{"id":"59540980-0f6f-4628-ac7d-f32b1661ac95","arxiv_id":"2505.05391","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A decoupled spatial-temporal Mamba architecture denoises event camera streams with 88.9K parameters and state-of-the-art accuracy on four benchmarks.","lead":"EDmamba splits event-camera denoising into separate spatial and temporal state-space branches, achieving a reported 88.9K-parameter model that runs 36 times faster than a Transformer baseline while improving average accuracy. Generalists may care because it suggests event-vision preprocessing can be made fast and tiny enough for real-time systems without sacrificing quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on unverified baseline comparability; Table 2 footnotes show some numbers are copied rather than re-run, so the 2.1 pp margin may reflect protocol differences.","rationale":"The paper's central contribution is the claim that decoupling spatial and temporal state-space branches yields both higher accuracy and higher efficiency than joint 4D models. The efficiency part is supported by direct measurements of parameter count, FLOPs, and inference time, which are credible. The accuracy part, however, depends entirely on comparisons with published baselines. In Tables 1 and 2, the paper reports small margins (0.2–3%) over EDformer and EDnCNN. These margins are exactly the kind that can be inverted by differences in training data, event sampling, evaluation protocol, or checkpoint provenance. The paper explicitly notes that some MESR numbers are 'derived from E-MLB [25]' rather than re-run, contradicting the Figure 1 caption's claim that all methods were evaluated under identical settings. This makes the four-benchmark SOTA claim vulnerable. A controlled re-benchmark with identical training data and evaluation code would settle whether the decoupled architecture genuinely leads to SOTA accuracy. Until then, the verdict should remain conditional.","tokens_in":12019,"tokens_out":9273,"duration_ms":86698,"concrete_test":"Obtain the official code/checkpoints for EDformer, AEDNet, and EDnCNN; retrain each on the identical ED24 training split and event-sampling configuration used for EDmamba, then evaluate all methods with a single evaluation script on DND21 and DVSCLEAN under the protocol of [16]. If the average AUC gap between EDmamba and EDformer on DND21 shrinks below 1 percentage point or reverses on any of the four benchmarks, the headline 'state-of-the-art' claim loses its support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that decoupled spatial/temporal SSM branches achieve state-of-the-art accuracy on four public benchmarks—depends on the comparability of baseline numbers in Tables 1 and 2. The paper states (Figure 1 caption) 'All methods were evaluated under identical settings,' yet Section 4.2's footnotes say for GET, EventZoom, and IETS, 'The result is derived from E-MLB [25], as the official code is not publicly available,' and Pre-Mamba's efficiency is 'derived from Pre-Mamba [12].' Thus the MESR comparisons in Table 2 are a mix of self-run and transcribed results. If the cited MESR values were computed with a different evaluation script, event-volume normalization, or parameter settings, the small margins (e.g., 0.976 vs 0.996 on E-MLB Daylight ND1) could flip. For Table 1, the paper says AUC follows [16] but does not state which baselines were retrained on ED24, whether the same train/validation split was used, or whether any baseline numbers were taken from original papers. Without this, the 1.15–3.01% relative AUC improvements over EDformer could be attributable to protocol differences (e.g., EDformer's official model may have been trained on a different subset of ED24 or with different event sampling). The paper also reports no run-to-run variance or significance tests, so the margins are within plausible noise for event denoising benchmarks. Because the entire novelty argument ('decoupled design gives both accuracy and efficiency') rests on these comparisons, this is the load-bearing weak point.