Pith. sign in

REVIEW 3 major objections 6 minor 61 references

DerainSplat: Feed-Forward Clean 3D Gaussian Splatting from Sparse Rainy Views

T0 review · 3 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read Two rainy views produce one clean 3D scene in a single pass

desk verdict A solid synthetic-domain feed-forward deraining 3DGS framework whose 'cross-dataset generalization' claim is not yet backed by real-rain quantitative evidence. read the letter →

arxiv 2608.02191 v1 pith:QNN2VAXN submitted 2026-08-03 cs.CV

classification cs.CV
keywords rainy3Dreconstructionfeed-forwardGaussiansplattingmulti-viewderainingweatherfactordecompositionsyntheticraindatasetsparseviewsynthesiscycleconsistencycost-volumematching
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 show that a clean 3D scene can be reconstructed from just two rainy photographs in a single forward pass, without per-scene optimization. Existing feed-forward 3D reconstruction methods collapse when the input is degraded by rain and haze; this framework claims to be the first tailored to such sparse rainy inputs. The route is to predict a set of weather factors (transmission, airlight, rain streaks, lens-drop masks) and convert them into two support maps that tell the geometry matcher and the appearance fuser which pixels to trust. In the paper's largest comparison, the method ranks first on all metrics and beats feed-forward baselines by more than 4 dB PSNR. The authors also build a large synthetic multi-view rain dataset, with paired clean/rainy images and privileged weather factors, and argue that models trained on it generalize to real rain.

What carries the argument

The load-bearing object is the weather net and its two support maps. A compact encoder-decoder predicts pixel-wise transmission, rain-streak layer and masks, lens-affected mask, plus a global airlight and color transform from each rainy view. Those factors are converted into (1) a scene support s_geo = h(t) ⊙ (1 - m_rs) ⊙ (1 - m_rd), which weights the cost volume along epipolar lines so the depth posterior trusts rain-free, well-see-through pixels; and (2) a radiance support s_rad = 1 - m_rd, which blends the reference view's color with depth-aligned source pixels. The geometry evidence derived from the support volume also attenuates Gaussian opacity. A rainy cycle consistency loss re-applie

What would settle it

A real-world multi-view capture of the same scene under rain and under clear sky, with aligned cameras; run DerainSplat (trained only on synthetic rain) and two-stage baselines on the rainy views and compare PSNR/SSIM/LPIPS against the clear captures. If the method's margin over two-stage or finetuned baselines drops below the synthetic gap, or if the predicted streak mask on real rain has significantly different angle/length/width statistics from the synthetic predictions, the central claim about realistic generalization is undermined.

Watch

Extended reading notes

Core claim

The central claim is that rainy-to-clean 3D reconstruction should be learned directly in a multi-view, weather-aware setting rather than as two-stage per-view deraining plus reconstruction. DerainSplat recovers clean 3D Gaussians from two rainy context views by first predicting disentangled weather factors with a compact weather net, then turning those factors into a scene support that down-weights degraded regions during cross-view cost-volume matching and a radiance support that drives depth-aligned appearance fusion to fill corrupted pixels. The resulting geometry evidence attenuates Gaussian opacity at unreliable locations, and a rainy cycle consistency loss re-renders clean views with t

Load-bearing premise

The four-stage synthetic rain pipeline—overcast dimming, depth-dependent haze, screen-blended streaks, and replayable lens drops—is representative enough of real rain that models trained on it generalize, since the only real-rain evidence in the paper is qualitative (no aligned clean targets).

Editorial extensions

If this is right

  • If the claim holds, embodied AI and autonomous driving can obtain clean 3D geometry and appearance from as few as two rainy camera views, with no per-scene fitting.
  • Reconstructing in one pass (about 0.47 s on a GPU) makes the approach practical for real-time systems, unlike per-scene optimizers that take minutes.
  • Ablations show both support maps contribute, and removing the rainy cycle loss costs about 1 dB PSNR, so the factor-supervised, self-consistent design is the source of the gain.
  • The four-stage synthetic dataset supplies paired clean-rain images and privileged factors, enabling weather-supervised training and providing a benchmark for future multi-view deraining methods.
  • Qualitative results on real-world rainy scenes, plus matching streak geometry statistics, suggest the synthetic pipeline is realistic enough to transfer.

