Pith. sign in

REVIEW 4 major objections 4 minor 52 references

Event-Based De-Snowing for Autonomous Driving

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Combining event-camera data with a single snowy frame removes occlusions by exploiting space-time snow streaks, outperforming image- and video-only de-snowers by 3 dB PSNR.

desk verdict The paper's own Table II contradicts the headline 3 dB claim on real data, but the DSEC-Snow dataset idea is worth a serious look. read the letter →

arxiv 2507.20901 v1 pith:533ZQ6JO submitted 2025-07-25 cs.CV

classification cs.CV
keywords eventcamerasde-snowingspatio-temporalstreaksimagereconstructionautonomousdrivingdatasetgenerationopticalflowadverseweather
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

Snow occlusions, unlike ordinary scene motion, carve a recognizable streak through the space-time volume of an event camera, and the paper's thesis is that those streaks carry enough information to reconstruct the occluded background. The authors build a network that reads streak geometry with an attention-based module, produces a per-pixel occlusion mask, and fuses the event-derived estimate with a single snowy RGB frame. To train and measure it, they create DSEC-Snow, a dataset that overlays real recorded snowflake events onto a driving dataset with clean ground truth, plus a controlled real-snow dataset and real driving sequences. The central claim is a gain of about 3 dB in PSNR over state-of-the-art image- and video-only de-snowers, and about 20% better downstream optical-flow performance.

What carries the argument

The load-bearing object is the snowflake streak in the spatio-temporal event volume: because event cameras report contrast changes with sub-millisecond latency, each occluding snowflake appears as a continuous line in x-y-t rather than a blur whose appearance depends on exposure time. The method's components are an EventNet (ConvLSTM plus U-Net) that extracts streak features and outputs a soft occlusion mask, a transformer backbone that fuses image and event features, and an adaptive pixel-wise fusion that keeps the original image where the mask says no occlusion. Together these components learn a data-driven version of the geometric relation $I_b = I_r - \sum p C$, where the sum of event polarities along the streak, scaled by the contrast threshold, recovers the occluded background intensity.

What would settle it

Run the trained model on a controlled real-snow rig with a known static background, an event camera, and a synchronized RGB camera, and compare against a clean ground-truth frame: if flakes whose motion includes ego-motion-induced depth parallax produce misaligned occlusion masks, PSNR will drop toward the image-only baseline, contradicting the claimed 3 dB advantage.

Watch

Extended reading notes

Core claim

The paper claims de-snowing can be turned from a spatial hallucination problem into a temporal measurement problem. Because an event camera reports brightness changes asynchronously with no exposure time, a falling snowflake always leaves a continuous streak in the x-y-t event volume, and the polarity of events tells whether the flake is brighter or darker than its background. Given the flake intensity and the sensor contrast threshold, the background intensity can in principle be recovered by integrating event polarity along the streak, with a warp when the camera is moving; the network learns this operation implicitly. On their benchmark, fusing event features with one image through a learned mask raises reconstruction quality above frame-based and video-based state of the art by about 3 dB PSNR, and the cleaned images improve downstream optical-flow and depth estimation.

Load-bearing premise

The dataset's realism rests on the assumption that snow captured in front of a black screen can be linearly overlaid on driving events, with flakes pruned only by a fixed contrast rule and warped by a homography; if real snow under camera ego-motion, depth, and changing illumination produces event patterns that do not match this composite, the learned streak-to-mask mapping, and the 3 dB gain, will not transfer to real driving.

Editorial extensions

If this is right

  • One snowy image plus ten milliseconds of events suffices for high-quality de-snowing, so performance no longer depends on high RGB frame rates.
  • DSEC-Snow provides the first synchronized image-event-snow dataset with clean ground truth, making event-based de-snowing trainable and reproducible.
  • Because the model learns from the streak signature rather than from snow appearance, it also reduces rain, lens-flare, and other transient occlusions in qualitative tests.
  • Desnowed outputs plug into off-the-shelf optical-flow and depth algorithms, improving their accuracy relative to the outputs of image-only de-snowers.

