{"id":"288d9793-d569-4a65-bd59-e7aa523a98dd","arxiv_id":"2505.21187","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A causal, density-based event subsampling method preserves classification accuracy better than random, spatial, temporal, event-count, and corner-based baselines in sparse regimes, except when event counts vary widely across videos.","lead":"The paper compares six ways to discard events from event cameras and measures how much classification accuracy survives. It finds that keeping events from dense, busy regions preserves accuracy best at low event budgets, and warns that naive row/column subsampling is easily thrown off by grid alignment.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The causal density method's sparse-regime advantage depends on a fixed, dataset-specific threshold; N-Cars (Fig. 9) shows this starves sparse videos, so the abstract's unqualified claim is not supported for high event-count-variance datasets.","rationale":"The reader's weakest assumption identifies the same load-bearing point: the causal density method's performance hinges on a fixed threshold and kernel parameters chosen before seeing each video, and this fails when per-video event counts vary widely, as on N-Cars. This is not an external disagreement with consensus; it is internally evidenced by the paper's own Fig. 9 and Table 1. The paper is honest about the failure and proposes a non-causal normalization or an implementation-dependent adaptive threshold, but the abstract's unqualified claim overstates the result. A conditional verdict is appropriate: the main comparisons are carefully run, the error bars are reported, and the density hypothesis is plausible on N-Caltech101 and DVS-Gesture, but the central claim needs the variance-dependent qualifier and ideally a causal adaptive threshold demonstration. The proposed concrete test would settle whether the N-Cars failure is a calibration artifact or a genuine limitation of causal density-based subsampling. No new objection beyond the reader's was found, so the verdict should remain unchanged.","tokens_in":15132,"tokens_out":7075,"duration_ms":83485,"concrete_test":"Run the N-Cars experiment with a causal adaptive threshold: lower f^(thresh) when a causal event-rate estimate is low (e.g., compare each density value to a running quantile of recent densities rather than a fixed dataset-wide value), retrain at the six Table 3 subsampling levels, and compare per-video accuracy stratified by post-subsampling event count against random subsampling. If adaptive-causal density reaches random-level nAUC (~0.82), the N-Cars failure is a threshold-calibration artifact and the density hypothesis survives in qualified form; if it does not, the central claim should be restricted to low-variance event-count datasets.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the central claim, density-based selection must improve sparse-regime accuracy because it retains informative events, not because a fixed threshold happens to fit a dataset's event-count distribution. The paper's own N-Cars experiment falsifies the unqualified version: with fixed f^(thresh) (Eq. 3, Table 3), Fig. 9 shows many videos reduced to near-zero events, those videos' accuracy collapses, and density-based nAUC is 0.789 versus 0.825 for random (Table 1). Matching only the average retained-event count (Figs. 4–5) hides this: the method redistributes events unevenly across videos, so the same ⟨N⟩ is not the same operating point. The Fig. 10 fix (mean-normalized densities) explicitly sacrifices causality, and the paper's suggested adaptive threshold would need a causal running estimate whose performance is untested. Thus the abstract's 'improved classification accuracy in sparse regimes' is established only for datasets with narrow per-video event-count variance, not as a general property of causal density-based subsampling.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper systematically compares six hardware-friendly event-camera subsampling methods—spatial, temporal, random, event count, corner-based, and a newly proposed causal density-based method—on three event-video classification benchmarks (N-Caltech101, DVS-Gesture, and N-Cars) using an EST representation with a ResNet34 classifier. The central claim is that the causal density-based method, which retains events from spatiotemporally dense regions via Eq. (2) and a random threshold in Eq. (3), achieves improved classification accuracy in sparse regimes. The paper also analyzes factors such as spatial-subsampling offset sensitivity and the failure of density-based subsampling on high-variance event-count datasets