{"id":"d014fc55-3675-4a7b-90eb-9857047d373a","arxiv_id":"2505.07715","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"HsVT is a hybrid ANN-SNN Transformer that achieves modest mAP improvements over RVT on GEN1, a new fall dataset, and an aircraft dataset, with fewer parameters.","lead":"The authors introduce HsVT, a hybrid spiking Transformer that detects objects in event-camera streams by pairing Transformer attention with spiking temporal modules. It reports small mAP gains over the RVT baseline and contributes a simulated fall-detection dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed gain over RVT on GEN1 is a 0.006 mAP gap with no error bars; 'significant' is not established without multiple-seed results.","rationale":"I focused on the GEN1 comparison because it is the only standard, public benchmark where HsVT is compared with RVT and is cited in the reader's strongest claim. The 0.006 mAP gap is not accompanied by any variance estimate; a single seed is not sufficient to establish a 'significant' improvement. The reader identified the simulated FALL dataset as the weakest assumption, which I agree is a serious internal inconsistency (Abstract 'captured using an event-based camera' vs Section 3.1 'convert ... using ESIM'). However, even if the FALL dataset were real, the primary quantitative claim on GEN1 remains unsupported due to missing error bars. Since the manuscript also has no code and no working dataset link, the experimental evidence cannot be audited. My recommendation of UNCHANGED reflects that the reader's CONDITIONAL verdict already requires additional evidence; the missing statistical support is precisely the kind of evidence that should be supplied before acceptance.","tokens_in":14037,"tokens_out":10494,"duration_ms":100159,"concrete_test":"Run HsVT-B and RVT-B on GEN1 with at least 5 different random seeds, keeping identical training settings (optimizer, schedule, batch size, and Δt=50ms), and report mean±std of mAP 50:95. If the 95% confidence intervals overlap or the mean gap is below 0.005, the abstract's 'significant performance improvements' claim on the main benchmark fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, 'significant performance improvements' (Abstract), rests on Table 7: HsVT-B achieves 0.478 mAP 50:95 on GEN1 versus RVT's 0.472, a delta of 0.006. Table 8 shows FALL/AIR margins of 0.004 to 0.028. No standard deviations, seed counts, or significance tests are reported anywhere. For event-based detection training, a 0.006 mAP difference is within normal run-to-run variance; without repeated runs, the GEN1 advantage may be noise. The only consistent wins over RVT then come from the simulated FALL dataset and the non-public AIR dataset, which cannot be independently verified. Thus the headline result is not currently supported by the evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HsVT, a hybrid ANN/SNN multi-stage vision transformer for event-based object detection. The architecture combines MaxViT-based spatial feature extraction (Block-SA and Grid-SA with SpikingMLP) with LSTM and a spiking temporal feature extraction (STFE) module in the final block. The authors introduce a new Fall Detection Dataset, generated by converting the frame-based Le2i dataset into event streams with the ESIM simulator, and they evaluate HsVT on GEN1, FALL, and a private aircraft-detection dataset. On GEN1, HsVT-B reaches 0.478 mAP 50:95 versus RVT's 0.472, while HsVT-T and HsVT-S reach 0.449 and 0.465. On FALL and AIR, HsVT variants are reported to outperform RVT by small margins. The paper also reports ablations over spiking neuron models, surrogate gradients, SNN components, and component placement, plus theoretical energy estimates.","tokens_in":14210,"tokens_out":4959,"duration_ms":47481,"significance":"If the performance claims were robust, HsVT would be a useful parameter-efficient hybrid design for event-based detection, and the public release of a privacy-preserving fall-detection event benchmark would be a community contribution. The paper's strengths are the systematic ablation study, the parameter counts for each variant, and the explicit energy-estimation methodology. However, the central quantitative claim of 'significant performance improvements' is not yet supported: the main GEN1 gain over RVT is 0.006 mAP, the FALL/AIR gains are also small, and no error bars, seed counts, or significance tests are reported. The FALL dataset is simulated rather than sensor-captured, contrary to the abstract's wording, which limits the real-world transferability of the FALL results. The two supporting comparisons on non-public or synthetic data cannot compensate for the lack of statistical evidence on GEN1.","major_comments":[{"comment":"The abstract and Section 5.4 claim 'significant performance improvements', but the GEN1 comparison in Table 7 shows HsVT-B at 0.478 mAP versus RVT at 0.472, a delta of 0.006, and HsVT-T and HsVT-S at 0.449 and 0.465, below RVT. No standard deviations, number of seeds, or significance tests are reported anywhere in the paper, so the claimed significance is not established; a 0.006 mAP difference is within normal run-to-run variance for event-based detection training. Please report multi-seed results with error bars or confidence intervals for all three datasets and adjust the wording of the central claim accordingly.","section":"Abstract; Section 5.4; Table 7"},{"comment":"Table 7 