Reading between the lines

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

  • Because streak contrast depends on the sensor's contrast threshold rather than exposure time, newer event sensors with lower thresholds should make fainter snow streaks visible; the paper does not test this scaling, but the geometric model predicts it directly.
  • The synthetic-data recipe should generalize to other fast transient occluders, such as pedestrians, vehicles, or debris, by recording foreground event streams in front of a black screen and overlaying them on arbitrary backgrounds.
  • A sharper test of the method's real-world premise would be to train on physically simulated snow with depth-dependent, ego-motion-corrected particle trajectories and compare against the homography-warped composite used here; that comparison would quantify how much of the reported 3 dB gain depends on the dataset simplification.
  • The claimed 20% optical-flow improvement is measured with one flow network on one dataset, so replicating the result across weather conditions and flow estimators would tell how broadly the downstream benefit holds.
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 / 4 minor

Summary. The paper proposes an event-based de-snowing method for autonomous driving. It introduces a synthetic dataset (DSEC-Snow) built by chroma-key compositing of separately recorded foreground snow events onto background frames and events from DSEC, plus a controlled real dataset (Slider-Snow) and qualitative real driving sequences. The method uses an EventNet (ConvLSTM plus U-Net) to extract spatio-temporal snow-streak features and a learned mask, a Snowformer-style transformer to fuse image and event features, and an adaptive fusion step to output the de-snowed image. The authors claim a 3 dB PSNR improvement over state-of-the-art frame-based and event-based methods on both synthetic and real datasets, and a 20% improvement in downstream optical flow.

Significance. The paper addresses a relevant real-world problem, and the idea of exploiting the distinctive streak signature of snow in event data is plausible and potentially valuable. The controlled Slider-Snow setup is a reasonable attempt to obtain real ground truth, and the ablation of naive fusion provides some evidence that events help in the synthetic setting. However, the central claims are not supported by the paper's own reported experiments. The headline 3 dB improvement is contradicted by Table II on the real Slider-Snow benchmark, the synthetic evaluation is partly circular because the same foreground event stream is used both to synthesize the snowy image and to construct the input events, and the baseline comparison protocol is not described with enough detail to rule out domain-shift artifacts. No code or dataset is released, so the results cannot be independently checked. As presented, the paper's significance is therefore not established.

major comments (4)
  1. [Abstract; Section X (Conclusion); Table II (Section VIII-A)] The abstract and conclusion claim a 3 dB PSNR improvement over state-of-the-art methods 'on both synthetic and real datasets.' Table II contradicts this on the real Slider-Snow dataset: Restormer and S2VD report 25.16 dB PSNR while the proposed method reports 21.55 dB, i.e., 3.61 dB lower, and the SSIM is also lower (0.6832 vs. 0.8939 and 0.8706). Section VIII-A's statement that the proposed method 'achieves the highest PSNR and SSIM across both datasets' is therefore inconsistent with the paper's own table. This is an internal falsification of the central claim, not a domain-gap speculation.
  2. [Section VII-A, Eq. (9), Algorithm 2, Section IX, Section XI] The DSEC-Snow evaluation is partially circular. In Eq. (9), the snowy image Z(x) is generated by overlaying the foreground snow events Esnow(x) onto a hazy background, and in Algorithm 2 the synthetic event stream E(x) is built from the same Esnow(x), pruned by a contrast rule that depends on the background intensity. Consequently, the event cue is perfectly aligned with the snow occlusions in the image by construction. A network can learn to exploit this artificial alignment, so the large DSEC-Snow margin (28.08 dB vs. 21.48 dB for SnowFormer) does not by itself demonstrate real-world transfer. The paper's own limitations weaken the transfer argument further: Section IX acknowledges that independent event streams are linearly combined without modeling lighting interactions or occlusion ordering, and Section XI states that snow-particle motion is not modeled according to ego-motion, only a homography is applied to foreground events.
  3. [Section VI (Baselines); Section VIII-B] The baseline comparison protocol is not described with enough detail to make the synthetic results meaningful. The paper does not state whether Restormer, SnowFormer, RLP, S2VD, and E2VID are retrained or fine-tuned on the DSEC-Snow training split; Section VIII-B says only that the proposed model is trained on synthetic DSEC-Snow and applied directly. If the baselines use their originally released weights, the comparison measures domain shift rather than the benefit of the event modality. This is a load-bearing issue because the claimed 3 dB improvement rests on the DSEC-Snow numbers.
  4. [Table III; Section VIII-A] Table III is internally inconsistent. The proposed method has the lowest EPE (22.64) but lower accuracy values than several baselines under the same metric family: AE<3 is 0.130 for Ours versus 0.284 for Restormer and 0.273 for RLP, and AE<5 is 0.270 versus 0.412 and 0.421. A method with lower EPE should generally have higher accuracy at fixed thresholds, so either the metric definitions or the reported numbers are erroneous. The abstract's claim of a 20% improvement in downstream tasks is also not clearly supported: relative EPE improvement over the best baseline (RLP, 27.33) is about 17.2%, and no definition of the 20% figure is provided.
minor comments (4)
  1. [Section XI (Dataset Statistics)] The dataset statistics are inconsistent: Section XI states the dataset 'consists of around 200 training and 50 test sequences,' then later says 'The dataset consists of 1000 training and 470 test pairs of images and events.' The dataset name is also used inconsistently across the paper (DSEC-Snow, Slider-Snow, SnowDriving in Table I).
  2. [Figures and cross-references] Several figure cross-references are incorrect. For example, Section VII-A says 'Some examples from these sequences are shown in Fig. 6,' but Fig. 6 is captioned as Slider-Snow, while Fig. 5 shows DSEC-Snow. Figure 8 is referenced in Section VII-C before Fig. 5 and Fig. 6 are introduced, and the figure numbering appears out of order.
  3. [Throughout] There are numerous typos and formatting errors, including 'evalutaing,' 'DSEC-Snowdataset,' 'V olume' in reference [9], duplicate references [36] and [37] for the same Time Lens++ paper, and the S2VD citation in Table III is given as '[2]' instead of '[35]'. The pseudocode in Algorithm 1 also uses potentially confusing notation, such as 'τX pC' without clear indexing.
  4. [Section IV-B, Eq. (7)] The loss function uses two terms with weights λ1 and λ2, but the text says λ1 and λ2 are 'empirically set to 1 and 0.2'; this is fine, but the paper does not report sensitivity to these weights or whether the perceptual loss is essential, despite the ablation study being limited to architecture changes.