such as N-Cars.","tokens_in":15392,"tokens_out":6661,"duration_ms":74821,"significance":"If the central claim holds, the paper provides a useful practical guide for hardware-level event-rate reduction and supports the hypothesis that high-density events carry more task-relevant information. The strengths include a systematic comparison across six methods, 18 independent runs (6 for N-Cars) with reported standard deviations, control of average event counts across methods, an informative offset-sensitivity analysis, and a commitment to release code. The paper is also honest about the N-Cars failure case. However, the abstract's unqualified sparse-regime claim is contradicted by the paper's own N-Cars results, and the density method's hyperparameter dependence is under-examined; these issues should be addressed before publication.","major_comments":[{"comment":"The abstract and §4.1 state that the causal density-based method demonstrates improved classification accuracy in sparse regimes, but the N-Cars experiment in §4.3 and Table 1 contradicts this unqualified claim: with the fixed threshold of Eq. (3), the causal density method achieves nAUC = 0.789 ± 0.011 versus 0.825 ± 0.003 for random subsampling, and Fig. 9 shows that many videos are reduced to near-zero events. The claim is therefore supported only for datasets with relatively narrow per-video event-count distributions, not as a general property of causal density-based subsampling. Please either qualify the abstract and §4.1 accordingly, or add and evaluate a causal adaptive-threshold variant (e.g., based on a running event-rate estimate) that does not forfeit causality.","section":"Abstract; §4.1, §4.3, Fig. 9, Table 1"},{"comment":"The density method's advantage depends on several dataset-specific choices: the threshold f^(thresh) is set separately for each dataset and each subsampling level, and the kernel parameters wd = 7 and tau = 30 ms are chosen per dataset based on 'scene dynamics and camera resolution.' Because random subsampling is also tuned to match the average event count, the comparison may partly reflect matching the event-count distribution rather than the intrinsic informativeness of dense-region events. Please report a sensitivity analysis of the sparse-regime accuracy with respect to f^(thresh) (e.g., a sweep or transfer across datasets), or otherwise justify that the advantage is not an artifact of per-level threshold tuning.","section":"§3.2.4, Table 3"},{"comment":"The headline comparisons are based on nAUC values and standard deviations, but no significance tests are reported. For example, in DVS-Gesture the causal density method (0.883 ± 0.009) and corner-based method (0.886 ± 0.008) are separated by 0.003, within one standard deviation, yet are described as achieving 'similarly high performance'; in N-Caltech101 the density-over-random gap is 0.015 with small standard deviations, which may be significant but is not tested. In addition, the N-Cars nAUC is computed only for runs with ⟨N⟩ > 50, a post hoc filter that removes the very sparse regime where the density method collapses. Please provide paired significance tests across the shared seeds and justify the ⟨N⟩ > 50 exclusion, or show that the conclusions are unchanged without it.","section":"§4.4, Eq. (4)–(5), Table 1"}],"minor_comments":[{"comment":"The statement that 'the exponential temporal filtering enables recursive computation of the density value f^i(p_i) using the previous value of f^{i-1}(p_{i-1})' is not directly justified by Eq. (2), because the spatial kernel s(x_i - x_j, y_i - y_j) depends on the coordinates of the current event. Please clarify the recursion (e.g., a per-pixel density map with exponential decay, consistent with the O(HW) memory in Table 2) or revise the memory-efficiency explanation.","section":"§3.2.4, Eq. (2)"},{"comment":"The corner threshold h^(thresh) for DVS-Gesture is listed as 0.077, 0.17, 0.5, 16.7, 3.33, 7.70 for levels 1–6. The value 16.7 at level 4 breaks the monotonic increase expected for decreasing event counts and is likely a typo; please correct it.","section":"Table 3"},{"comment":"There are several typographical errors: 'casual density-based' in §4.1 should be 'causal density-based'; 'of of spatial filtering' in §3.2.4 should read 'of spatial filtering'; and 'appyling' in the supplementary material should be 'applying.'","section":"§4.1, §3.2.4, Supplementary §7"},{"comment":"The figure would be clearer if the histogram bars were labeled as a histogram rather than a 'bar plot,' and if the accuracy lines included confidence intervals or individual run markers, since the N-Cars averages use only six runs and