{"id":"b6837f48-4d25-4714-9428-34238fbc5d41","arxiv_id":"2505.05307","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PRE-Mamba is the first point-based event deraining framework, using a 4D event cloud and a multi-scale state space model to classify and remove rain events while preserving temporal precision.","lead":"PRE-Mamba removes rain streaks from event camera data using a lightweight point-based state space model, claiming faster and more accurate deraining than prior frame-based approaches. It also introduces EventRain-27K, a new benchmark dataset for event camera deraining under synthetic and real rain.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim depends on unvalidated EventRain-27K ground-truth labels; the synthetic label-generation step is unspecified and the KNN alignment used for the self-recorded data can systematically tag background events as rain.","rationale":"The reader's weakest assumption is exactly the reliability of the EventRain-27K labels, and my review confirms that this is the most load-bearing point. The architecture, efficiency measurements, and qualitative results are plausible and the paper gives credit for releasing code and dataset, but the quantitative SOTA claim rests entirely on supervised labels whose generation is under-specified in Section 4. The synthetic label-generation step is not described at all, and the KNN alignment for the artificial data is described in one sentence with no thresholds, no validation, and no sensitivity analysis. Since SR/NR/DA are per-event classification metrics, any systematic bias in the labels transfers directly to the reported numbers. The internal inconsistency between the abstract's dataset counts (18K labeled + 9K real) and Section 4's counts (over 7K + over 7K + over 9K) further weakens reproducibility. The proposed oracle-label comparison is a concrete, executable check that would settle whether this concern lands. Because the reader already made acceptance conditional on this issue, my read does not move the verdict; it sharpens the condition that must be met before accepting the SOTA claim.","tokens_in":16521,"tokens_out":5541,"duration_ms":61096,"concrete_test":"Take a held-out subset of the synthetic pipeline where the rendered rain mask is available before Vid2E. Build oracle per-event labels by running Vid2E separately on the clean video and on the rain-streak layer, then compare these oracle labels with the EventRain-27K labels used for training/evaluation. Report per-event precision and recall of the rain class. If rain-class precision or recall is below about 0.9, or if the error is strongly correlated with rain density, retrain and re-evaluate PRE-Mamba and the baselines on the oracle labels; if the reported SR/NR/DA values or the relative ranking shift materially, the SOTA claim does not survive without dataset relabeling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PRE-Mamba's headline numbers (0.95 SR, 0.91 NR, 0.93 DA) and the SOTA comparison are computed as per-event classification accuracy on EventRain-27K. Section 4 is the load-bearing point: for the synthetic data it says only that clean KITTI/SPAC videos were made rainy and then passed through Vid2E; it never states how the per-event rain/background ground truth is derived. The natural reading is that labels come from differencing Vid2E outputs of rainy and clean videos, but contrast-threshold event generation makes such a difference invalid: rain alters background contrast, shifting or suppressing background events, so a simple difference mislabels occlusion/contrast-change events as rain and also misses some background events that should be retained. For the self-recorded artificial data, Section 4 says background events were identified by KNN spatiotemporal alignment with rain-free captures. With the slight camera vibration used during capture, many background events in the rainy take will have no close neighbor in the rain-free take and will be labeled rain; foreground droplets also occlude background, adding further label noise. No validation of label accuracy is provided. Because SR/NR/DA are computed against these labels, systematic mislabeling—especially labeling background as rain—would directly inflate noise-removal scores and distort the baseline comparison. There is also an internal size mismatch: the abstract says 18K labeled synthetic plus 9K real sequences, while Section 4 lists over 7K synthetic, over 7K artificial, and over 9K real samples, which impedes independent verification of the dataset.