Circularity Check

1 steps flagged · score 6.0 of 10

DSEC-Snow benchmark is circular by construction: the same foreground snow event stream Esnow generates both the snowy image (Eq. 9) and the input event stream (Algorithm 2), so the reported 3 dB synthetic gain is not independent evidence.

  1. self definitional [Section VII-A, Eq. (9) and Algorithm 2]
    "Z(x) = Ihaze(x) + α · Aug(Esnow(x)) ... For each event esnow ∈ Esnow(x): Given event location (x, y) and timestamp t If |Ihaze(x, y) − Isnow(x, y)| > C, add esnow to E(x)"

    The synthetic snowy image Z(x) and the synthetic input event stream E(x) are both constructed from the same recorded snow event stream Esnow(x). Therefore, on DSEC-Snow, the event input is, by construction, a perfect label of where the snow occlusions were overlaid. The network can learn to read snow positions directly from events, and the large margin over image-only baselines (28.08 vs 21.48 dB for SnowFormer) measures this built-in alignment rather than a validated real-world ability to detect snow. This makes the DSEC-Snow half of the headline '3 dB on both synthetic and real datasets' self-referential rather than an independent test.

full rationale

The synthetic DSEC-Snow evaluation is partially circular: Eq. (9) builds the snowy image from Esnow, and Algorithm 2 builds the event input from the same Esnow, so the event cue perfectly coincides with the artifact by construction. This is a concrete same-variable reduction, not a speculation about intent. The real Slider-Snow results are not circular, but they independently contradict the headline claim: Table II lists Restormer and S2VD at 25.16 dB PSNR while the proposed method is 21.55 dB, so the 'both synthetic and real datasets' claim is internally falsified; I treat that as a correctness problem rather than a circularity step. Self-citations (DSEC [16], [51]) are not load-bearing: DSEC is an external driving dataset and [51] supports only an SNR remark. Sections IX and XI disclose additional synthetic-to-real limitations (linear compositing, no snow ego-motion), which further weaken transfer but are limitations, not circular reasoning. Because one of the two quantitative pillars of the central claim reduces to a construction, the score is 6 rather than lower; the architecture itself still has independent content, so 8-10 would overstate the circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central result depends on several unmeasured constants and simplifying assumptions in dataset construction. The most important are the contrast threshold C and snow intensity Ir used to synthesize events and to run the model-based baseline, plus the linear compositing and foreground-only assumptions that make the synthetic events artificially aligned with image snow.