are highly variable in the sparse regime.","section":"Fig. 9"},{"comment":"The integral in Eq. (4) is written as ∫ acc(#events) d(log10 #events), but the x-axis of the figures is the average number of events per video. Please clarify whether the nAUC is computed per run on the per-video event counts or on the averaged curve, as this affects the interpretation of the reported standard deviations.","section":"§4.4, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reports its own counterexample to the abstract's central claim, which is a sign of scientific honesty but also means the central claim must be narrowed or supported by a causal adaptive variant. The authors should be given the opportunity to revise; the empirical effort and reproducibility practices are otherwise solid, and the paper would be a useful contribution once the N-Cars caveat is fully integrated into the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid empirical paper, worth refereeing, but the abstract sells the density method a bit harder than the data support. The genuinely new thing is the systematic accuracy-versus-event-count comparison of six hardware-friendly subsampling rules on CNN classification, with event counts matched on average, plus the causal density rule with random thresholding. On N-Caltech101 and DVS-Gesture the density method does hold the sparse-regime edge, and the corner baseline is a sensible inclusion. The offset-sensitivity analysis for spatial subsampling is a genuinely useful result for hardware designers: it shows naive row/column dropping is fragile in a way temporal subsampling is not. The complexity table is also helpful.\n\nThe soft spots are real but mostly acknowledged. The fixed density threshold is tuned per dataset and per level, so the method's advantage over random is partly a statement about how well a threshold can be chosen in advance, not a free lunch. The N-Cars result is the load-bearing caveat: with a fixed threshold, dense-region selection starves sparse videos and underperforms random (nAUC 0.789 vs 0.825). The mean-normalized fix restores performance but explicitly gives up causality. That means the abstract's 'improved classification accuracy in sparse regimes' should be qualified to datasets with moderate per-video event-count variance. The authors do flag this in Section 4.3 and the limitations section, so it is a framing problem, not a hidden one.\n\nWhat is missing: no significance tests, and we only get standard deviations across runs. Given the reported error bars, the density-vs-random difference is probably real on the two main datasets, but a paired test would settle it. Also, the code is promised but not actually released, which matters for a comparison paper like this; the parameter tables are detailed enough to reimplement, but 'will be released' is not the same as shipped. The citation pattern is fine: the self-citation to [3] is used for experimental setup, not to prop the claim.\n\nWho this is for: people designing event-camera hardware or picking subsampling rules for edge CNNs. They will get concrete guidance. It does not reorganize the field, and the density hypothesis is not tested beyond CNN classification, but the paper is honest about that.\n\nMy recommendation: yes, send it to peer review, and ask for the qualified abstract, a significance test, and the code. The central direction holds; it needs tightening, not rejection.","headline":"A careful, useful empirical comparison of event-camera subsampling methods whose sparse-regime claim is real but dataset-conditional; the N-Cars failure is the paper's own strongest caveat.","tokens_in":15881,"tokens_out":2297,"would_cite":true,"duration_ms":25001,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Events from dense spatiotemporal regions carry the information that matters, and a causal density-based subsampler that keeps them improves event-video classification accuracy in sparse regimes.","keywords":["event camera","subsampling","density-based filtering","event-based vision","video classification","convolutional neural network","data efficiency","neuromorphic vision"],"falsifier":"Run the causal density-based method on a dataset with deliberately high per-video event-count variance while holding a fixed threshold: if many videos drop below a small event count and classification accuracy falls toward chance while random subsampling holds up, the fixed-threshold assumption is the decisive factor. Conversely, on a dataset where every video has similar total event counts, the density-based advantage should persist; measuring the accuracy gap as a function of event-count variance would settle