Reading between the lines

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

  • The same two-support-map pattern (one for geometry trust, one for appearance trust) could be applied to other transient or localized degradations—fog layers, snow, glare, or even dynamic occluders—since the cost-volume and fusion mechanisms are not rain-specific.
  • The paper's real-rain evidence is only qualitative, so the strongest test of the framework is a quantitative real benchmark with aligned clean targets; if the 4 dB margin narrows or vanishes on such data, the contribution is mainly the dataset and training recipe, not the architecture.
  • The privileged weather factors could become a common pretraining signal for other rain-aware 3D models, turning DerainSplat's dataset into a reusable resource beyond the paper's own method.
  • One could test the weather-net's disentanglement by synthesizing a scene with known factors and checking whether the predicted factors recover them; failure would indicate the supervision is being fit to the dataset rather than to physical rain formation.
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 DerainSplat, a feed-forward 3D Gaussian Splatting framework that takes sparse rainy views (two views in the reported setting) and predicts a clean 3D Gaussian scene in a single forward pass. To support training, the authors construct a large multi-view synthetic deraining dataset by a four-stage pipeline that applies overcast illumination, depth-dependent haze, rain streaks, and lens raindrops to clean images from RealEstate10K, ACID, and Mip-NeRF360, storing privileged weather factors as supervision. The framework uses a weather net to predict these factors and converts them into two support maps: a scene support that modulates cross-view matching and Gaussian opacity, and a radiance support that drives depth-aligned appearance fusion. A rainy-cycle consistency loss re-renders the clean context views with the predicted factors and aligns them with the rainy inputs. Experiments on synthetic Rainy RE10K, Rainy ACID, and Rainy Mip-NeRF360 report large gains over feed-forward and per-scene baselines, while real-world rainy scenes are evaluated only qualitatively.

Significance. If the reported results hold on independent data, DerainSplat would be a useful contribution to an underexplored problem: clean feed-forward 3D reconstruction from rainy images. The explicit decomposition into weather factors, the support-map mechanism, and the rainy-cycle consistency are well motivated, and the ablations show clear contributions of the major components. The dataset construction with privileged factors is a valuable resource if released. However, the central limitation is that every quantitative benchmark is generated by the same synthesis pipeline used for training, and real-world validation is only qualitative; therefore the 'strong cross-dataset generalization' claim is not yet substantiated. The technical core is sound and the paper fits the journal's scope, but the evidence needs strengthening or the claims need to be tempered.