free parameters (5)
  • contrast threshold C = not specified (paper says typically around 15%)
    Used in Equations 1-3 for model-based recovery and in Algorithm 2 to decide whether a snow event is added. The actual sensor threshold is unknown and illumination-dependent, so it is a free parameter.
  • snowflake intensity Ir = not specified
    Assumed known in Equation 1-2 and needed in Algorithm 2's contrast check |Ihaze - Isnow| > C. Snowflake intensity depends on distance, illumination, and background; not measured.
  • atmospheric light A = not specified
    Appears in Equation 8 haze rendering; value chosen per scene but never reported, affecting all synthetic images.
  • alpha (ambient illumination blend) = not specified
    Equation 9 controls snow overlay strength; follows Chen et al. [8] but the exact values used are not reported.
  • snow speed and density augmentation scales = not specified
    Supplementary describes scaling timestamps and staggering overlays to control speed and density; exact values not reported.
assumptions (5)
  • domain assumption Snowflakes are brighter than the background and always in the foreground
    Supplementary Section XI states 'snow particles tend to be brighter than most objects' and 'snowflakes will always be in the foreground'; this underlies event pruning in Algorithm 2 and the sign convention in Equation 1.
  • ad hoc to paper Background and foreground event streams can be linearly combined without physical interaction
    Explicitly stated as a limitation in Section IX: 'it relies on the assumption that independently captured motions ... can be linearly combined without introducing artifacts.' This is needed to build the synthetic event stream.
  • ad hoc to paper Event generation is governed by a fixed contrast threshold with known snow intensity
    Algorithm 2 adds a snow event only when |Ihaze - Isnow| > C, ignoring real sensor noise, illumination-dependent threshold, and partial occlusions. This is the model that creates the correlation between image snow and events.
  • domain assumption Static scene depth during fusion; snow motion does not include ego-motion
    Supplementary says 'we simplify this problem and only apply homography transformation to the foreground events' and 'we do not model the motion of the snow particles according to ego-motion of the camera.' This makes the synthetic events differ from driving snow.
  • standard math Standard atmospheric scattering model [44] for haze
    Equation 8 uses the atmospheric scattering model; accepted prior art, not the paper's contribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Event-Based De-Snowing for Autonomous Driving." pith.science (2026). https://pith.science/paper/533ZQ6JO

@misc{pith2026250720901,
  author       = {Pith},
  title        = {Pith review of: Event-Based De-Snowing for Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/533ZQ6JO}},
  note         = {Machine review of arXiv:2507.20901}
}
abstract

Adverse weather conditions, particularly heavy snowfall, pose significant challenges to both human drivers and autonomous vehicles. Traditional image-based de-snowing methods often introduce hallucination artifacts as they rely solely on spatial information, while video-based approaches require high frame rates and suffer from alignment artifacts at lower frame rates. Camera parameters, such as exposure time, also influence the appearance of snowflakes, making the problem difficult to solve and heavily dependent on network generalization. In this paper, we propose to address the challenge of desnowing by using event cameras, which offer compressed visual information with submillisecond latency, making them ideal for de-snowing images, even in the presence of ego-motion. Our method leverages the fact that snowflake occlusions appear with a very distinctive streak signature in the spatio-temporal representation of event data. We design an attention-based module that focuses on events along these streaks to determine when a background point was occluded and use this information to recover its original intensity. We benchmark our method on DSEC-Snow, a new dataset created using a green-screen technique that overlays pre-recorded snowfall data onto the existing DSEC driving dataset, resulting in precise ground truth and synchronized image and event streams. Our approach outperforms state-of-the-art de-snowing methods by 3 dB in PSNR for image reconstruction. Moreover, we show that off-the-shelf computer vision algorithms can be applied to our reconstructions for tasks such as depth estimation and optical flow, achieving a $20\%$ performance improvement over other de-snowing methods. Our work represents a crucial step towards enhancing the reliability and safety of vision systems in challenging winter conditions, paving the way for more robust, all-weather-capable applications.

Figures

Figures reproduced from arXiv: 2507.20901 by the authors.

