{"id":"c89136af-873c-4f65-aaff-93da12b32bfd","arxiv_id":"2505.11232","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"AW-GATCN uses adaptive segmentation, multi-factor edge weighting, and graph attention to remove noise from event data and improve object recognition accuracy on four benchmarks.","lead":"This paper introduces a graph-based neural network that denoises event camera data before classifying objects. It reports accuracy gains of up to 8.79% over prior graph methods and up to 6.26% over Euclidean-based approaches on four event-based benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The variance-maximizing threshold in Algorithm 2 is not shown to separate signal from noise; Table III reports recognition accuracy, not noise-removal quality, and the δ loop is under-specified.","rationale":"This pass agrees with the reader that the adaptive denoising heuristic is the weakest link, but sharpens it in two ways. First, Table III's 17-20% 'improvement' is a change in recognition accuracy, not a direct measure of noise-removal quality; a graph sparsification control is missing. Second, Algorithm 2 is not fully specified because edge weights are real-valued and no discretization step for δ is given, so the exact threshold cannot be reproduced from the text. These concerns do not contradict the reported accuracies; the gains could be real. The paper deserves credit for a clear ablation, consistent gains, and fixed parameter values, so conditional acceptance is appropriate. A synthetic-noise precision/recall test would settle whether the mechanism is genuinely denoising or primarily sparsifying, and would also resolve the algorithmic ambiguity by forcing a well-defined discretization.","tokens_in":11116,"tokens_out":6951,"duration_ms":73689,"concrete_test":"Create synthetic corrupted event streams by taking clean N-Caltech101 sequences and adding a known amount of background noise so every event has a ground-truth signal/noise label. Reproduce the full pipeline, implementing Algorithm 2 with δ ranging over the sorted distinct edge weights in each voxel (the only well-defined finite reading). For each voxel, record whether the selected T retains signal and noise equally, and compute precision/recall of retained true-signal events against thresholds at fixed edge-weight percentiles and at the MST upper bound. If Algorithm 2's T is not close to the best precision/recall operating point, or if it removes signal and noise at similar rates, the denoising claim is unsupported; a recognition-only re-run cannot settle it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: adaptive noise removal and recognition accuracy. The load-bearing assumption is in Algorithm 2 / Eq. (10): choosing δ that maximizes the variance of the normalized degree distribution is taken to preserve meaningful event structure while filtering noise. The paper's only support is Table III, which compares recognition accuracy with and without denoising and labels the difference 'noise reduction performance'. That is not a measure of noise removal: a sparser graph can improve a GAT classifier even if the removed edges are signal edges, so the ablation cannot distinguish denoising from generic graph sparsification. The mechanism is also under-specified: Algorithm 2 iterates 'δ = 0 to t' with no discretization step; edge weights are continuous, so the threshold T is not uniquely defined by the text. If T does not track actual noise, the joint-denoising claim fails and the accuracy gains may not transfer to event data with different noise statistics, even though the four reported numbers themselves are not internally contradicted.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes AW-GATCN, a graph-based method for joint denoising and object recognition of event camera data. The pipeline consists of three main components: adaptive event-point segmentation based on normalized density and a square-root-law voxelization (Algorithm 1), an adaptive denoising procedure that builds weighted graphs using Euclidean distance, velocity magnitude difference, angular difference, and polarity consistency and then selects a threshold by maximizing the variance of the normalized degree distribution (Algorithm 2, Eq. (10)), and a graph attention convolutional network that uses the resulting edge weights in the attention computation (Eqs. (13)--(14)). The method is evaluated on N-Caltech101, CIFAR10-DVS, MNIST-DVS, and N-CARS, where it is reported to achieve 83.77%, 76.79%, 99.30%, and 96.89% accuracy, respectively, outperforming prior graph-based methods by up to 8.79%. An ablation study on D-END, N-END, and N-CARS compares four weight configurations and reports accuracy improvements of 12.2--19.57 percentage points from the proposed denoising step.","tokens_in":11367,"tokens_out":4597,"duration_ms":48335,"significance":"If the reported results hold, the accuracy gains on N-Caltech101 and CIFAR10-DVS over prior graph-based methods such as NVS and EvS are meaningful, and the idea of adapting the graph threshold from the degree distribution is a simple and potentially useful contribution. The claims are empirical and falsifiable: the four benchmark numbers can be checked by independent implementation. However, the paper provides no code and no variance estimates, and the central denoising claim is not directly evidenced. The paper's contribution would be strengthened substantially by a concrete specification of Algorithm 2, a direct noise-removal evaluation, and