi Ren, Xinwei Gao, Rui Wang, and Xiaochun Cao. Restoring images in adverse weather condi- tions via histogram transformer. In ECCV, 2024. 2

  40. [48]

    Logit standardization in knowledge distillation

    Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, and Xi- aochun Cao. Logit standardization in knowledge distillation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 15731–15740, 2024. 3

  41. [49]

    A hybrid transformer- mamba network for single image deraining

    Shangquan Sun, Wenqi Ren, Juxiang Zhou, Jianhou Gan, Rui Wang, and Xiaochun Cao. A hybrid transformer- mamba network for single image deraining. arXiv preprint arXiv:2409.00410, 2024. 3

  42. [50]

    Perceptual adversarial networks for image-to-image transformation

    Chaoyue Wang, Chang Xu, Chaohui Wang, and Dacheng Tao. Perceptual adversarial networks for image-to-image transformation. IEEE TIP, 2018. 2

  43. [51]

    A model- driven deep neural network for single image rain removal

    Hong Wang, Qi Xie, Qian Zhao, and Deyu Meng. A model- driven deep neural network for single image rain removal. In CVPR, 2020. 2

  44. [52]

    Should we encode rain streaks in video as de- terministic or stochastic? In ICCV, 2017

    Wei Wei, Lixuan Yi, Qi Xie, Qian Zhao, Deyu Meng, and Zongben Xu. Should we encode rain streaks in video as de- terministic or stochastic? In ICCV, 2017. 1, 2

  45. [53]

    Semi-supervised transfer learning for image rain re- moval

    Wei Wei, Deyu Meng, Qian Zhao, Zongben Xu, and Ying Wu. Semi-supervised transfer learning for image rain re- moval. In CVPR, 2019. 2

  46. [54]

    Rainmamba: Enhanced locality learning with state space models for video deraining

    Hongtao Wu, Yijun Yang, Huihui Xu, Weiming Wang, Jinni Zhou, and Lei Zhu. Rainmamba: Enhanced locality learning with state space models for video deraining. In ACM MM,

  47. [55]

    Fusionmamba: Dynamic feature enhancement for mul- timodal image fusion with mamba

    Xinyu Xie, Yawen Cui, Tao Tan, Xubin Zheng, and Zitong Yu. Fusionmamba: Dynamic feature enhancement for mul- timodal image fusion with mamba. Visual Intelligence, 2(1): 37, 2024. 3

  48. [56]

    Image deraining with frequency-enhanced state space model

    Shugo Yamashita and Masaaki Ikehara. Image deraining with frequency-enhanced state space model. arXiv preprint arXiv:2405.16470, 2024. 3

  49. [57]

    Tan, Wenhan Yang, and Dengxin Dai

    Wending Yan, Robby T. Tan, Wenhan Yang, and Dengxin Dai. Self-aligned video deraining with transmission-depth consistency. In CVPR, 2021. 3

  50. [58]

    Frame- consistent recurrent video deraining with dual-level flow

    Wenhan Yang, Jiaying Liu, and Jiashi Feng. Frame- consistent recurrent video deraining with dual-level flow. In CVPR, 2019. 3

  51. [59]

    Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu

    Wenhan Yang, Robby T. Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep net- works. IEEE TPAMI, 2020. 6

  52. [60]

    Tan, Shiqi Wang, and Jiaying Liu

    Wenhan Yang, Robby T. Tan, Shiqi Wang, and Jiaying Liu. Self-learning video rain streak removal: When cyclic consis- tency meets temporal correspondence. In CVPR, 2020. 1, 3, 5, 6

  53. [61]

    Tan, Jiashi Feng, Shiqi Wang, Bin Cheng, and Jiaying Liu

    Wenhan Yang, Robby T. Tan, Jiashi Feng, Shiqi Wang, Bin Cheng, and Jiaying Liu. Recurrent multi-frame deraining: Combining physics guidance and adversarial learning. IEEE TPAMI, 2022. 3, 6, 7, 8

  54. [62]

    Tan, Shiqi Wang, Alex C

    Wenhan Yang, Robby T. Tan, Shiqi Wang, Alex C. Kot, and Jiaying Liu. Learning to remove rain in video with self- supervision. IEEE TPAMI, 2022. 3

  55. [63]

    Rajeev Yasarla and Vishal M. Patel. Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. In CVPR, 2019. 2

  56. [64]

    Tan, Rei Kawakami, Yasuhiro Mukaigawa, and Katsushi Ikeuchi

    Shaodi You, Robby T. Tan, Rei Kawakami, Yasuhiro Mukaigawa, and Katsushi Ikeuchi. Adherent raindrop mod- eling, detection and removal in video. IEEE TPAMI, 2016. 2

  57. [65]

    Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992,

    Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992,

  58. [66]

    Semi-supervised video deraining with dynamical rain gener- ator

    Zongsheng Yue, Jianwen Xie, Qian Zhao, and Deyu Meng. Semi-supervised video deraining with dynamical rain gener- ator. In CVPR, 2021. 1, 2, 3, 6

  59. [67]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming–Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2

  60. [68]

    He Zhang, Vishwanath Sindagi, and Vishal M. Patel. Im- age de-raining using a conditional generative adversarial net- work. IEEE TCSVT, 2020. 2

  61. [69]

    Enhanced spatio-temporal interaction learning for video deraining: A faster and better framework

    Kaihao Zhang, Dongxu Li, Wenhan Luo, Wenqi Ren, and Wei Liu. Enhanced spatio-temporal interaction learning for video deraining: A faster and better framework. IEEE TPAMI, 2022. 3, 6, 7, 8

  62. [70]

    Rain removal in video by combining tempo- ral and chromatic properties

    Xiaopeng Zhang, Hao Li, Yingyi Qi, Wee Kheng Leow, and Teck Khim Ng. Rain removal in video by combining tempo- ral and chromatic properties. In ICME, 2006. 2

  63. [71]

    Freqmamba: Viewing mamba from a frequency perspective for image deraining

    Zou Zhen, Yu Hu, and Zhao Feng. Freqmamba: Viewing mamba from a frequency perspective for image deraining. arXiv preprint arXiv:2404.09476, 2024. 3

  64. [72]

    Global tracking transformers

    Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Kr¨ahenb¨uhl. Global tracking transformers. In CVPR, 2022. 6, 8

  65. [73]

    Vision mamba: Efficient visual representation learning with bidirectional state space model

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 3

  66. [74]

    Freqmamba: Viewing mamba from a frequency perspective for image de- raining

    Zhen Zou, Hu Yu, Jie Huang, and Feng Zhao. Freqmamba: Viewing mamba from a frequency perspective for image de- raining. In ACM MM, 2024. 2

Pith tools

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