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes EDmamba, an event-denoising network that represents raw events as 4D point clouds, extracts coarse geometric and polarity features, and then routes them through two decoupled state-space branches: a Spatial-SSM for location-dependent artifacts such as hot pixels, and a Temporal-SSM for temporally inconsistent background activity. These branches are assembled in a U-Net-style encoder-decoder with shared Spatial-Temporal State Space Blocks. The paper reports 88.98K parameters, 2.27 GFLOPs, and 0.0685 s per 100K events on a single GPU, and claims state-of-the-art AUC on DND21 and DVSCLEAN plus strong MESR on E-MLB and DND21. Ablations on DND21 isolate the contributions of the geometric features, polarity features, and the two SSM branches, including a Joint-SSM control variant.","tokens_in":12350,"tokens_out":5507,"duration_ms":55109,"significance":"The core design idea is attractive and practically relevant: separating spatial and temporal denoising is a natural response to the different physical origins of hot-pixel artifacts and background-activity noise, and the reported parameter and latency figures are unusually low. The paper also includes a self-designed Joint-SSM control, which is a useful methodological step, and it evaluates on both labeled and unlabeled real-world benchmarks. If the comparisons against prior methods are made fully fair, reproducible, and statistically grounded, the result would be a meaningful advance for real-time event denoising. At present, however, the headline state-of-the-art claim is not yet supported because the baseline tables mix re-run and transcribed results, several reported differences are small, and no variance information is provided.","major_comments":[{"comment":"The claim that EDmamba establishes new state-of-the-art accuracy on four public benchmarks is not supported by Table 2: on E-MLB Daylight ND1, GET reaches 1.051 versus 0.976 for EDmamba; on Daylight ND16, TS reaches 0.980 versus 0.985; and on DND21 MESR, EventZoom reaches 1.059 versus 1.057. The authors should either qualify the state-of-the-art claim to specific benchmarks and metrics, or explain how these non-best values are compatible with the abstract and Section 1 statements.","section":"Section 4.2, Table 2"},{"comment":"The Figure 1 caption states that all methods were evaluated under identical settings, but Table 2 footnotes say that IETS, GET, and EventZoom results are derived from E-MLB [25] and Table 3 says Pre-Mamba efficiency is derived from Pre-Mamba [12]. The paper must state explicitly which baselines were re-run by the authors and which were transcribed from prior papers. For re-run baselines, the authors should report the training split, noise-injection protocol, event-sampling scheme, evaluation script, and hardware configuration. Without this information, the 1.15-3.01% relative AUC improvements over EDformer in Table 1 and the small MESR margins in Table 2 could plausibly be caused by protocol differences rather than by the proposed method.","section":"Section 4.1, Section 4.2, Tables 1-2, Figure 1"},{"comment":"The efficiency comparison is not fully coherent. The 'Rel. Speed' column uses TS as the 1.0x reference for filtering methods but EDnCNN as the 1.0x reference for learning-based methods, so the relative speed values cannot be compared across the two groups. In addition, Pre-Mamba's inference time is 'derived' rather than measured, and no details are given on whether the baseline runtimes were obtained with the same operator-level implementation, batch size, and memory configuration. The 36x speedup claim over EDformer should be supported with measured runtimes under a common benchmarking protocol, or explicitly labeled as an estimate.","section":"Section 4.3, Table 3"},{"comment":"Ablations are reported for a single noise level (5 Hz/pixel) and without error bars or multiple seeds. Given that several differences are only 0.22-0.36 AUC points (e.g., w/o Geometry Feat. on Hotel-bar), the robustness of the ablation conclusions is uncertain. The same lack of variance information applies to the headline comparisons in Tables 1 and 2, where the reported margins are often smaller than typical run-to-run variation for event-denoising benchmarks.","section":"Section 4.4, Table 4"}],"minor_comments":[{"comment":"Table 1 contains a formatting error: several DVSCLEAN entries are concatenated (e.g., '0.78270.93940.76790.93010.7615'), making the table difficult to read.","section":"Table 1"},{"comment":"The learning rate is described as '8 x 10^-5 per sample'; please clarify whether this is a per-sample or per-batch learning rate and how it interacts with the batch size of 128.","section":"Section 4, Implementation Details"},{"comment":"The text says one event is randomly sampled from each non-empty voxel during training; please state what sampling is used at inference time, since random sampling would introduce run-to-run variation in the reported metrics.","section":"Section 3.2, Event Sampling"},{"comment":"Figure 1 has a legend with 'Not mentioned' appearing twice, which leaves the marker encoding of parameter counts unclear; a cleaner legend would improve readability.","section":"Figure 1"},{"comment":"The abstract uses '2.1 percentage points' while Section 1 says '2.08% improvement to denoising accuracy'; these should be reconciled and the relative versus absolute distinction made explicit.","section":"Abstract and Section 1"},{"comment":"The paper does not state whether code or trained models will be released; providing them would substantially improve reproducibility, especially for baseline numbers that are currently derived from the literature.