a clarification of how the hyperparameters were chosen relative to the reported test sets.","major_comments":[{"comment":"Algorithm 2 is under-specified because the loop 'for δ = 0 to t' has no discretization step. The edge weights in Eq. (6) are continuous, so there are infinitely many candidate values of δ between 0 and t, and the argmax in Eq. (10) is not uniquely defined by the text. The authors should specify whether δ is swept over the sorted unique edge weights, over a fixed grid, or over some other finite set. Without this, the denoising step and the reported results are not reproducible.","section":"Section III-B, Algorithm 2 and Eq. (10)"},{"comment":"The edge weight in Eq. (6) sums four heterogeneous terms: Euclidean distance, velocity magnitude difference, angular difference, and a binary polarity term. These quantities have different units and ranges, yet no normalization or scaling is described. Consequently, the fixed coefficients α=0.7, β=0.1, γ=0.1, δ=0.1 are not meaningful unless each term is normalized to a common scale, and the threshold T in Algorithm 2 will depend on the arbitrary scales of the input coordinates and timestamps. The authors should specify how each term is normalized or otherwise made commensurable.","section":"Section III-B, Eq. (6)"},{"comment":"Table III is presented as evidence of 'noise reduction performance' or 'denoising effectiveness,' but it only compares recognition accuracy with and without the denoising step. An accuracy improvement from removing edges is not a measurement of noise removal: the procedure could act as a generic graph sparsifier that helps the classifier even if it removes signal edges. To support the joint-denoising claim, the authors need a direct evaluation of noise-removal quality, for example by adding synthetic noise with known ground truth and reporting precision/recall of retained events, or by comparing against established event-denoising methods on noisy event streams.","section":"Section IV-C2 and Table III"},{"comment":"The weight parameters are tuned on N-CARS and the same N-CARS result is then reported as a main result. Table II shows that Comb 3 achieved 96.89% on N-CARS, and Table I reports 96.89% for AW-GATCN on N-CARS. If the parameter selection in Section IV-C1 used the N-CARS test data, then this is test-set tuning and the reported N-CARS accuracy is optimistically biased. The authors should clarify how N-CARS was split during parameter selection or restrict the hyperparameter choice to validation folds.","section":"Section IV-C1, Table II, and Table I"},{"comment":"The main comparison in Table I reports only single accuracy numbers for each method, despite stating that all models are trained with 5-fold cross-validation. Without standard deviations, confidence intervals, or the per-fold results, it is not possible to assess whether the reported gains, some of which are small (e.g., 99.3% vs. 99.4% for AsyncNet on MNIST-DVS), are statistically significant. The authors should report the mean and variance across folds for their own method, and ideally for the compared methods.","section":"Section IV-A and Table I"}],"minor_comments":[{"comment":"The computed Nvoxels is a real number (a square root), but it is used as a count of voxels. The manuscript should specify whether rounding, ceiling, or floor is applied before the clamping operation.","section":"Algorithm 1, line 12"},{"comment":"The preprocessing constants Nmin, Cscale, Nmin_vox, and Nmax_vox appear in Algorithm 1 but their values are never reported. Without these values, the segmentation and the subsequent graph construction are not fully reproducible.","section":"Section IV-A"},{"comment":"Eq. (9) reuses the symbol Φδ for both the raw and the normalized degree distribution, which makes the normalization step confusing. A different symbol, for example Φ̂δ, should be used.","section":"Section III-B, Eqs. (8)--(9)"},{"comment":"The attention coefficient uses 1/wij, but nothing in the text rules out wij = 0, for example when two events have identical coordinates and timestamps or when the polarity term is zero. The authors should state how division by zero is avoided.","section":"Section III-C, Eq. (13)"},{"comment":"The abstract claims 'improving noise reduction performance by up to 19.57%,' but Table III reports improvement in recognition accuracy, not in noise reduction. The wording should be changed to 'improving recognition accuracy' or a direct noise-removal metric should be reported.","section":"Abstract and Section IV-C2"},{"comment":"The description of CIFAR10-DVS states that it contains 'a sixth of the original CIFAR10 dataset, totaling 60,000 samples (6,000 per class),' which is internally inconsistent: one sixth of 60,000 is 10,000. The authors should correct the dataset statistics or clarify the version used.","section":"Section IV-A, dataset description"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is an honest engineering effort: it takes known graph-attention machinery, adds an adaptive segmentation step and a multi-factor edge weight, and reports consistent accuracy gains across four event-camera benchmarks. The specific new idea is choosing the denoising threshold by maximizing the variance of the normalized degree distribution, and the ablations show that the full pipeline beats its own no-denoising baseline by double digits on some datasets. That is real evidence that the method does something useful. The authors also give credit where it is due, citing the earlier AGSF work that inspired the adaptive graph formulation.