Figure 1
Figure 1. Event-based video snow removal in challenging nighttime scenes: (a) Sample image from our dataset captured while driving in snowfall, and (b) corresponding event data highlighting the motion of snowflakes. The presence of dense, dynamic snow and low visibility presents significant challenges for conventional image restoration methods, as illustrated by the results of (c) Restormer [1] and (d) Snowformer [2]. Our pro… view at source ↗
Figure 2
Figure 2. Effect of snow on images (left) and events (right) (Left) The appearance of snowfall in an image depends on multiple factors, such as snowflake size, density, ambient illu￾mination, and camera exposure settings. Two such examples are shown within the red and blue regions. (Right) The same snow occlusions in event data are visualized on the image plane and x-y-t volume (bottom-right). Irrespective of the snowflake si… view at source ↗
Figure 3
Figure 3. Overview of our data-driven method for recon￾structing deoccluded images using event-camera data. Events within the camera’s exposure time are segmented into non-overlapping spatio-temporal windows, converted into voxel-grid representations, and processed alongside RGB im￾ages through modality-specific feature extraction. Event fea￾tures are extracted via EventNet, which also produces a spatial mask. The image recon… view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Overview of the synthetic snow dataset generation process. Given a clean background image J(x) and its corresponding event stream Ehaze(x) and foreground snow event stream Esnow(x), we generate a synthetic snow-occluded image Z(x) and a synthetic event stream E(x). The…
Figure 5
Figure 5. Figure 5: Example scenes from our DSEC-Snowdataset. It consists of synchronized RGB frames(Left), Events (Middle) and Groundtruth (Right). B. Slider-Snowdataset To evaluate the model on real data, we propose to collect our own dataset. Evaluating with real snowfall, however, is …
Figure 6
Figure 6. Figure 6: Examples scenes from our Slider-Snowdataset It consists of synchronized RGB frames(Left), Events (Middle) and Groundtruth (Right). propose to use a controlled setup to evaluate the performance in the real world, using a snow machine. We use a linear slider to move the …
Figure 7
Figure 7. Figure 7: Overview of the experimental setups (a) Our setup for recording the foreground occlusion events used for gen￾erating DSEC-Snowsequence. (b) Experimental setup for our controlled real-world dataset. The scene was printed on a poster and the camera was placed on a linear…
Figure 8
Figure 8. Figure 8: Samples from our real-world snowfall driving sequences The images are recorded using the synchronized and aligned setup of RGB camera (top) and event camera (bottom) and mounted on the dashboard of the car while driving in the snowfall. (a) Image (b) Restormer [1] (c) …
Figure 9
Figure 9. Figure 9: Qualitative comparison of image desnowing results on synthetic data. (a) Input images with synthetic snow,(b)-(c) image restoration baselines , (d) our result, and (e) ground truth (GT) images. The proposed approach recovers clearer scene structures and more faithfully…
Figure 10
Figure 10. Figure 10: Effect of occlusion density on image reconstruction quality Each row corresponds to a different level of occlusion density, increasing from top to bottom. As occlusion density increases, visibility of scene details and robustness of the image￾only algorithm degrade si…
Figure 11
Figure 11. Figure 11: Comparing the image reconstruction, optical flow estimation for event-only baseline E2VID, video baseline S2VD, and image baseline Restormer with our method on the Slider-Snowdataset. Zoom Zoom (a) Image (b) Restormer [1] (d) Ours (c) GT [PITH_FULL_IMAGE:figures/full…
Figure 13
Figure 13. Figure 13: Qualitative comparison on real driving scenes. (a) Input images, (b)-(c) image restoration baselines, and (d) our results. Rows labeled ”Zoom” show magnified regions with dense snow and intense light sources, highlighting the ability of our method to reduce snow occlu…
Figure 14
Figure 14. Figure 14: Qualitative comparison on the MUSES Dataset [50] with rain and snow occlusions. (a) Input images affected by rain or snow occlusions, and (b) results from our method. The examples show the effectiveness of our approach in handling adverse weather occlusions across dif…
Figure 15
Figure 15. Figure 15: Dataset statistics of DSEC-Snowdataset. The his [PITH_FULL_IMAGE:figures/full_fig_p013_15.png]
Figure 16
Figure 16. Figure 16: Comparison of our method with state-of-the-art de-snowing methods on DSEC-Snowdataset. We show samples from our method and compare it with state-of-the-art de-snowing methods on the Slider-Snowdataset in [PITH_FULL_IMAGE:figures/full_fig_p014_16.png]
Figure 17
Figure 17. Figure 17: Comparison of our method with state-of-the-art de-snowing methods on Slider-Snowdataset. [7] W.-T. Chen, H.-Y. Fang, C.-L. Hsieh, C.-C. Tsai, I. Chen, J.-J. Ding, S.-Y. Kuo et al., “All snow removed: Single image desnowing algo￾rithm using hierarchical dual-tree compl…
Figure 18
Figure 18. Figure 18: Qualitative results of our method on downstream task - depth estimation. We show the input image and depth map from the DSEC-Snow dataset, along with the depth map generated by our method. The depth map is estimated using the input image and events, demonstrating the …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 50 canonical work pages

  1. [1]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” in CVPR, 2022

  2. [2]

    Snowformer: Scale- aware transformer via context interaction for single image desnowing,

    S. Chen, T. Ye, Y . Liu, E. Chen, J. Shi, and J. Zhou, “Snowformer: Scale- aware transformer via context interaction for single image desnowing,” arXiv preprint arXiv:2208.09703 , 2022