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PRE-Mamba, a point-based event-camera deraining framework that operates directly on raw event streams via a 4D event cloud representation, a Spatio-Temporal Decoupling and Fusion module (STDF), a Multi-Scale State Space Model (MS3M), and a frequency-domain regularization loss. The authors construct EventRain-27K, a dataset with synthetic, self-recorded artificial, and real-world rain event sequences, and report state-of-the-art per-event classification accuracy (average SR/NR/DA of 0.95/0.91/0.93) with low parameter count and high inference speed. The paper also claims generalization to snow. The central claim is empirical: the proposed architecture outperforms adapted event-denoising baselines on the newly introduced dataset.","tokens_in":16837,"tokens_out":2712,"duration_ms":25766,"significance":"If the results hold, the paper is a useful contribution: it is the first point-based event deraining framework, introduces a public dataset and code, and demonstrates that a selective state-space model can process raw event streams at scale with linear complexity. The architecture is described in sufficient detail to be reimplemented, and the ablations isolate the contribution of each proposed component. The efficiency numbers are attractive for embedded and real-time applications. The main significance is contingent on the reliability of the EventRain-27K ground-truth labels, because the headline SR/NR/DA metrics and the SOTA comparison are computed directly against those labels; the current manuscript does not provide enough evidence that the labels are trustworthy.","major_comments":[{"comment":"The labeling protocol for the synthetic dataset is unspecified. The text states only that clean KITTI/SPAC videos were made rainy and then passed through Vid2E, but it never states how per-event rain/background labels are derived. If the labels are obtained by differencing the event streams of rainy and clean videos, the procedure is invalid because rain alters the effective scene contrast, shifting or suppressing background events; the resulting labels would systematically misclassify occlusion and contrast-change events as rain, directly inflating SR and NR. The authors must specify the exact label-generation rule, and if differencing is used, they must either justify it with a physical model or replace it with a valid procedure.","section":"Section 4, Synthetic Dataset"},{"comment":"The KNN-based spatiotemporal alignment used to label background events in the self-recorded artificial data is not validated. Since the authors deliberately applied slight camera vibration during capture, many background events in the rainy take will have no close spatiotemporal neighbor in the rain-free take and will be labeled as rain; foreground droplets also occlude background, adding further label noise. Because Table 1 reports per-event accuracy against these labels, systematic mislabeling would distort the absolute numbers and, more importantly, the comparison with baselines. Please provide a quantitative label-quality assessment (e.g., human-annotated subset, consistency analysis, or an alternative alignment method) and discuss how label noise affects the reported SR/NR/DA.","section":"Section 4, Self-recorded Artificial Dataset"},{"comment":"The dataset size is internally inconsistent. The abstract and introduction state that EventRain-27K comprises 18K labeled synthetic and 9K unlabeled real-world sequences, totaling 27K. Section 4 instead says the dataset contains over 7K self-generated synthetic samples, over 7K self-recorded artificial samples, and over 9K real rain samples, totaling approximately 23K. This discrepancy must be resolved, as the reader cannot tell how many labeled sequences were actually used for training and evaluation.","section":"Abstract and Section 4"},{"comment":"The efficiency claim is numerically inconsistent. The abstract reports 0.4s/M events, and the text says the model scales to 1M events in 0.398s, but Table 2 reports 0.0987s per 100K events, which is 0.987s per 1M events. Moreover, the 'Relative speed' column is confusing: EDnCNN is assigned 1.0x while TS, which is 155x faster than EDnCNN by inference time, is also assigned 1.0x. Please correct these numbers and define the reference method clearly.","section":"Section 5.2, Table 2 and Abstract"},{"comment":"The comparison protocol is under-specified. The paper says only that EDnCNN was retrained with the authors' ground-truth labels; it does not state whether AEDNet and EDformer were likewise retrained on EventRain-27K or used with their pretrained weights. If the latter, the comparison is not a controlled evaluation of deraining architecture quality, because the baselines were trained on different label distributions and may simply be mismatched to the test set. Please report the exact training protocol for every learning-based baseline and, ideally, provide per-method confidence intervals across multiple runs.","section":"Section 