\n\nThe soft spots are real, and the stress-test note lands. First, Table III reports recognition accuracy with and without denoising, not any direct measure of noise removal. A sparser graph can help a GAT classifier even if the removed edges are signal, so the ablation cannot distinguish true denoising from generic sparsification. Second, Algorithm 2 is under-specified: it iterates 'δ = 0 to t' but the edge weights are continuous, so the threshold T is not uniquely defined without a discretization step. That is a load-bearing detail, not a cosmetic one. Third, the weight parameters α, β, γ, δ are tuned on D-END, N-END, and N-CARS, and then the same configuration is reported on N-CARS in the main table—that is a selection overlap on one of the four headline numbers. Fourth, the paper reports single accuracy numbers after 5-fold cross-validation without any variance, and no code is released. These are not fatal flaws, but they make independent verification impossible.\n\nI would not call the central results circular. The threshold heuristic is empirical, and the accuracy gains could transfer to other event data, but the current write-up does not give the reader enough to know whether the heuristic tracks actual noise or just prefers a certain graph density. The citation pattern is fine; the self-citation is to the authors' own earlier work and is relevant.\n\nWho should read this: anyone working on graph-based event camera processing, especially denoising or recognition pipelines. It is not a major methodological breakthrough, but it is a competent combination of ideas with credible benchmark numbers. For peer review, I would send it out with a request for code, a proper discretization of Algorithm 2, and a direct noise-removal evaluation (e.g., event-wise precision/recall on a dataset with ground-truth noise). With those additions, the paper would be solid; without them, the central denoising claim remains a plausible but unverified heuristic.","headline":"A plausible engineering paper whose variance-maximizing denoising rule is the load-bearing but least-supported piece; worth refereeing with requests for code and a better-specified threshold.","tokens_in":11812,"tokens_out":1641,"would_cite":false,"duration_ms":18964,"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":"Adaptive graph-based denoising prunes event-camera noise by maximizing degree-distribution variance and lifts recognition accuracy on four benchmarks.","keywords":["event cameras","event data denoising","graph attention networks","object recognition","adaptive thresholding","degree distribution variance","neuromorphic vision","voxel segmentation"],"falsifier":"Take a clean event stream with known labels, inject Poisson background events at several controlled rates, and check two things: whether the per-voxel threshold $T$ increases with the injection rate, and whether recognition accuracy with denoising stays above accuracy without denoising at every rate. If $T$ is flat or accuracy collapses at moderate injection rates, the variance-maximizing heuristic is not tracking noise.","tokens_in":10946,"feed_emoji":"⚡","tokens_out":14609,"duration_ms":125190,"temperature":0.7,"pith_summary":"Event cameras emit sparse, asynchronous streams that mix object structure with large amounts of redundant and noisy events. The paper claims that this noise can be removed and the remaining events classified by building a graph per local voxel, weighting each edge by a combination of distance, velocity, angular difference, and polarity, and then pruning edges with an adaptive threshold chosen to maximize the variance of the normalized degree distribution. It reports accuracies of 83.77% on N-Caltech101, 76.79% on CIFAR10-DVS, 99.30% on MNIST-DVS, and 96.89% on N-CARS, beating previous graph-based approaches by up to 8.79% and improving recognition accuracy by up to 19.57% when the denoising step is switched on in ablations. A sympathetic reader would care because the method avoids fixed-radius neighbor rules and frame conversion, which are the usual weak points in event-based recognition.","feed_headline":"Event-camera accuracy climbs to 99.3% with adaptive graph pruning","feed_subtitle":"On four event-camera benchmarks, pruning noisy graph edges by a variance-maximizing threshold beats graph rivals by up to 8.79%.","key_machinery":"The load-bearing mechanism is the adaptive edge-pruning step: a multifactorial edge weight that combines Euclidean distance, velocity magnitude difference, angular difference, and polarity consistency, together with a per-voxel threshold $T$ set by maximizing the variance of the normalized degree distribution. The threshold is meant to keep the graph's structure most differentiated while discarding weakly connected noise. The same edge weights are then inverted inside a graph attention convolution, so the network's feature aggregation is steered by the same measure of event-point correlation that drives denoising.","core_discovery":"On its own terms, the paper's central discovery is that the graph formulation itself can do the denoising, if the graph is built and pruned adaptively. Starting from an event stream $E=\\{(x_k,y_k,t_k,p_k)\\}$, the method segments events into density-balanced windows and then voxels, constructs a graph per voxel with edge weights $w_{ij}=\\alpha D_{ij}+\\beta\\Delta v_{ij}+\\gamma\\theta_{ij}+\\delta P_{ij}$, and removes edges whose weight exceeds