{"id":"55823944-0090-49a0-b0e2-c535e8115977","arxiv_id":"2412.20803","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A frequency-aware network operating on raw-like event clouds matches or beats prior event-based models on nine benchmarks while using roughly 0.1 G MACs, far below frame and voxel baselines.","lead":"FECNet is a lightweight neural network that processes event-camera data directly as event clouds, preserving polarity and time order, and uses Fourier-domain filters to cut computation. It reports state-of-the-art accuracy on nine event-based benchmarks while running at about 230 frames per second, which could make event cameras more practical for low-power robotics and edge devices.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is pinned to a single anomalous UCF101-DVS number (0.916 vs prior 0.787) obtained with an unspecified random 20% split; until the split and seed are verified the empirical claim is not load-bearing safe.","rationale":"I do not think the frequency-filter attribution is the largest risk: a global frequency filter is equivalent to a long circular convolution, so it is a legitimate long-range mixer, and Table 7 shows removing it hurts performance. The real exposure is that the paper's only large SOTA margin is on a dataset whose split is described only as a random 20% draw with no seed. This is a reproducibility and correctness risk, not an internal inconsistency. The paper otherwise has useful benchmarking and a plausible efficiency story; the right gate is conditional acceptance on split verification and code release. My recommendation is therefore unchanged from the reader's CONDITIONAL verdict, but for a different primary reason.","tokens_in":13910,"tokens_out":9214,"duration_ms":96625,"concrete_test":"Run FECNet on UCF101-DVS using the exact split provided by the UCF101-DVS release or by Bi et al. (or the split used by VMST-Net), with three seeds and a fixed downsampling procedure, and re-run VMST-Net under the same protocol. If FECNet's accuracy drops below roughly 0.85, or if VMST-Net rises to within a few points, the advertised +16.4% SOTA gain is a split artifact; also test for train/test sample overlap from the same source video.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that FECNet achieves SOTA on all three action-recognition datasets depends disproportionately on UCF101-DVS, where FECNet reports 0.916 versus VMST-Net's 0.787, a +16.4 percentage-point jump. The gains on DVS128 Gesture (0.989 vs 0.988) and Daily DVS (0.993 vs 0.991) are marginal. For UCF101-DVS, the paper states in Sec. 4.1 that datasets without official divisions were split by \"randomly selected 20% of the samples as the testset,\" but it gives no random seed, no stratification, and no check for source-video overlap between train and test. If the split is not the standard split used by the baselines, or if clips from the same source video appear on both sides, the reported accuracy is not comparable and the SOTA claim collapses irrespective of whether the Temporal-FA module is the right explanation. This is more load-bearing than the frequency-filter attribution question, because it attacks the empirical headline rather than only the mechanism story.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a point-cloud-style Event Cloud representation for event cameras, with an event-based Grouping and Sampling module (D-FPS, EF-KNN, CES), Spatial and Temporal Frequency-aware modules based on the Fourier transform, and a residual block. The authors evaluate the resulting FECNet on object classification, action recognition, and human pose estimation, reporting strong accuracy with very low MACs (e.g., 0.109 G on DVS128 Gesture) and near-SOTA or SOTA results across nine datasets. The claimed headline is that FECNet achieves SOTA on all three action-recognition datasets, with a particularly large gain on UCF101-DVS (0.916 vs 0.787). The manuscript also includes an event-number scalability study and an ablation of the main modules. However, the paper has internal inconsistencies (SECNet vs FECNet; ten vs nine datasets), and the largest empirical claim rests on a random, unseeded test split whose comparability to prior work is not established.","tokens_in":14126,"tokens_out":3903,"duration_ms":40093,"significance":"If the empirical results are reproducible, the paper is a useful contribution: the architecture combines raw-event fidelity with dramatically reduced MACs, reports throughput above 200 FPS, shows favorable preprocessing time for Event Cloud versus