5.2, Baselines"}],"minor_comments":[{"comment":"The frequency loss divides by max(|F(P)-F(Y)|, epsilon) but it is not clear what happens when the predicted and ground-truth spectra are both zero at a particular frequency; please clarify the epsilon handling and the choice of the L1/L2 normalization.","section":"Equation (6)"},{"comment":"The 'Relative speed' column should state the reference method and whether the ratios are derived from the reported inference times; currently the entries appear mutually inconsistent.","section":"Table 2"},{"comment":"The caption contains a typo: 'isualizes' should be 'visualizes'.","section":"Figure 7 caption"},{"comment":"The snow generalization claim is supported only by qualitative examples. While this is acceptable as a supplementary demonstration, the text should explicitly say that no quantitative snow evaluation was performed, so readers are not misled.","section":"Section 5.4"}],"recommendation":"major_revision","confidential_remarks":"The core concern is that the dataset labeling protocol—the foundation of all quantitative claims—is not adequately specified or validated. The synthetic label-generation step in particular appears to be glossed over, and the KNN alignment for the self-recorded data is likely to inject systematic label noise. I would like to see the authors make the labeling code and label-quality metrics public as part of the revision. There is also a discrepancy between the abstract's 18K+9K and Section 4's 7K+7K+9K composition that needs to be resolved before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe one thing you should know: this is the first point-based event camera deraining method, and it ships a new dataset, so it will get attention. The architecture is a reasonable adaptation of Mamba to point events: a 4D event cloud with dual temporal scales, spatio-temporal decoupling, multi-scale SSM, and a frequency loss. The ablations show each piece contributes, and the efficiency numbers (0.26M params, ~0.4s/M events) are credible and useful. That part is solid.\n\nThe soft spot is the ground truth, and the stress-test note is right about it. The SOTA claim (0.95 SR, 0.91 NR, 0.93 DA) is per-event classification accuracy on EventRain-27K, and the labels are load-bearing. For the synthetic data, the paper says rainy KITTI/SPAC videos are run through Vid2E, but never explains how per-event rain/background labels are derived. The natural reading—differencing events from rainy and clean videos—is invalid because rain changes local contrast and shifts/suppresses background events, so a simple difference mislabels contrast-change events as rain. For the self-recorded artificial data, the KNN spatiotemporal alignment with rain-free captures is vulnerable to the slight camera vibration the authors mention: many background events in the rainy take will lack a close neighbor and be tagged as rain. No validation of label accuracy is provided. If labels systematically mark background as rain, the noise-removal numbers are inflated and the baseline comparison is distorted. There's also a size mismatch: the abstract says 18K labeled synthetic sequences, but Section 4 describes over 7K synthetic plus over 7K artificial labeled samples. This needs to be fixed.\n\nA second issue: the baseline comparison is not fully controlled. Only EDnCNN is explicitly retrained with the new labels; AEDNet and EDformer appear to be used off-the-shelf or with their own training, so the comparison is not apples-to-apples. The real-world and snow results are purely qualitative, which is fine as supporting evidence but should not be oversold.\n\nThe central idea holds up: point-based deraining is worth pursuing, and the efficiency claim is plausible. But as written, the empirical contribution is conditional on label validity. If the authors can clarify the label generation, add a validation (even a small human-annotated subset), and retrain all baselines, this would be a genuinely useful paper.\n\nMy recommendation: send it to peer review. It deserves referee time, and the label concern is fixable. But I would not cite the headline numbers until the labels are confirmed.","headline":"A promising first point-based event deraining framework with a new benchmark, but the headline numbers rest on unvalidated synthetic labels and the SOTA comparison is not fully controlled.","tokens_in":17376,"tokens_out":3685,"would_cite":false,"duration_ms":32434,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PRE-Mamba claims that event-camera deraining can be done directly on raw event points, reporting 0.95 signal retention, 0.91 noise removal, and 0.4 seconds per million events on the new EventRain-27K benchmark.","keywords":["event camera deraining","point-based event processing","state space