  3. [3]

    Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather,

    M. Bijelic, T. Gruber, F. Mannan, F. Kraus, W. Ritter, K. Dietmayer, and F. Heide, “Seeing through fog without seeing fog: Deep multimodal sensor fusion in unseen adverse weather,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2020

  4. [4]

    Desnownet: Context-aware deep network for snow removal,

    Y .-F. Liu, D.-W. Jaw, S.-C. Huang, and J.-N. Hwang, “Desnownet: Context-aware deep network for snow removal,” IEEE Transactions on Image Processing, vol. 27, no. 6, pp. 3064–3073, 2018

  5. [5]

    Deep dense multi-scale network for snow removal using semantic and geometric priors,

    K. Zhang, R. Li, Y . Yu, W. Luo, and C. Li, “Deep dense multi-scale network for snow removal using semantic and geometric priors,” IEEE Transactions on Image Processing , 2021

  6. [6]

    Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal,

    W.-T. Chen, H.-Y . Fang, J.-J. Ding, C.-C. Tsai, and S.-Y . Kuo, “Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal,” in European Conference on Computer Vision , 2020. 15 (a) Input (b) Events (c) Model-based (d) Restormer [1] (e) RLP [20] (f) SnowFormer [2] (g) S2VD [35] (h) E...

  7. [7]

    All snow removed: Single image desnowing algo- rithm using hierarchical dual-tree complex wavelet representation and contradict channel loss,

    W.-T. Chen, H.-Y . Fang, C.-L. Hsieh, C.-C. Tsai, I. Chen, J.-J. Ding, S.-Y . Kuo et al. , “All snow removed: Single image desnowing algo- rithm using hierarchical dual-tree complex wavelet representation and contradict channel loss,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 4196–4205

  8. [8]

    Snow removal in video: A new dataset and a novel method,

    H. Chen, J. Ren, J. Gu, H. Wu, X. Lu, H. Cai, and L. Zhu, “Snow removal in video: A new dataset and a novel method,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2023, pp. 13 211–13 222

Show all 52 references
  1. [9]

    When does a camera see rain?

    K. Garg and S. Nayar, “When does a camera see rain?” in Tenth IEEE International Conference on Computer Vision (ICCV’05) V olume 1, vol. 2, 2005, pp. 1067–1074 V ol. 2

  2. [10]

    Video snow removal based on self-adaptation snow detection and patch-based gaussian mixture model,

    B. Yang, Z. Jia, J. Yang, and N. K. Kasabov, “Video snow removal based on self-adaptation snow detection and patch-based gaussian mixture model,” IEEE Access , vol. 8, pp. 160 188–160 201, 2020

  3. [11]

    Video desnowing and deraining based on matrix decomposition,

    W. Ren, J. Tian, Z. Han, A. Chan, and Y . Tang, “Video desnowing and deraining based on matrix decomposition,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017, pp. 2838– 2847

  4. [12]

    Online rain/snow removal from surveillance videos,

    M. Li, X. Cao, Q. Zhao, L. Zhang, and D. Meng, “Online rain/snow removal from surveillance videos,” IEEE Transactions on Image Pro- cessing, vol. 30, pp. 2029–2044, 2021

  5. [13]

    Video rain/snow removal by transformed online multiscale convolutional sparse coding,

    M. Li, X. Cao, Q. Zhao, L. Zhang, C. Gao, and D. Meng, “Video rain/snow removal by transformed online multiscale convolutional sparse coding,” ArXiv, vol. abs/1909.06148, 2019. [Online]. Available: https://api.semanticscholar.org/CorpusID:202572625

  6. [14]

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

    J.-H. Kim, J.-Y . Sim, and C.-S. Kim, “Video deraining and desnowing using temporal correlation and low-rank matrix completion,” IEEE Transactions on Image Processing , vol. 24, no. 9, pp. 2658–2670, 2015

  7. [15]

    Event-based vision: A survey,

    G. Gallego, T. Delbruck, G. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. Davison, J. Conradt, K. Daniilidis, and D. Scara- muzza, “Event-based vision: A survey,”IEEE Trans. Pattern Anal. Mach. Intell., 2020

  8. [16]

    Dsec: A stereo event camera dataset for driving scenarios,

    M. Gehrig, W. Aarents, D. Gehrig, and D. Scaramuzza, “Dsec: A stereo event camera dataset for driving scenarios,” IEEE Robot. Autom. Lett. , 2021

  9. [17]