major comments (3)
  1. [Sec. 4.1, Sec. 4.2, Abstract] All quantitative benchmarks (Rainy RE10K, Rainy ACID, and Rainy Mip-NeRF360) are generated by the proposed four-stage synthesis pipeline, as stated in Sec. 4.1. This means the reported >4 dB gains and the 'strong cross-dataset generalization' claim in the Abstract come from in-distribution evaluation with respect to the exact formation model the model was trained on. The Rainy Mip-NeRF360 split is described as a cross-dataset test, but it is also generated by the same pipeline, so it does not test generalization to unseen weather formation. Real-world evidence is only qualitative (Fig. 10) or indirect (Table 3 compares streak geometry, not reconstruction quality). I recommend evaluating on independent rainy multi-view data, ideally real captures with some ground-truth proxy or at least a different synthesis pipeline, or revising the generalization claims to match the evidence.
  2. [Sec. 3.1, Alg. 1 line 4] The text states that 'both γ and the depth normalization range are shared across views of a scene, so the same 3D point receives consistent haze.' However, Algorithm 1 line 4 computes d̃_{s,v} ← Norm(D_{s,v}) per view. If Norm is a per-view min–max normalization, the transmission t_{s,v} = exp(−β_s d̃_{s,v}) will not be view-consistent for the same 3D point, contradicting the stated design goal. This is load-bearing because multi-view consistency of the weather factors is the basis for the support-map modulation of cross-view matching. Please clarify the normalization scope and ensure the implementation matches the text.
  3. [Sec. 4.1 and Sec. 4.3] The real-rain evaluation is explicitly limited to qualitative results, as the manuscript acknowledges in Sec. 4.1: 'Real-world rainy scenes lack aligned clean targets and are used for qualitative evaluation only.' The paper nevertheless uses this evaluation to support the 'strong cross-dataset generalization' claim in the Abstract and Sec. 4.3. While the lack of aligned real clean targets is understandable, the current evidence does not support a quantitative generalization claim. At minimum, the paper should state clearly that all quantitative conclusions are in-distribution and that real-world performance is unverified quantitatively.
minor comments (6)
  1. [Sec. 4.2, Table 2] The text says DerainSplat 'enables real-time reconstruction,' but Table 2 reports 0.47 s per scene. This is substantially faster than the per-scene optimization methods (≈400 s) but is not real-time in the usual interactive sense. Suggest rewording to 'near real-time' or 'much faster than per-scene optimization.'
  2. [Sec. 4.1, Implementation Details] Typo: 'Training is conducted uses with a batch size of 10' should be 'Training is conducted with a batch size of 10.'
  3. [Fig. 4 caption, Sec. 3.2] Typo: 'apearance' should be 'appearance.' Also, the math symbols in the Fig. 4 caption are garbled (e.g., '𝒎𝒓𝒔),𝒎𝒓𝒅*').
  4. [Sec. 3.3, Eq. (11)] The notation for the lens mask loss is unclear: λ_d, λ_g, and the area slack δ_guard appear without explicit definitions. Please define these hyperparameters and their roles (e.g., in the text or caption).
  5. [Table 3] The 'Overlap' metric between predicted real streak geometry and synthetic ground truth is not a reconstruction-quality metric. The claim that 'the synthetic streak geometry is physically consistent with real rainfall' would be more convincing with a direct visual comparison of predicted masks on real and synthetic images, or a more rigorous distributional test.
  6. [Sec. 3.1, Alg. 1] Minor typo: 'Visualization exmaple' should be 'Visualization example.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the synthetic-benchmark dependence is an external-validity limitation, not a circular derivation.

full rationale

The derivation chain is self-contained and does not reduce to its inputs. The four-stage synthesis pipeline (Sec. 3.1, Eqs. 1-3) is used only to generate training pairs and privileged weather factors; the model is trained with clean-target supervision (Eq. 9), direct weather-factor supervision (Eq. 10), and a rainy-cycle consistency (Eqs. 13-14) that requires the predicted factors and rendered clean view to jointly explain the observed rainy input. None of these losses defines the target as a function of the prediction: the clean target is a held-out image, and the weather factors are generator-side ground truths, not model outputs. The support maps and opacity modulation (Eqs. 4-6) are architectural mechanisms, not fitted parameters later renamed as predictions. There are no load-bearing self-citations and no imported uniqueness theorems. The paper itself flags the main limitation: "All synthetic benchmarks are generated by the proposed four-stage synthesis pipeline" (Sec. 4.1) and "Real-world rainy scenes lack aligned clean targets and are used for qualitative evaluation only" (Sec. 4.3). This means the quantitative generalization evidence is in-distribution with respect to the rain formation model, which is a correctness/generalization concern, not a circularity: the test images and clean targets are not constructed from the model's fitted outputs or predictions, and the reported PSNR is measured on held-out scene splits. Therefore the circularity score is 0.

Assumptions & free parameters 6 free parameters · 6 assumptions · 3 invented entities

The framework rests on a fairly long list of hand-chosen synthesis parameters and domain assumptions: the weather formation model itself, the depth source, the loss scheduling, and the unverified transfer of synthetic factors to real scenes. The most honest reading is that the paper contributes a method plus a synthetic benchmark, with real-world evidence still qualitative.