a threshold $T$ chosen per voxel as $T=\\arg\\max_\\delta \\sigma^2_\\delta$, where $\\sigma^2_\\delta$ is the variance of the normalized degree distribution at threshold $\\delta$. The surviving graph feeds two graph-attention convolutional layers whose attention coefficients include $1/w_{ij}$, so strongly correlated pairs dominate feature aggregation. The paper reports that this joint denoising-and-recognition pipeline surpasses prior graph-based methods on all four tested benchmarks, with the largest gains on the noisiest datasets.","pith_inferences":["The variance-maximizing threshold may serve as an unsupervised noise-level indicator: on a dataset with known injected noise, the chosen $T$ should rise with noise rate, which would let the same mechanism calibrate event-camera preprocessing without labels.","Although the paper evaluates offline recognition, the per-voxel threshold computation is local and cheap enough that a streaming or online variant is a natural next step for latency-sensitive applications such as drone navigation.","The fixed weight parameters $\\alpha=0.7,\\beta=0.1,\\gamma=0.1,\\delta=0.1$ are chosen by search on END and N-CARS; making them learnable end-to-end could transfer better to new event-camera domains, but that extension goes beyond what the paper tests."],"forward_implications":["If the reported results hold, event-based recognition can be improved without dense frame conversion, preserving the sparsity and temporal resolution of the raw stream.","The adaptive threshold makes denoising self-tuning per voxel, so the method should transfer across scene densities without re-tuning a global radius.","The ablation comparison shows the balanced multifactor weight (Comb 3) beats the distance-only configuration by 4.50 points on D-END and 6.26 points on N-CARS.","The denoising ablation attributes most of the gain to the noise filter: accuracy rises by 17.19% on D-END, 19.57% on N-CARS, and 12.2% on N-END when denoising is enabled."],"supporting_citations":[{"why":"Graph-based object classification for neuromorphic sensing is the direct graph baseline that the paper's accuracy comparisons extend and improve upon.","marker":"[11]"},{"why":"Space-time event clouds for gesture recognition provide the event-cloud graph representation that motivates converting event streams into graphs.","marker":"[12]"},{"why":"Graph attention networks supply the attention mechanism that AW-GATCN adapts by weighting attention with inverse edge weights.","marker":"[15]"},{"why":"Adaptive graph formulation for shape representation is the source of the adaptive denoising idea that the paper reworks for event data.","marker":"[26]"},{"why":"The N-Caltech101 dataset, created by converting static images to event streams, is one of the four main recognition benchmarks.","marker":"[27]"},{"why":"The CIFAR10-DVS event-stream dataset is one of the four main recognition benchmarks.","marker":"[28]"},{"why":"The MNIST-DVS event dataset is one of the four main recognition benchmarks.","marker":"[29]"},{"why":"The E-MLB multilevel event denoising benchmark supplies the D-END and N-END datasets used in the ablation studies.","marker":"[30]"}],"fun_headline_variants":["Adaptive graph denoising lifts event-camera recognition to 99.3%","Event-camera noise slashed by adaptive graph pruning, accuracy up 8.79%","Graph pruner removes event noise, raises recognition by up to 8.79%","Adaptive graph weighting beats Euclidean by 6.26% on event recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the threshold $T$ maximizing the variance of the normalized degree distribution separates meaningful event structure from noise; if that heuristic does not track the actual noise level, denoising will discard signal or keep noise and the accuracy gains will not carry over to new datasets.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive graph denoising lifts event-camera recognition to 99.3%","Event-camera noise slashed by adaptive graph pruning, accuracy up 8.79%","Graph pruner removes event noise, raises recognition by up to 8.79%","Adaptive graph weighting beats Euclidean by 6.26% on event recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000619,"raw_usage":{"total_tokens":2877,"prompt_tokens":954,"completion_tokens":1923,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":1833}},"tokens_in":570,"tokens_out":1923,"duration_ms":14862,"temperature":1.0,"reasoning_tokens":1833,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:54:12.811557+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a clean event stream with known labels, inject Poisson background events at several controlled rates, and check two things: whether the per-voxel threshold $T$ increases with the injection rate, and whether recognition accuracy with denoising stays above accuracy without denoising at every rate. If $T$ is flat or accuracy collapses at moderate injection rates, the variance-maximizing heuristic is not tracking noise.","supporting_citations":[{"cited_title":"Converting Static Image Datasets to Spiking Neuromorphic Datasets Using Saccades","cited_arxiv_id":"1507.07629","evidence_quote":"The N-Caltech101 dataset, created by converting static images to event streams, is one of the four main recognition benchmarks."},{"cited_title":"CIFAR10-DVS: An Event-Stream Dataset for Object Classification,","cited_arxiv_id":null,"evidence_quote":"The CIFAR10-DVS event-stream dataset is one of the four main recognition benchmarks."}],"review_version":1}