    Detection and removal of rain from videos,

    K. Garg and S. Nayar, “Detection and removal of rain from videos,” in Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004. , vol. 1, 2004, pp. I–I

  10. [18]

    Automatic single-image-based rain streaks removal via image decomposition,

    L.-W. Kang, C.-W. Lin, and Y .-H. Fu, “Automatic single-image-based rain streaks removal via image decomposition,” IEEE Transactions on Image Processing, vol. 21, no. 4, pp. 1742–1755, 2012

  11. [19]

    Single image deraining: From model-based to data-driven and beyond,

    W. Yang, R. T. Tan, S. Wang, Y . Fang, and J. Liu, “Single image deraining: From model-based to data-driven and beyond,” vol. 43, no. 11, 2021, pp. 4059–4077

  12. [20]

    Learning rain location prior for nighttime deraining,

    F. Zhang, S. You, Y . Li, and Y . Fu, “Learning rain location prior for nighttime deraining,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2023, pp. 13 148– 13 157

  13. [21]

    Pixel based temporal analysis using chromatic property for removing rain from videos

    P. Liu, J. Xu, J. Liu, and X. Tang, “Pixel based temporal analysis using chromatic property for removing rain from videos.” Computer and Information Science , vol. 2, no. 1, pp. 53–60, 2009

  14. [22]

    Rain removal in video by combining temporal and chromatic properties,

    X. Zhang, H. Li, Y . Qi, W. K. Leow, and T. K. Ng, “Rain removal in video by combining temporal and chromatic properties,” in Proceedings of the IEEE International Conference on Multimedia and Expo (ICME) , 2006, pp. 461–464

  15. [23]

    P. C. Barnum, S. Narasimhan, and T. Kanade, “Analysis of rain and 16 (a) Input Image (b) Ours - Image (c) Input-Depth (d) Ours - Depth Fig. 18: Qualitative results of our method on downstream task - depth estimation. We show the input image and depth map from the DSEC-Snow dat...

  16. [24]

    Rain or snow detection in image sequences through use of a histogram of orientation of streaks,

    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,” International Journal of Computer Vision , vol. 93, no. 3, pp. 348–367, 2011

  17. [25]

    Utilizing local phase information to remove rain from video,

    V . Santhaseelan and V . K. Asari, “Utilizing local phase information to remove rain from video,” International Journal of Computer Vision , vol. 112, no. 1, pp. 71–89, 2015

  18. [26]

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

    Y .-L. Chen and C.-T. Hsu, “A generalized low-rank appearance model for spatio-temporally correlated rain streaks,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV) , 2013, pp. 1968– 1975

  19. [27]

    A novel tensor-based video rain streaks removal approach via utilizing discriminatively intrinsic priors,

    T.-X. Jiang, T.-Z. Huang, X.-L. Zhao, L.-J. Deng, and Y . Wang, “A novel tensor-based video rain streaks removal approach via utilizing discriminatively intrinsic priors,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017, pp. 2818– 2827

  20. [28]

    Video desnowing and deraining based on matrix decomposition,

    W. Ren, J. Tian, Z. Han, A. Chan, and Y . Tang, “Video desnowing and deraining based on matrix decomposition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017, pp. 4210–4219

  21. [29]

    Video rain streak removal by multiscale convolutional sparse coding,

    M. Li, Q. Xie, Q. Zhao, W. Wei, S. Gu, J. Tao, and D. Meng, “Video rain streak removal by multiscale convolutional sparse coding,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2018, pp. 6644–6653

  22. [30]

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

    J. Chen, C.-H. Tan, J. Hou, L.-P. Chau, and H. Li, “Robust video content alignment and compensation for rain removal in a cnn framework,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2018, pp. 6286–6295

  23. [31]

    Erase or fill? deep joint recurrent rain removal and reconstruction in videos,

    J. Liu, W. Yang, S. Yang, and Z. Guo, “Erase or fill? deep joint recurrent rain removal and reconstruction in videos,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2018, pp. 3233–3242

  24. [32]

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

    ——, “D3r-net: Dynamic routing residue recurrent network for video rain removal,” IEEE Transactions on Image Processing , vol. 28, no. 2, pp. 699–712, 2018

  25. [33]

    Frame-consistent recurrent video de- raining with dual-level flow,

    W. Yang, J. Liu, and J. Feng, “Frame-consistent recurrent video de- raining with dual-level flow,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2019, pp. 1661– 1670

  26. [34]

    A semi-direct approach to structure from motion,

    H. Jin, P. Favaro, and S. Soatto, “A semi-direct approach to structure from motion,” The Visual Computer , vol. 19, no. 6, pp. 377–394, 2003