frames/voxels, and demonstrates scaling to much larger event counts than prior point-based methods. The complexity argument for replacing convolution with frequency-domain Hadamard products is concrete, and the ablation study, while incomplete, gives evidence that polarity, the event-based G&S module, and temporal frequency filtering each contribute. The code links in both abstract and full text are a strength, though the two links currently point to differently named repositories and must be reconciled. The main risk is not the mechanism but the reproducibility and comparability of the benchmark numbers, particularly the UCF101-DVS split and the absence of error bars for close margins.","major_comments":[{"comment":"The UCF101-DVS result that carries the SOTA claim (0.916 vs 0.787 for VMST-Net) is obtained on a \"randomly selected 20%\" test split with no seed, no stratification, and no check for source-video overlap between train and test. Because the reported accuracy advantage is +16.4 percentage points, this single number is load-bearing; without releasing the exact split (or a reproducible file list) and confirming that the same split is used for all baselines, the claimed SOTA on action recognition is not verifiable.","section":"4.1, Table 4"},{"comment":"The manuscript is internally inconsistent about its own identity and scope: the arXiv title and abstract describe \"SECNet\" evaluated on \"ten\" datasets, while the full text (title, Section 4.1, Table 1, Section 5) describes \"FECNet\" evaluated on \"nine\" datasets. This is not just a typographical issue: the abstract's claim of ten datasets cannot be checked, and the naming mismatch prevents readers from connecting the paper to the code repositories (SECNet_ICML in the abstract, FECNet in the full text). The authors must harmonize these and either account for the missing tenth dataset or correct the abstract.","section":"Title, Abstract, 4.1, Table 1"},{"comment":"None of the main accuracy tables report variance or the number of seeds. Given that two of the three action-recognition gains are marginal (0.989 vs 0.988 on DVS128 Gesture; 0.993 vs 0.991 on Daily DVS), the differences could be within run-to-run noise; without mean ± std over multiple seeds, or a clear statement that all runs use fixed and reported seeds, the empirical claims are not supported at the precision implied by the tables.","section":"Tables 2, 4, 6"},{"comment":"The paper attributes the long-sequence gains to the Temporal-FA module, but the evidence is aggregate: Table 8 compares FECNet with an LSTM variant and point-based networks, not with a temporal convolution or attention module at matched capacity, and no analysis shows which frequencies are preserved or discarded. Since the frequency-aware mechanism is the paper's central design novelty, a matched-capacity temporal-convolution ablation (or a learned-filter frequency analysis, e.g., plotting the magnitude of the learned filters) is needed to support the attribution; without it, the accuracy gains could be due to the Event Cloud representation or other modules rather than to the temporal frequency filtering.","section":"3.3, Table 8"}],"minor_comments":[{"comment":"The summation notation is garbled (e.g., \"T −1X\" and \"j 2π T kn\"), and \"specturm\" should be \"spectrum\"; these formatting errors obscure the DFT definitions and should be corrected.","section":"3.3, Eq. (10)-(12)"},{"comment":"The column header \"ALS-DVS\" should read \"ASL-DVS\", and the caption contains typos such as \"Excperimemt\"; these should be fixed throughout the tables and text.","section":"Table 2"},{"comment":"The batch size is described as \"dynamically set to take full use of the GPU's memory\", but no actual batch size, number of workers, or training seed is given; this makes the experimental setup hard to reproduce even with the code.","section":"4.2"},{"comment":"There are typos such as \"ACCuracy\" and the claim of \"one-thousandth of the MACs\" is stated without naming the specific baseline; Table 5 compares only a subset of methods, so the claim should be made relative to the particular entries shown.","section":"4.5, Table 5"},{"comment":"Reference [21] for the tonic package is incomplete: it lacks a proper title and should include a version or publication venue.","section":"References, [21]"},{"comment":"The captions and pseudocode contain incomplete sentences and inconsistent terminology, e.g., \"Classifer Regressor\" and \"Get category ˆc Get 13 skeleton points ˆp\"; these should be cleaned up.","section":"Algorithm 