model","4D event cloud","spatio-temporal decoupling and fusion","frequency-domain regularization","rain noise removal","event dataset"],"falsifier":"Take a held-out set of real rainy events recorded with a synchronized high-speed camera, manually label every event as rain or background by checking the optical ground truth, and run PRE-Mamba on it. If its SR/NR/DA falls to near chance or well below the reported 0.95/0.91 while the same labels look correct to human inspectors, the central claim is falsified.","tokens_in":16362,"feed_emoji":"🌧️","tokens_out":6969,"duration_ms":67217,"temperature":0.7,"pith_summary":"The paper is trying to establish that rain noise in event cameras can be removed without sacrificing the sensor's native temporal precision. Its claim is that a point-based network operating on raw events outdoes frame- or voxel-based deraining and denoising methods on a new benchmark, with only 0.26M parameters and linear-time processing. A sympathetic reading: if true, event cameras become usable in rain without giving up the microsecond-resolution, asynchronous advantages that make them attractive for high-speed robotics and autonomous driving. The method's three load-bearing ideas are a 4D event-cloud representation that keeps two temporal scales, a spatiotemporal decoupling and fusion module, and a multi-scale state-space backbone, regularized in the frequency domain.","feed_headline":"Rain-proofing event cameras at 0.4 seconds per million events","feed_subtitle":"A tiny point-based model removes rain from raw event streams while keeping microsecond timing.","key_machinery":"The central object is the 4D Event Cloud, $(x,y,z,T_n,p)$, which serializes raw events along z-order and Hilbert curves so a selective state-space model can ingest them while preserving both intra-window (microsecond) and inter-window (global) temporal structure. Around it, the Spatio-Temporal Decoupling and Fusion module (STDF) extracts spatial features and modulates them by intra- and inter-window temporal features; the Multi-Scale State Space Model (MS3M) runs parallel appearance and motion branches, adds a multi-scale spatial pathway, and applies a cross-product fusion before the SSM scan. The frequency-domain regularizer $L_{\\text{fft}}$ aligns the FFT amplitude and phase of predicted and ground-truth event labels, enforcing physically consistent rain patterns. Together these carry the claim by showing that each component contributes to the final DA score in ablations.","core_discovery":"PRE-Mamba treats event deraining as per-event binary classification: each event is labeled rain or background, rather than reconstructing a clean frame. The central discovery is that this classification works better when each event is represented as a point in a 4D cloud $(x,y,z,T_n,p)$, where $z$ is time normalized inside a fixed window and $T_n$ is the window index, and processed by a state-space model with separate intra-window and inter-window branches. The paper reports average SR/NR/DA of 0.95/0.91/0.93 on its EventRain-27K test sets, processes 0.4 seconds per million events with 0.26M parameters, and shows qualitative generalization across rain intensities, viewpoints, and snow.","pith_inferences":["Beyond the paper: if the per-event labels are reliable, the same 4D-cloud-plus-state-space recipe should transfer to other sparse weather noise such as hail, spray, or dust, where the spatiotemporal signature differs from ordinary sensor noise.","Beyond the paper: the dual-temporal representation may be useful outside deraining, for example as a general input format for event-based segmentation or tracking where intra-window and inter-window motion both matter.","Beyond the paper: a strong practical test would be measuring downstream task performance (tracking, SLAM, or object detection) on derained real rainy events, since SR/NR/DA on labeled synthetic data may not capture true utility.","Beyond the paper: the frequency-domain loss on event labels could be adapted to any sparse point-classification problem with imbalanced positive/negative ratios, not just rain."],"forward_implications":["Rain removal can be done event-by-event on raw streams, so downstream high-speed vision consumes cleaned events without frame conversion.","At 0.26M parameters and 0.4 seconds per million events, the approach fits resource-constrained and near-real-time settings.","Because performance degrades gracefully from 5 to 150 mm/h rain and transfers to snow, a single model may cover varied weather rather than requiring per-intensity retraining.","The linear-complexity backbone avoids the quadratic attention bottleneck that blocks point transformers at high event rates.","EventRain-27K supplies labeled synthetic, artificial, and real sequences, giving subsequent