free parameters (6)
  • scene-level sunny score gamma = per-scene value in [0,1]
    Estimated from clean views and used to set overcast intensity and beta in the haze stage; a fitted quantity that changes the whole synthesis.
  • haze coefficient beta = beta = 0.5 + 0.8 * gamma
    Hand-chosen linear function of gamma; controls transmission map t = exp(-beta * normalized_depth) and hence haze strength.
  • airlight color A = not fully specified; sampled/input
    A global color used in the atmospheric scattering model; its distribution is not described, affecting all synthetic rainy images.
  • rain streak motion kernel parameters = theta0 ~ U(-8,8), sigma_g in {1.3,2.2,3.0} degrees
    Hand-set distributions for streak angle and blur; determine streak morphology in the synthetic dataset.
  • lens raindrop state L = per-scene alpha, f, sigma, highlights
    Generated per scene without specified distributions; controls the lens-drop corruption model in Eq. 3.
  • loss and annealing hyperparameters = lambda_w, lambda_r, lambda_p, lambda_d, lambda_g, eta=0.1, lambda=0.1, s_p=5000, s_c=5000
    Hand-chosen weights and schedules that affect training stability and final results; no sensitivity analysis is provided.
assumptions (6)
  • domain assumption Atmospheric scattering model: I_h = I_o * t + A * (1 - t)
    Assumed haze formation model, cited to Narasimhan and Nayar 2002 and used as Eq. 1 in the synthesis pipeline.
  • domain assumption Rain streaks combine via screen blending
    Eq. 2 assumes rain streaks are composited with screen blending, which is a particular and not universally valid model.
  • ad hoc to paper Lens raindrops replace pixels via coverage mask with warp/blur/highlights
    Eq. 3 introduces a synthetic lens-drop model adapted from prior raindrop papers; its realism is assumed, not validated quantitatively.
  • domain assumption Depth Anything 3 monocular depths are accurate and multi-view consistent
    Section 3.1 Stage II relies on DA3 to build depth-dependent transmission and to keep haze consistent across views.
  • domain assumption Synthetic weather factors and supervision transfer to real rain
    The method is trained on synthesized factors, and real-world validation is only qualitative, so this transfer assumption is load-bearing.
  • standard math Differentiable 3DGS rendering with sparse views is an adequate scene model
    The framework adopts MVSplat-style feed-forward 3DGS; the rendering equation (Eq. 7) is standard and cited to Kerbl et al. 2023.
invented entities (3)
  • Weather factor set (t, A, S, mrs, mrd, gamma)
    purpose: Disentangle rainy appearance into transmission, airlight, rain streak layer, and masks to guide reconstruction
    These factors are internal parameterizations supervised only by synthetic ground truth from the authors' own pipeline; no independent real-world measurement validates them.
  • Scene support sgeo and radiance support srad
    purpose: Modulate cross-view cost-volume matching and depth-aligned appearance fusion
    Computational constructs derived from predicted weather factors; their usefulness is shown only via ablation on synthetic data.
  • Lens censor mask u_c
    purpose: Exclude lens-affected pixels from the rainy-cycle consistency loss
    Annealed from ground-truth to predicted mask during training; a model-internal mechanism without external validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DerainSplat: Feed-Forward Clean 3D Gaussian Splatting from Sparse Rainy Views." pith.science (2026). https://pith.science/paper/QNN2VAXN

@misc{pith2026260802191,
  author       = {Pith},
  title        = {Pith review of: DerainSplat: Feed-Forward Clean 3D Gaussian Splatting from Sparse Rainy Views},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QNN2VAXN}},
  note         = {Machine review of arXiv:2608.02191}
}
read the original abstract

Although image deraining has advanced substantially, existing methods mainly focus on 2D image restoration. As spatial intelligence applications such as embodied AI and autonomous driving continue to emerge, reconstructing clean 3D scenes from sparse rainy views in a feed-forward manner becomes increasingly important. Existing feed-forward 3D Gaussian Splatting (3DGS) methods often assume clean inputs and collapse under rainy conditions. To this end, we present \textbf{\textit{DerainSplat}}, a feed-forward framework that reconstructs clean 3D scenes from only a few rainy views. To support this task, we build a large-scale multi-view derain dataset through a four-stage synthesis pipeline that sequentially models overcast illumination, depth-dependent haze, rain streaks, and lens raindrops, producing privileged weather factors. We introduce a weather net that predicts the weather factors from rainy context and yields two support maps. Scene support modulates cross-view cost-volume matching, while radiance support drives depth-aligned appearance fusion to fill corrupted pixels. The derived geometry evidence further attenuates Gaussian opacity to reduce spurious structures. A rainy cycle consistency re-renders clean views using the predicted factors and aligns them with rainy inputs. Extensive experiments show that \textbf{\textit{DerainSplat}} outperforms existing methods on various datasets, including RealEstate10K, ACID, Mip-NeRF360, and real-world rainy scenes, with strong cross-dataset generalization.

Figures

Figures reproduced from arXiv: 2608.02191 by the authors.