1, Figure 3"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the central empirical claim depends on a single anomalous number on UCF101-DVS, and the split is not reproducible from the manuscript. The title/abstract inconsistency (SECNet vs FECNet, ten vs nine datasets) and the two different code repository links suggest the paper may have been assembled from different versions or submissions; this needs editorial attention as much as technical revision. I would like to see the exact split, the variance over seeds, and a matched-capacity temporal baseline before considering acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about this paper. The architecture is a genuine step forward for event-cloud methods, and the efficiency numbers are impressive. But the central claim—SOTA on all three action recognition datasets—leans almost entirely on one UCF101-DVS result, and that result comes from an unseeded, unspecified random 20% split. Until the authors pin down the split and rule out source-video leakage, the empirical headline is not safe to trust.\n\nWhat is actually new: the paper treats events as order-preserving 2S-1T-1P clouds and builds a grouping/sampling module that keeps polarity in the structure rather than as an input channel. The D-FPS scaling, EF-KNN, and CES are sensible engineering. The spatial/temporal frequency-aware modules replace convolutions with Hadamard products along the feature and time axes; the complexity argument is plausible, and the ablation shows removing the temporal-FA module hurts. The benchmark coverage is broad: nine datasets, three tasks, plus preprocessing-time measurements and a hardware table. That is real work, and at 0.109 G MACs with roughly 230 FPS on DVS128 Gesture, the efficiency story is believable.\n\nSoft spots, in order of severity. First, the UCF101-DVS number. The paper says it randomly selected 20% of samples as test, with no seed, no stratification, and no check for same-source-video overlap. VMST-Net's 0.787 may come from a different split; if so, 0.916 is not comparable and the SOTA claim collapses. The gains on DVS128 Gesture and Daily DVS are marginal. Second, the text is internally sloppy: the abstract says SECNet on ten datasets while the full text says FECNet on nine, and the GitHub link changes name. This is fixable but it undermines confidence in the stated scope. Third, there are no error bars and no released code or seeds, so none of the headline numbers are independently reproducible right now. Fourth, \"pioneer\" is an overstatement; the components are known from point-cloud and frequency-domain literature, and the contribution is the assembly and empirical demonstration. The heavy self-citation cluster to the group's own point-cloud papers is noticeable but not disqualifying.\n\nIs the Temporal-FA module the right explanation for the gains? The paper does not analyze which frequencies matter or compare against a matched-capacity temporal convolution. But the ablation gives it some support, so this is a secondary concern.\n\nThis paper deserves a serious referee. Send it out, but require the authors to release exact split seeds, stratification, and a video-level leakage check for UCF101-DVS. If the split is non-standard, the SOTA claim must be revised. I would not cite the headline numbers until that is resolved.","headline":"A real efficiency-first architecture, but the SOTA claim is hinged on one unseeded UCF101-DVS split and should not be trusted until the authors provide the split and a leakage check.","tokens_in":14691,"tokens_out":3290,"would_cite":false,"duration_ms":33711,"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":"FECNet claims state-of-the-art action recognition from raw event clouds at 0.1 G MACs.","keywords":["event cameras","event cloud representation","frequency domain learning","action recognition","point cloud networks","efficient neural networks","neuromorphic vision"],"falsifier":"Run FECNet on the event-number sweep of Table 8 with the Temporal-FA module replaced by a one-dimensional temporal convolution or a GRU of equal parameter count: if accuracy on N-Caltech101 or UCF101-DVS stays within noise, or if the learned global filters are near-uniform and removing them leaves results unchanged, the paper's attribution of long-sequence gains to the frequency-domain filter is settled false.","tokens_in":13679,"feed_emoji":"⚡","tokens_out":4206,"duration_ms":43109,"temperature":0.7,"pith_summary":"The paper argues that raw event-camera outputs, treated as an ordered 'event cloud' of