point-based deraining methods a common training and evaluation ground."],"supporting_citations":[{"why":"Supplies the linear-time selective state-space backbone that gives MS3M its efficiency.","marker":"[32]"},{"why":"Provides the motion-aware dual-temporal design that the inter-window branch is adapted from.","marker":"[70]"},{"why":"Converts rainy videos into event streams, producing training labels for the synthetic portion of EventRain-27K.","marker":"[81]"},{"why":"Supplies clean driving videos used as base content for synthetic rainy sequences.","marker":"[77]"},{"why":"Adds diverse clean video scenes that widen the synthetic rain training distribution.","marker":"[78]"},{"why":"Adds physically modeled rain streaks to clean videos before event conversion.","marker":"[79]"},{"why":"Supplements the rain-rendering pipeline for the synthetic dataset.","marker":"[80]"},{"why":"Transformer-based event denoiser used as a baseline; its quadratic attention motivates the linear-complexity design.","marker":"[66]"},{"why":"Learning-based denoising baseline whose labels are replaced with EventRain-27K's for a fair comparison.","marker":"[86]"},{"why":"Voxel-based event deraining method used for qualitative comparison and as the frame-based contrast.","marker":"[11]"}],"fun_headline_variants":["Point-based state space model clears rain from event cameras","4D event cloud plus state space model derains cameras in 0.4s/M events","Tiny 0.26M-param model derains event streams at 0.4s per M events","Per-event rain detection: 4D state space model for event cameras"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison depends on the ground-truth labels that say which events are rain and which are background; if those labels are wrong or systematically biased, the reported accuracy numbers do not mean what they appear to mean.","fun_headline_variants_meta":{"raw":{"variants":["Point-based state space model clears rain from event cameras","4D event cloud plus state space model derains cameras in 0.4s/M events","Tiny 0.26M-param model derains event streams at 0.4s per M events","Per-event rain detection: 4D state space model for event cameras"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001441,"raw_usage":{"total_tokens":5788,"prompt_tokens":910,"completion_tokens":4878,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":4789}},"tokens_in":526,"tokens_out":4878,"duration_ms":33547,"temperature":1.0,"reasoning_tokens":4789,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:07:05.947564+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of real rainy events recorded with a synchronized high-speed camera, manually label every event as rain or background by checking the optical ground truth, and run PRE-Mamba on it. If its SR/NR/DA falls to near chance or well below the reported 0.95/0.91 while the same labels look correct to human inspectors, the central claim is falsified.","supporting_citations":[{"cited_title":"Mambamos: Lidar-based 3d moving object segmentation with motion- aware state space model","cited_arxiv_id":null,"evidence_quote":"Provides the motion-aware dual-temporal design that the inter-window branch is adapted from."},{"cited_title":"Video to events: Recycling video datasets for event cameras","cited_arxiv_id":null,"evidence_quote":"Converts rainy videos into event streams, producing training labels for the synthetic portion of EventRain-27K."},{"cited_title":"Robust video content alignment and compensation for rain removal in a cnn framework","cited_arxiv_id":null,"evidence_quote":"Adds diverse clean video scenes that widen the synthetic rain training distribution."},{"cited_title":"Rain rendering for evaluating and improving robustness to bad weather","cited_arxiv_id":null,"evidence_quote":"Adds physically modeled rain streaks to clean videos before event conversion."},{"cited_title":"Adobe after effects software, 2023","cited_arxiv_id":null,"evidence_quote":"Supplements the rain-rendering pipeline for the synthetic dataset."},{"cited_title":"Edformer: Transformer-based event denois- ing across varied noise levels","cited_arxiv_id":null,"evidence_quote":"Transformer-based event denoiser used as a baseline; its quadratic attention motivates the linear-complexity design."},{"cited_title":"Event Probability Mask (EPM) and Event Denoising Convolutional Neural Network (EDnCNN) for Neuromorphic Cameras","cited_arxiv_id":"2003.08282","evidence_quote":"Learning-based denoising baseline whose labels are replaced with EventRain-27K's for a fair comparison."},{"cited_title":"Distill drops into data: Event- based rain-background decomposition network","cited_arxiv_id":null,"evidence_quote":"Voxel-based event deraining method used for qualitative comparison and as the frame-based contrast."}],"review_version":1}