  27. [35]

    Semi-supervised video derain- ing with dynamical rain generator,

    Z. Yue, J. Xie, Q. Zhao, and D. Meng, “Semi-supervised video derain- ing with dynamical rain generator,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021

  28. [36]

    Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion,

    S. Tulyakov, A. Bochicchio, D. Gehrig, S. Georgoulis, Y . Li, and D. Scaramuzza, “Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion,” IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2022

  29. [37]

    Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion,

    ——, “Time lens++: Event-based frame interpolation with parametric non-linear flow and multi-scale fusion,” IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2022

  30. [38]

    Seeing behind dynamic occlusions with event cameras,

    R. Zou, M. Muglikar, N. Messikommer, and D. Scaramuzza, “Seeing behind dynamic occlusions with event cameras,” 2023

  31. [39]

    Unsupervised video deraining with an event camera,

    J. Wang, W. Weng, Y . Zhang, and Z. Xiong, “Unsupervised video deraining with an event camera,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2023, pp. 10 831–10 840

  32. [40]

    Photoshop, https://www.adobe.com/products/photoshop.html

    A. Photoshop, https://www.adobe.com/products/photoshop.html

  33. [41]

    Pytorch: An imperative style, high-performance deep learning library,

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: An imperative style, high-pe...

  34. [42]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” CoRR, vol. abs/1412.6980, 2014. [Online]. Available: https://api.semanticscholar.org/CorpusID:6628106

  35. [43]

    High speed and high dynamic range video with an event camera,

    H. Rebecq, R. Ranftl, V . Koltun, and D. Scaramuzza, “High speed and high dynamic range video with an event camera,” IEEE Trans. Pattern Anal. Mach. Intell. , 2019

  36. [44]

    Benchmarking single-image dehazing and beyond,

    B. Li, W. Ren, D. Fu, D. Tao, D. Feng, W. Zeng, and Z. Wang, “Benchmarking single-image dehazing and beyond,” IEEE Transactions on Image Processing , vol. 28, no. 1, pp. 492–505, 2018

  37. [45]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” ArXiv preprint, 2021

  38. [46]

    Benchmarking single-image dehazing and beyond,

    B. Li, W. Ren, D. Fu, D. Tao, D. Feng, W. Zeng, and Z. Wang, “Benchmarking single-image dehazing and beyond,” Trans. Img. Proc., vol. 28, no. 1, p. 492–505, Jan. 2019. [Online]. Available: https://doi.org/10.1109/TIP.2018.2867951

  39. [47]

    V2E: From video frames to realistic DVS event camera streams,

    T. Delbruck, Y . Hu, and Z. He, “V2E: From video frames to realistic DVS event camera streams,” arXiv e-prints , 2020

  40. [48]

    Finateu, A

    T. Finateu, A. Niwa, D. Matolin, K. Tsuchimoto, A. Mascheroni, E. Reynaud, P. Mostafalu, F. Brady, L. Chotard, F. LeGoff, H. Takahashi, H. Wakabayashi, Y . Oike, and C. Posch, “A 1280x720 back-illuminated stacked temporal contrast event-based vision sensor with 4.86 µm pixels,...

  41. [49]

    RAft: Recurrent All-Pairs Field Transforms for Optical Flow,

    Z. Teed and J. Deng, “RAft: Recurrent All-Pairs Field Transforms for Optical Flow,” in Eur . Conf. Comput. Vis. (ECCV) , 2020

  42. [50]

    Muses: The multi-sensor seman- tic perception dataset for driving under uncertainty,

    T. Br ¨odermann, D. Bruggemann, C. Sakaridis, K. Ta, O. Liagouris, J. Corkill, and L. Van Gool, “Muses: The multi-sensor seman- tic perception dataset for driving under uncertainty,” arXiv preprint arXiv:2401.12761, 2024

  43. [51]

    Event cameras meet spads for high-speed, low- bandwidth imaging,

    M. Muglikar, S. Somasundaram, A. Dave, E. Charbon, R. Raskar, and D. Scaramuzza, “Event cameras meet spads for high-speed, low- bandwidth imaging,” 2025, pp. 1–12

  44. [2019]

    Available: https://proceedings.neurips.cc/paper files/ paper/2019/file/bdbca288fee7f92f2bfa9f7012727740-Paper.pdf

    [Online]. Available: https://proceedings.neurips.cc/paper files/ paper/2019/file/bdbca288fee7f92f2bfa9f7012727740-Paper.pdf

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.