spatial, temporal, and polarity coordinates, can support state-of-the-art action recognition while using roughly one-thousandth of the multiply-accumulate operations of frame-based models. It proposes FECNet, a network that learns local groupings from event features rather than from raw coordinates alone, and that extracts spatial and temporal features with Fourier-domain filters instead of convolutions and recurrent or attention modules. If the claim holds, event-cloud processing becomes a practical backbone for low-power, high-speed event-based perception across classification, action recognition, and human pose estimation.","feed_headline":"Event-cloud network tops action benchmarks at 0.1 G MACs","feed_subtitle":"Fourier-domain filters let a lightweight network handle long event sequences with high accuracy.","key_machinery":"The central object is the Event Cloud: an ordered, downsampled sequence of raw events with four coordinates (x, y, t, polarity), treated as a one-dimensional temporal signal rather than as a permutation-invariant point set. The key mechanism is the Fourier transform: the Spatial-FA module replaces a convolution over the feature dimension with a Hadamard product in the frequency domain, reducing per-point complexity from O($d^{2}$) to O(d log d), and the Temporal-FA module applies a learnable global filter to the spectrum along the event-time axis to model long-range temporal relationships. Around this mechanism sit the Differentiation Farthest Point Sampling, Event Feature-based K-Nearest Neighbour, and Coordinates Evolution Strategy, which construct local groups that respect the different meanings of the four event coordinates. Together they keep the input close to the raw sensor output while keeping the model cheap enough to handle tens of thousands of events.","core_discovery":"FECNet's central claim is that the finest-grained event representation, an ordered Event Cloud with coordinates (x, y, t, polarity), can be scaled to long event sequences without a computational explosion by moving feature extraction into the frequency domain. The paper reports state-of-the-art results on all three action recognition datasets it tests (DVS128 Gesture 98.9%, Daily DVS 99.3%, UCF101-DVS 91.6%), exceeding the prior best on UCF101-DVS by 16 percentage points, while keeping computational cost at 0.109 G MACs and throughput at about 230 FPS on DVS128 Gesture. On object classification and human pose estimation it reports accuracy comparable to or better than existing point-based and voxel-based methods, with the smallest model size among pose estimation baselines. The claimed source of these gains is the combination of an Event Cloud that preserves polarity and fine temporal order, a redesigned grouping-and-sampling module that treats each coordinate according to its role, and Fourier-domain filters that capture long-range dependencies globally at low cost.","pith_inferences":["A natural test the paper leaves implicit is to inspect the learned Temporal-FA filters on trained models: if the filters are nearly uniform or the accuracy gain vanishes when they are removed, the frequency-domain mechanism is not the true source of the reported long-sequence improvements.","Because the method relies on temporal order rather than permutation invariance, the same frequency-aware grouping could transfer to other ordered spatiotemporal point streams, such as timestamped LiDAR returns, where polarity is absent but time ordering still matters.","Since FECNet's accuracy keeps rising as the event count grows while a point-cloud baseline plateaus, a practical extension would be a runtime scheduler that adapts the number of events fed to the network based on scene activity, trading accuracy against latency."],"forward_implications":["Event Cloud can be processed at near-raw fidelity, so temporal information lost in frame or voxel conversion need not be sacrificed for efficiency.","Frequency-domain filtering along the time axis can replace attention and LSTM components for long event sequences at a fraction of the computational cost.","Polarity, integrated structurally during grouping rather than as an extra input channel, improves accuracy across classification, action recognition, and pose estimation.","The reported trade-off of 0.109 G MACs, roughly 230 FPS, and 98.9% accuracy on DVS128 Gesture makes event-cloud networks feasible for real-time, low-power vision systems."],"supporting_citations":[{"why":"Supplies the global learnable frequency filter that the Spatial-FA and Temporal-FA modules adapt from