whether the claim is about density or about threshold calibration.","tokens_in":14955,"feed_emoji":"⚡","tokens_out":4846,"duration_ms":49685,"temperature":0.7,"pith_summary":"The paper asks which events to keep when an event camera produces far more events than a downstream classifier can afford to process. It systematically compares six hardware-friendly subsampling strategies—spatial, temporal, random, event-count, corner-based, and a new causal density-based method—on three event-video classification datasets. The central claim is that events arriving in spatiotemporally dense regions carry more class-relevant information, so keeping those events yields higher CNN classification accuracy at a fixed event budget, particularly in sparse regimes. Results on N-Caltech101 and DVS-Gesture support this, while N-Cars exposes a failure mode: a fixed dataset-wide density threshold starves naturally sparse videos. If the claim holds, density-aware selection is a better default than random or spatial subsampling for rate reduction in event-camera pipelines.","feed_headline":"Density-based subsampling beats random on sparse event video","feed_subtitle":"A causal density filter keeps task-relevant events, raising CNN accuracy at fixed event budgets on two benchmarks.","key_machinery":"The central object is the causal spatiotemporal density filter of Eq. (2): each event's density is the sum over past same-polarity events of a Gaussian spatial kernel times an exponential decay in time, computed recursively per polarity. A random threshold f^(thresh) with u ~ U(0,1) decides whether to keep each event, preventing greedy selection from only the densest region. The paper's case rests on comparing this filter against spatial, temporal, random, event-count, and Harris corner-based subsampling, all tuned so the average number of events per video matches across methods.","core_discovery":"The paper introduces a causal density-based subsampling method that scores each incoming event via a recursive spatiotemporal density filter—a Gaussian spatial kernel combined with an exponential temporal decay—and retains events whose density exceeds a randomly jittered threshold. Evaluated with an EST voxel representation and ResNet34 on N-Caltech101, DVS-Gesture, and N-Cars, this method achieves the highest normalized area-under-curve accuracy-versus-event-count on N-Caltech101 and ties with corner-based Harris selection on DVS-Gesture, beating random, spatial, temporal, and event-count baselines in the sparse regime. The evidence is read as supporting the hypothesis that dense regions carry more task-relevant information. On N-Cars, fixed-threshold density selection collapses many videos to near-zero events, and accuracy drops; normalizing density values or using an adaptive threshold restores performance at the cost of strict causality.","pith_inferences":["The density hypothesis could extend beyond classification to detection or optical flow, since dense event regions often correspond to moving foreground; the paper only tests classification.","A testable extension is to replace the fixed threshold with a causal, adaptive threshold based on a running event-rate estimate and measure whether the N-Cars accuracy benefit survives without losing causality.","The random-threshold trick is not independently ablated; separating the diversity benefit of random thresholding from the density benefit itself would clarify the mechanism behind the accuracy gain.","The accuracy–event-count curves may understate the density method's value at very high subsampling rates because the fixed threshold creates a floor of near-empty videos; per-video normalization in the training loop would be a more realistic deployment recipe."],"forward_implications":["When an event budget is fixed, keeping events from dense spatiotemporal regions preserves more class-relevant information than keeping a random subset, at least for CNN classification on N-Caltech101 and DVS-Gesture.","Simple input-independent methods (spatial, temporal, random) outperform the event-count downscaling baseline, so hardware comparisons should include these trivial baselines before adopting more complex subsamplers.","Spatial subsampling is highly sensitive to row/column offset choice, while temporal subsampling is more robust, making temporal subsampling the safer naive hardware option.","A fixed density threshold cannot handle datasets with large per-video event-count variance; normalization or adaptive thresholding is required, but strict causality is then lost.","Corner-based subsampling matches density-based accuracy on DVS-Gesture, so corners are also a competitive