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the most serious issue is the provenance of the baseline numbers. I recommend asking the authors to provide a detailed table indicating which entries in Tables 1-3 were re-run, which were taken from original papers, and which were derived. The self-cited Pre-Mamba efficiency number is particularly worth checking because it comes from the same group and is used in the speedup calculation. I do not see evidence of fabrication, but the reporting needs to be stricter before the state-of-the-art claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid architecture paper with a real efficiency story, but the 'state-of-the-art on four benchmarks' claim leans on baseline numbers that weren't all re-run, so the headline margins are softer than they look.\n\nWhat's new: decoupling spatial and temporal SSM branches for event denoising is a genuine design choice, and the 88.9K parameter / 2.27 GFLOP / 0.0685s per 100K events numbers are striking if they hold. The ablation and the Joint-SSM control make the decoupling argument credible: removing either branch hurts, and the joint variant does worse with more parameters. That's real evidence the architecture, not just scale, is doing the work.\n\nThe soft spots are concentrated in the evaluation, exactly where the stress-test note lands. The paper says 'all methods were evaluated under identical settings,' but Table 2's footnotes admit that GET, EventZoom, and IETS numbers are 'derived from E-MLB' because official code isn't available. Table 1 likewise follows the protocol of [16] but never says which baselines were retrained on ED24, what splits were used, or whether any numbers were transcribed from original papers. On top of that there are no error bars or significance tests, which matters when the claimed margins over EDformer are 1-3% relative AUC. On E-MLB, the paper is not uniformly best anyway - EDnCNN and EDformer win several cells. So the 2.1 percentage point headline is contingent on protocol equivalence that the paper doesn't demonstrate.\n\nThe physical noise model in Eq. 3 is motivational, not load-bearing; the classifier f_theta in Eq. 4 is not fully specified. I don't treat those as fatal - they're typical for this kind of paper - but they're minor overreach if the authors want to claim a derivation-level contribution.\n\nThe citation pattern is fair: the self-cited Pre-Mamba is flagged as 'derived,' and it's used as an efficiency baseline, not to define the loss. That's acceptable.\n\nBottom line: this deserves a serious referee. The architecture is clean, the efficiency numbers are worth checking, and the ablation supports the central design claim. But the referee should ask for code, a re-run or clear provenance of every baseline, and error bars before the SOTA sentence can stand. I'd also ask the authors to soften 'four public benchmarks' to what's actually supported.\n\nRecommendation: send to peer review, conditional on evaluation transparency demands. It's a useful paper for the event-camera community and for anyone building real-time preprocessing; the claimed gains are plausible but need verification.","headline":"Decoupled spatial/temporal SSM design is genuinely new and the efficiency numbers are striking, but the SOTA claim rests on baseline numbers that were partly transcribed, not re-run.","tokens_in":12904,"tokens_out":3449,"would_cite":true,"duration_ms":30868,"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":"Spatial and temporal event noise can be suppressed independently, and a decoupled state-space network does it faster and more accurately than joint 4D models.","keywords":["event camera denoising","state space models","Mamba","spatiotemporal decoupled architecture","background activity noise","hot pixel noise","event stream processing","real-time efficiency"],"falsifier":"Re-run the Table 1 comparison from scratch: train every learning-based baseline and EDmamba on the same ED24 split, evaluate each on DND21 and DVSCLEAN with multiple