image models to event streams.","marker":"[40]"},{"why":"Provides the PointNet++ hierarchical grouping and sampling baseline that FECNet reworks for event coordinates.","marker":"[39]"},{"why":"TTPOINT is the prior point-cloud action recognition method that FECNet compares against and the main baseline for the event-number scalability study.","marker":"[41]"},{"why":"EventNet is an early point-cloud event network used as a classification baseline that motivates the Event Cloud representation.","marker":"[45]"},{"why":"Supplies the DVS128 Gesture dataset, one of the three action recognition benchmarks where FECNet reports state-of-the-art accuracy.","marker":"[1]"},{"why":"DHP19 is the event-based human pose dataset used to evaluate FECNet's pose estimation performance.","marker":"[5]"},{"why":"Provides dataset split settings for several benchmarks and is a strong voxel-based comparison method.","marker":"[4]"},{"why":"Provides the neuromorphic object classification datasets NMNIST and N-Caltech101 that are central to the classification experiments.","marker":"[35]"}],"fun_headline_variants":["Fourier-domain event cloud network slashes MACs to 0.1 G","Event cloud with Fourier filters hits 99% action accuracy","SECNet: event cloud network delivers 16-point accuracy jump on UCF101-DVS","Frequency-domain features make event-based classification scalable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a global learnable filter applied to the Fourier spectrum along the event-time axis can capture long-range temporal dependencies better than attention or LSTM, at matched cost, without losing localization.","fun_headline_variants_meta":{"raw":{"variants":["Fourier-domain event cloud network slashes MACs to 0.1 G","Event cloud with Fourier filters hits 99% action accuracy","SECNet: event cloud network delivers 16-point accuracy jump on UCF101-DVS","Frequency-domain features make event-based classification scalable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000321,"raw_usage":{"total_tokens":1810,"prompt_tokens":948,"completion_tokens":862,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":785}},"tokens_in":564,"tokens_out":862,"duration_ms":7363,"temperature":1.0,"reasoning_tokens":785,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:10:19.204163+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FECNet on the event-number sweep of Table 8 with the Temporal-FA module replaced by a one-dimensional temporal convolution or a GRU of equal parameter count: if accuracy on N-Caltech101 or UCF101-DVS stays within noise, or if the learned global filters are near-uniform and removing them leaves results unchanged, the paper's attribution of long-sequence gains to the frequency-domain filter is settled false.","supporting_citations":[{"cited_title":"Pointnet++: Deep hierarchical feature learning on point sets in a metric space","cited_arxiv_id":null,"evidence_quote":"Provides the PointNet++ hierarchical grouping and sampling baseline that FECNet reworks for event coordinates."},{"cited_title":"Ttpoint: A tensorized point cloud network for lightweight action recognition with event cam- eras","cited_arxiv_id":null,"evidence_quote":"TTPOINT is the prior point-cloud action recognition method that FECNet compares against and the main baseline for the event-number scalability study."},{"cited_title":"Eventnet: Asynchronous recursive event processing","cited_arxiv_id":null,"evidence_quote":"EventNet is an early point-cloud event network used as a classification baseline that motivates the Event Cloud representation."},{"cited_title":"A low power, fully event-based gesture recognition system","cited_arxiv_id":null,"evidence_quote":"Supplies the DVS128 Gesture dataset, one of the three action recognition benchmarks where FECNet reports state-of-the-art accuracy."},{"cited_title":"Dhp19: Dynamic vi- sion sensor 3d human pose dataset","cited_arxiv_id":null,"evidence_quote":"DHP19 is the event-based human pose dataset used to evaluate FECNet's pose estimation performance."},{"cited_title":"Graph-based spatio-temporal feature learning for neuromorphic vision sensing","cited_arxiv_id":null,"evidence_quote":"Provides dataset split settings for several benchmarks and is a strong voxel-based comparison method."},{"cited_title":"Converting static image datasets to spiking neuromorphic datasets using saccades","cited_arxiv_id":null,"evidence_quote":"Provides the neuromorphic object classification datasets NMNIST and N-Caltech101 that are central to the classification experiments."}],"review_version":1}