Figure 1
Figure 1. (a) Given each rainy context view, DerainSplat predicts a clean 3D Gaussian scene in a single forward pass. (b) We build a large-scale multi-view derain dataset with paired clean–rain images and privileged weather factors. DerainSplat renders cleaner novel views than existing feed-forward baselines, with fewer rain/haze artifacts and stronger structural fidelity. Rainy Input CSUD Restore MVSplat Ours [PITH_FULL_IMA… view at source ↗
Figure 2
Figure 2. Failure of a two-stage strategy. Individually plau￾sible restored views lack cross-view consistency. ing, and a radiance support that drives depth-aligned appear￾ance fusion to fill corrupted pixels with reliable cross-view observations. The resulting geometry evidence further at￾tenuates Gaussian opacity at unreliable locations, reducing spurious structures without discarding valid scene content. The clean reconstr… view at source ↗
Figure 3
Figure 3. Stage-wise visualizations of the derain dataset syn￾thesis process. Starting from clean images, we sequentially synthesize overcast appearance, depth-dependent haze, rain streaks, and lens raindrops to obtain final rainy observations. HSV space and scale its saturation and brightness by γ, so brighter scenes are dimmed and desaturated more aggres￾sively: I o = T (I c ; γ). The sunny score is shared across all views … view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overview of DerainSplat. Rainy context views are processed by the weather net to estimate weather factors (t,ˆ A, ˆ S, ˆ mˆrs , mˆrs) and support maps that guide cost-volume matching, apearance fusion, and opacity, yielding clean 3DGS. Final Rain Transmission Rain Stre…
Figure 5
Figure 5. Figure 5: Visualization of weather factor. Algorithm 1: Procedure for Synthetic Derain Dataset Require: Clean multi-view scenes {(I c s , Cs)} M s=1, I c s = {I c s,v} Ns v=1 and Cs = {(Ks,v, Ts,v)} Ns v=1; depth estimator E (DA3); tone operator T (·); airlight color A ∈ [0, 1]3…
Figure 7
Figure 7. Figure 7: Qualitative comparison of novel views on rainy RealEstate10K and ACID datasets. Compared with feed-forward baselines, DerainSplat yields cleaner novel views with fewer rain/haze artifacts and clearer scene structures. Inputs GT PixelSplat MVSplat TransSplat Ours [PITH…
Figure 8
Figure 8. Figure 8: Finetuned qualitative comparison on rainy RE10K. 1) Rainy RE10K (Zhou et al. 2018) contains 13,768/1,455 (train/test) rainy outdoor scenes and serves as the primary benchmark for training and in-domain evaluation. 2) Rainy ACID (Liu et al. 2021) contains approximately …
Figure 11
Figure 11. Figure 11: Comparison of rainy-cycle consistency. The upper and lower rows are produced without and with the cycle loss, respectively; (a) and (b) trained 1K and 10K steps. with scene reliability, suppressing weather-induced spuri￾ous Gaussian. Radiance Support denotes support-w…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 6 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    Mvsplat360: Feed-forward 360 scene synthesis from sparse views , author=. Advances in Neural Information Processing Systems , volume=

  2. [2]

    European Conference on Computer Vision (ECCV) , year=

    NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis , author=. European Conference on Computer Vision (ECCV) , year=

  3. [3]

    arXiv preprint arXiv:2512.03004 , year=

    DGGT: Feedforward 4D Reconstruction of Dynamic Driving Scenes using Unposed Images , author=. arXiv preprint arXiv:2512.03004 , year=

  4. [4]

    arXiv preprint arXiv:2504.20995 , year=

    TesserAct: learning 4D embodied world models , author=. arXiv preprint arXiv:2504.20995 , year=

  5. [5]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal =

    Kerbl, Bernhard and Kopanas, Georgios and Leimk. 3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal =. 2023 , url =

  6. [6]

    Charatan and D

    S. Charatan and D. K. Finkel and T. J. B. Rees. PixelSplat: Scalable 3D Gaussian Splatting from Image Pairs. CVPR 2024. 2024

  7. [7]

    J. M. Xie and D. F. Jin. MVSplat: Cost-Volume Plane Sweeping for Efficient Sparse-View 3D Reconstruction. ECCV 2024. 2024

  8. [8]