informative-event prior for gesture data."],"supporting_citations":[{"why":"Supplies the EST voxel-grid event representation and its end-to-end learning procedure used for all CNN inputs.","marker":"[15]"},{"why":"Provides the random-subsampling training recipe and accuracy baselines this work extends and compares against.","marker":"[3]"},{"why":"Defines the Event Count spatial downscaling baseline method and its thresholding mechanism.","marker":"[19]"},{"why":"Supplies the luvHarris corner detector with the Threshold-Ordinal Surface representation used for corner-based subsampling.","marker":"[18]"},{"why":"Provides the DVS-Gesture dataset used as one of the three evaluation benchmarks.","marker":"[2]"},{"why":"Provides the N-Caltech101 dataset used as the static-image saccade benchmark.","marker":"[39]"},{"why":"Provides the N-Cars dataset used to expose the fixed-threshold failure mode in high-variance event counts.","marker":"[44]"},{"why":"Supplies the ResNet34 architecture used as the CNN classifier for all subsampling methods.","marker":"[25]"}],"fun_headline_variants":["Density-based subsampling beats random on sparse event video","Causal density filter improves event camera accuracy in sparse data","Keep dense events for better CNN accuracy on sparse streams","Event subsampling: density-based selection wins on sparse benchmarks","Density beats random for sparse event camera classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes one fixed density threshold and one set of kernel sizes chosen before seeing any video can serve every video in a dataset; N-Cars shows this assumption fails when per-video event counts vary widely, starving sparse videos of events.","fun_headline_variants_meta":{"raw":{"variants":["Density-based subsampling beats random on sparse event video","Causal density filter improves event camera accuracy in sparse data","Keep dense events for better CNN accuracy on sparse streams","Event subsampling: density-based selection wins on sparse benchmarks","Density beats random for sparse event camera classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000263,"raw_usage":{"total_tokens":1578,"prompt_tokens":903,"completion_tokens":675,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":597}},"tokens_in":519,"tokens_out":675,"duration_ms":7726,"temperature":1.0,"reasoning_tokens":597,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:33:12.223230+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the causal density-based method on a dataset with deliberately high per-video event-count variance while holding a fixed threshold: if many videos drop below a small event count and classification accuracy falls toward chance while random subsampling holds up, the fixed-threshold assumption is the decisive factor. Conversely, on a dataset where every video has similar total event counts, the density-based advantage should persist; measuring the accuracy gap as a function of event-count variance would settle whether the claim is about density or about threshold calibration.","supporting_citations":[{"cited_title":"End-to-End Learning of Repre- sentations for Asynchronous Event-Based Data","cited_arxiv_id":null,"evidence_quote":"Supplies the EST voxel-grid event representation and its end-to-end learning procedure used for all CNN inputs."},{"cited_title":"Pushing the boundaries of event subsampling in event-based video classification using CNNs","cited_arxiv_id":"2409.08953","evidence_quote":"Provides the random-subsampling training recipe and accuracy baselines this work extends and compares against."},{"cited_title":"Event Data Downscaling for Embedded Computer Vision:","cited_arxiv_id":null,"evidence_quote":"Defines the Event Count spatial downscaling baseline method and its thresholding mechanism."},{"cited_title":"luvHarris: A Practical Corner Detector for Event-Cameras","cited_arxiv_id":null,"evidence_quote":"Supplies the luvHarris corner detector with the Threshold-Ordinal Surface representation used for corner-based subsampling."},{"cited_title":"A Low Power, Fully Event-Based Gesture Recognition System","cited_arxiv_id":null,"evidence_quote":"Provides the DVS-Gesture dataset used as one of the three evaluation benchmarks."},{"cited_title":"Cohen, and Nitish Thakor","cited_arxiv_id":null,"evidence_quote":"Provides the N-Caltech101 dataset used as the static-image saccade benchmark."},{"cited_title":"HATS: Histograms of Aver- aged Time Surfaces for Robust Event-Based Object Classifi- cation","cited_arxiv_id":null,"evidence_quote":"Provides the N-Cars dataset used to expose the fixed-threshold failure mode in high-variance event counts."}],"review_version":1}