seeds, and report the spread; if the best baseline's mean AUC overlaps EDmamba's within one standard deviation, the claimed accuracy lead is not established.","tokens_in":11823,"feed_emoji":"⚡","tokens_out":5720,"duration_ms":54420,"temperature":0.7,"pith_summary":"This paper tries to establish that event-camera noise separates into spatially structured artifacts (like hot pixels and leakage) and temporally inconsistent firings (background activity), so denoising can be done by two independent lightweight state-space branches instead of one joint 4D model. If true, the practical payoff is large: the network is small enough (88.98K parameters, 2.27 GFLOPs) and fast enough (100K events in 68.5ms) to run in real time, while matching or beating prior Transformer and CNN denoisers in accuracy. The paper reports state-of-the-art AUC on four public benchmarks, outscoring the strongest prior model by about 2.1 percentage points. A sympathetic reader takes this as evidence that architectural specialization to noise physics can beat simply scaling a joint spatiotemporal backbone.","feed_headline":"Two Mamba branches clean event streams 36x faster than transformers","feed_subtitle":"Decoupled spatial and temporal state-space models beat joint 4D denoising with 88.98K parameters.","key_machinery":"The central object is EDmamba, a U-Net-style encoder-decoder whose core is a pair of decoupled Mamba branches. The Spatial-SSM flattens the spatial domain with space-filling curves (Z-order and Hilbert) to preserve local geometry while processing sequences with bidirectional state updates, so fixed-pattern noise is suppressed through learned spatial smoothing. The Temporal-SSM sorts events by normalized timestamp and runs a bidirectional Mamba block to capture forward and backward motion patterns, exploiting the causal consistency of real motion versus the acausal randomness of temporal noise. A coarse feature extraction module separates each event's representation into geometry and polarity streams before the branches, and the two branches interact through shared spatial-temporal state-space blocks rather than entangled attention.","core_discovery":"The paper claims that spatial and temporal event noise originate from different physical mechanisms and can therefore be suppressed independently, and that acting on this separation improves both accuracy and efficiency. Its proposed network, EDmamba, processes raw events as a 4D event cloud $(x, y, t, p)$ using a coarse feature extractor that encodes geometry and polarity separately, then routes the features through two decoupled Mamba branches: a Spatial-SSM that learns location-conditioned filters to silence persistent spatial artifacts, and a Temporal-SSM that models causal motion continuity to eliminate bursty background events. The paper reports that this decoupled design reaches higher AUC than prior models on DND21 and most DVSCLEAN sequences, while using far fewer parameters and running 36 times faster than a Transformer-based denoiser.","pith_inferences":["The same decoupling insight likely transfers to other event-stream tasks where spatial artifacts and temporal jitter have distinct causes, such as event-based tracking, SLAM preprocessing, or high-speed obstacle avoidance, where removing background activity before downstream perception could improve robustness.","Because the spatial and temporal branches rely on linear state-space scans rather than quadratic attention, a testable extension is deploying the model on low-power neuromorphic or embedded hardware to check whether the 36x speed advantage over Transformers persists at the edge.","A natural experimental follow-up would vary the relative strength of spatial versus temporal noise in a controlled synthetic dataset and measure whether the two branches degrade independently; the paper's current ablations only remove whole branches, not manipulate noise types separately.","The paper's logic suggests that adding a third specialized branch for polarity-flip noise, rather than folding polarity into the coarse features, could yield further gains, though this goes beyond what the authors tested."],"forward_implications":["If the central claim is correct, event denoising no longer needs heavy spatiotemporal attention: a small decoupled state-space network can reach or beat prior accuracy while processing 100K events in 68ms, making denoising practical for real-time UAV and autonomous-driving pipelines.","The decoupling principle appears to generalize across noise levels, since the reported gains hold for shot noise rates of 1-10 Hz/pixel on DND21 