    Zhang and J

    T. Zhang and J. Y. Joo. TranSplat: Transformer-Based Sparse-View 3D Reconstruction. AAAI 2025. 2025

Show all 61 references
  1. [9]

    T. L. Hamid and S. J. R. McKinley. AnySplat: Feed-Forward Gaussian Reconstruction from Unconstrained Views. arXiv:2505.23716. 2025

  2. [10]

    Li and G

    P. Li and G. Wang. WeatherGS: 3D Scene Reconstruction in Adverse Weather with Gaussian Splatting. ICRA 2025. 2025

  3. [11]

    Zhang and Y

    S. Zhang and Y. Shi. DerainNet: A Deep Learning Approach to Rain Removal. IEEE Transactions on Image Processing. 2021

  4. [12]

    He and X

    H. He and X. Li. DehazeNet: A Deep Network for Single Image Dehazing. IEEE Transactions on Image Processing. 2016

  5. [13]

    Zhang and L

    H. Zhang and L. Zhu. CSUD: Multi-Scale Joint Rain Removal for Sparse Views. CVPR 2024. 2024

  6. [14]

    Srinivasan and Matthew Tancik and Jonathan T

    Ben Mildenhall and Pratul P. Srinivasan and Matthew Tancik and Jonathan T. Barron and Ravi Ramamoorthi and Ren Ng , title =. CVPR , year =

  7. [15]

    arXiv , volume =

    Wei-Ting Chen and Yifan Wang and Sy-Yen Kuo and Gordon Wetzstein , title =. arXiv , volume =. 2023 , url =

  8. [16]

    arXiv , volume =

    Andrea Ramazzina and Mario Bijelic and Stefanie Walz and Alessandro Sanvito and Dominik Scheuble and Felix Heide , title =. arXiv , volume =. 2023 , url =

  9. [17]

    European Conference on Computer Vision , pages=

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images , author=. European Conference on Computer Vision , pages=. 2024 , organization=

  10. [18]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  11. [19]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  12. [20]

    arXiv preprint arXiv:2511.12040 , year=

    SRSplat: Feed-Forward Super-Resolution Gaussian Splatting from Sparse Multi-View Images , author=. arXiv preprint arXiv:2511.12040 , year=

  13. [21]

    2023 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Visual language maps for robot navigation , author=. 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2023 , organization=

  14. [22]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Drivingforward: Feed-forward 3d gaussian splatting for driving scene reconstruction from flexible surround-view input , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  15. [23]

    arXiv preprint arXiv:1805.09817 , year=

    Stereo magnification: Learning view synthesis using multiplane images , author=. arXiv preprint arXiv:1805.09817 , year=

  16. [24]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Infinite nature: Perpetual view generation of natural scenes from a single image , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  17. [25]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

  18. [26]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  19. [27]

    arXiv preprint arXiv:1412.6980 , year=

    Adam: A method for stochastic optimization , author=. arXiv preprint arXiv:1412.6980 , year=

  20. [28]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Depthsplat: Connecting gaussian splatting and depth , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  21. [29]

    ACM Transactions on Graphics (ToG) , volume=

    Local light field fusion: Practical view synthesis with prescriptive sampling guidelines , author=. ACM Transactions on Graphics (ToG) , volume=. 2019 , publisher=

  22. [30]

    arXiv preprint arXiv:2511.10647 , year=

    Depth anything 3: Recovering the visual space from any views , author=. arXiv preprint arXiv:2511.10647 , year=

  23. [31]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Mip-Nerf 360: Unbounded anti-aliased neural radiance fields , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  24. [32]

    arXiv preprint arXiv:2408.11540 , year=

    DeRainGS: Gaussian Splatting for Enhanced Scene Reconstruction in Rainy Environments , author=. arXiv preprint arXiv:2408.11540 , year=

  25. [33]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Weathergs: 3d scene reconstruction in adverse weather conditions via gaussian splatting , author=. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2025 , organization=

  26. [34]

    International Journal of Computer Vision , volume =

    Vision and Rain , author =. International Journal of Computer Vision , volume =. 2007 , doi =

  27. [35]

    IEEE Transactions on Image Processing , volume =

    Automatic Single-Image-Based Rain Streaks Removal via Image Decomposition , author =. IEEE Transactions on Image Processing , volume =. 2012 , doi =