and for both 50% and 100% noise injection on DVSCLEAN.","Ablations removing either branch degrade accuracy, with the Temporal-SSM contributing more in motion-heavy scenes, so interpreting temporal modeling as motion-continuity enforcement is a direct corollary of the paper's experiments.","The reported MESR gains on unlabeled real-world datasets imply the denoising behavior transfers beyond synthetic training data to daylight and nighttime conditions without requiring labels.","The Joint-SSM control variant, which shares one Mamba block for both spatial and temporal sequences, needs more parameters and runs slower while achieving lower accuracy, supporting the paper's claim that noise-specific specialization is the source of the gain."],"supporting_citations":[{"why":"The Transformer-based EDformer provides the main accuracy and latency baseline, and its ED24 dataset is used for supervised training.","marker":"[10]"},{"why":"Supplies the DND21 dataset and the AUC evaluation protocol that the paper follows for quantitative comparison.","marker":"[16]"},{"why":"Provides the AEDNet PointNet-based method and the DVSCLEAN synthetic dataset used for evaluation.","marker":"[9]"},{"why":"The CNN-based EDnCNN is one of the strongest learning-based baselines and the closest competitor on several DVSCLEAN sequences.","marker":"[8]"},{"why":"The Mamba selective state-space model is the building block for the paper's Spatial-SSM and Temporal-SSM branches.","marker":"[11]"},{"why":"Pre-Mamba is the prior 4D event state-space model that the paper compares against to show the efficiency cost of joint spatiotemporal modeling.","marker":"[12]"},{"why":"Provides the E-MLB dataset and the MESR metric used for label-free real-world generalization evaluation.","marker":"[25]"}],"fun_headline_variants":["Decoupled Mamba SSMs clean event streams with 88.9K params","Spatial and temporal Mamba branches cut denoising cost 36x","EDmamba: Event denoising via independent spatial-temporal SSMs","Two lightweight Mamba streams beat 4D transformers by 2.1 points","Event denoising decoupled: 88.9K params, 36x faster, SOTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art claim rests on the assumption that every comparison number in Table 1 came from identically run, fair evaluations, and the paper gives no code, no variance estimates, and no statement about whether non-learned baselines were re-run, so a reader cannot yet check whether the reported 1-3% margins exceed run-to-run noise.","fun_headline_variants_meta":{"raw":{"variants":["Decoupled Mamba SSMs clean event streams with 88.9K params","Spatial and temporal Mamba branches cut denoising cost 36x","EDmamba: Event denoising via independent spatial-temporal SSMs","Two lightweight Mamba streams beat 4D transformers by 2.1 points","Event denoising decoupled: 88.9K params, 36x faster, SOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1458,"prompt_tokens":933,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":416}},"tokens_in":549,"tokens_out":525,"duration_ms":5015,"temperature":1.0,"reasoning_tokens":416,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:05:11.983367+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Table 1 comparison from scratch: train every learning-based baseline and EDmamba on the same ED24 split, evaluate each on DND21 and DVSCLEAN with multiple seeds, and report the spread; if the best baseline's mean AUC overlaps EDmamba's within one standard deviation, the claimed accuracy lead is not established.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Transformer-based EDformer provides the main accuracy and latency baseline, and its ED24 dataset is used for supervised training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DND21 dataset and the AUC evaluation protocol that the paper follows for quantitative comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the AEDNet PointNet-based method and the DVSCLEAN synthetic dataset used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The CNN-based EDnCNN is one of the strongest learning-based baselines and the closest competitor on several DVSCLEAN sequences."},{"cited_title":"PRE-Mamba: A 4D State Space Model for Ultra-High-Frequent Event Camera Deraining","cited_arxiv_id":"2505.05307","evidence_quote":"Pre-Mamba is the prior 4D event state-space model that the paper compares against to show the efficiency cost of joint spatiotemporal modeling."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the E-MLB dataset and the MESR metric used for label-free real-world generalization evaluation."}],"review_version":1}