  28. [36]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages =

    Removing Rain from a Single Image via Discriminative Sparse Coding , author =. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages =. 2015 , doi =

  29. [37]

    IEEE Transactions on Image Processing , volume =

    Clearing the Skies: A Deep Network Architecture for Single-Image Rain Removal , author =. IEEE Transactions on Image Processing , volume =. 2017 , doi =

  30. [38]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Deep Joint Rain Detection and Removal From a Single Image , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2017 , url =

  31. [39]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Density-Aware Single Image De-Raining Using a Multi-Stream Dense Network , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2018 , url =

  32. [40]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages =

    Recurrent Squeeze-and-Excitation Context Aggregation Net for Single Image Deraining , author =. Proceedings of the European Conference on Computer Vision (ECCV) , pages =. 2018 , doi =

  33. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Progressive Image Deraining Networks: A Better and Simpler Baseline , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2019 , url =

  34. [42]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Multi-Stage Progressive Image Restoration , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2021 , url =

  35. [43]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Restormer: Efficient Transformer for High-Resolution Image Restoration , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2022 , url =

  36. [44]

    2019 IEEE International Conference on Image Processing (ICIP) , year =

    Unsupervised Single Image Deraining with Self-Supervised Constraints , author =. 2019 IEEE International Conference on Image Processing (ICIP) , year =. doi:10.1109/ICIP.2019.8803238 , url =

  37. [45]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2025 , url =

  38. [46]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Spatial Attentive Single-Image Deraining with a High Quality Real Rain Dataset , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2019 , url =

  39. [47]

    International Journal of Computer Vision , volume =

    Vision and the Atmosphere , author =. International Journal of Computer Vision , volume =. 2002 , doi =

  40. [48]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Single Image Haze Removal Using Dark Channel Prior , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2011 , doi =

  41. [49]

    IEEE Transactions on Image Processing , volume =

    Single-Image Real-Time Rain Removal Based on Depth-Guided Non-Local Features , author =. IEEE Transactions on Image Processing , volume =. 2021 , doi =

  42. [50]

    European Conference on Computer Vision (ECCV) , year =

    Not Just Streaks: Towards Ground Truth for Single Image Deraining , author =. European Conference on Computer Vision (ECCV) , year =

  43. [51]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Attentive Generative Adversarial Network for Raindrop Removal From a Single Image , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2018 , doi =

  44. [52]

    IEEE Transactions on Image Processing , volume =

    Rethinking Image Deraining via Rain Streaks and Vapors , author =. IEEE Transactions on Image Processing , volume =. 2020 , doi =

  45. [53]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    From Sky to the Ground: A Large-scale Benchmark and Simple Baseline Towards Real Rain Removal , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  46. [54]

    AAAI Conference on Artificial Intelligence , year =

    GTAV-NightRain: Photometric Realistic Large-scale Dataset for Night-time Rain Streak Removal , author =. AAAI Conference on Artificial Intelligence , year =

  47. [55]

    arXiv preprint arXiv:2404.11401 , year =

    RainyScape: Unsupervised Rainy Scene Reconstruction using Decoupled Neural Rendering , author =. arXiv preprint arXiv:2404.11401 , year =

  48. [56]

    arXiv preprint arXiv:2511.06734 , year =

    Rethinking Rainy 3D Scene Reconstruction via Perspective Transforming and Brightness Tuning , author =. arXiv preprint arXiv:2511.06734 , year =

  49. [57]

    Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , pages =

    Real-Time Adherent Raindrop Modelling and Removal for Automotive Cameras , author =. Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , pages =. 2019 , doi =

  50. [58]

    Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) , year =

    Adherent Raindrop Removal with Self-Supervised Attention Maps and Spatio-Temporal Generative Adversarial Networks , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW) , year =

  51. [59]

    Proceedings of the Asian Conference on Computer Vision Workshops (ACCVW) , year =

    Realistic Modeling of Water Droplets for Monocular Adherent Raindrop Recognition , author =. Proceedings of the Asian Conference on Computer Vision Workshops (ACCVW) , year =

  52. [60]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Adherent Raindrop Detection and Removal in Video , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2013 , doi =

  53. [61]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Waterdrop Stereo , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2017 , doi =

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.