contains internally inconsistent values: STAT is listed as 49.9 and SpikSSD as 40.8, while every other row uses mAP values between 0 and 1. These entries appear to be percentages and should likely read 0.499 and 0.408. As printed, the table cannot support the comparative claims; please correct these values and verify all baseline numbers against their original sources.","section":"Table 7"},{"comment":"The abstract states that the Fall Detection Dataset was 'captured using an event-based camera', but Section 3.1 explicitly says the dataset was generated by converting the frame-based Le2i videos into event streams with the ESIM simulator. This is a direct contradiction. Because FALL is one of the two datasets where HsVT is reported to consistently outperform RVT, the simulated nature of the events must be stated accurately, and the paper should discuss how well ESIM-simulated events approximate real event-camera data in this deployment scenario.","section":"Section 3.1; Abstract"},{"comment":"On FALL and AIR, Table 8 reports HsVT outperforming RVT by margins of only 0.004 to 0.028 mAP, and the AIR dataset is not publicly available. Without repeated runs and statistical tests, the claim that 'HsVT consistently outperforms RVT on both datasets' is not supported by the evidence presented. Moreover, because the FALL benchmark is synthetic and AIR is proprietary, the only independently checkable comparison where HsVT beats RVT is the 0.006 GEN1 gap, which is not statistically established.","section":"Section 5.4; Table 8"}],"minor_comments":[{"comment":"There is a duplicated article in 'The The Aircraft Detection Dataset'; please fix this typo.","section":"Section 5.1"},{"comment":"'fast interference' should be 'fast inference' in the description of RVT.","section":"Section 2.3"},{"comment":"'identifie' is a typo; it should be 'identify'.","section":"Section 3.1"},{"comment":"The column header 'FALL' followed by '1000ms 200ms AIR' is confusing; please label the columns as separate dataset/time-interval conditions. Also, the 200ms FALL mAP in Table 4 (0.490 for ATan) differs from the value 0.487 in Table 2 for the same interval, and this discrepancy should be reconciled.","section":"Table 4"},{"comment":"The Tiny row reports E_SNN = 0.017 mJ, but Equation (6) with SOP = 1156.00M gives approximately 1.04 mJ, matching the scale of the Small and Base rows; this looks like a typo and should be corrected.","section":"Table 9"},{"comment":"The reference list includes Spikformer (Zhou et al., 2022) but the text does not cite it; please either cite it where relevant or remove it.","section":"References"},{"comment":"The dataset release is only described as 'our Dropbox repository'; for a public benchmark, please provide a persistent identifier such as a DOI or an institutional repository link.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper lacks a code repository, and the dataset link is a non-persistent Dropbox URL. The Table 7 baseline values should be checked carefully against the original papers, since two entries are clearly off by a factor of 100. The central quantitative claim is not supported by the reported statistics, but the issues are fixable with additional experiments and honest reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nQuick take: HsVT is a competent, incremental extension of RVT that swaps some LSTM cells for a spiking module (STFE) and adds a simulated fall-detection benchmark. The work is coherent, the ablations are genuinely informative, and the energy estimates are a plus. But the claim of \"significant performance improvements\" in the abstract is not established. On GEN1, the best variant beats RVT by 0.006 mAP, and the tiny/small variants are below RVT. No error bars, no seeds, no significance tests anywhere. For event-based detection training, 0.006 mAP is within normal run-to-run variance. So the only consistent wins come from the synthetic FALL set and the non-public AIR set, which makes the load-bearing evidence thin.\n\nWhat is actually new: the STFE module, a spiking LSTM-like temporal feature extractor placed in the last block, is a reasonable idea, and the placement ablations (Tables 5-6) show it helps and where it works best. The FALL benchmark, despite being generated from Le2i frames via ESIM, could be useful if the data link worked and the paper were honest about the capture method. Right now the abstract says \"captured using an event-based camera\" while Section 3.1 says it was converted with a simulator. That contradiction needs fixing.\n\nSoft spots: Table 7 has likely typos (SpikSSD 40.8 and STAT 49.9 should be 0.408 and 0.499), the AIR dataset is unavailable for verification, and the dt=200ms choice for FALL is tuned to a synthetic stream, so it may not transfer to a real sensor. The paper also ships no code, which makes the numbers hard to check. None of these are fatal, but they erode confidence in the specific values.\n\nWho gets value: people working on hybrid ANN-SNN detectors for event cameras, and anyone who wants a simulated fall-detection benchmark. A serious referee should engage with this. The architecture is worth discussing, the ablations are mostly transparent, and the issues are fixable in revision. The authors need to release code, fix the abstract, run multiple seeds, and tone down \"significant.\" If that happens, the marginal gains would be a credible incremental step.\n\nMy advice: send to review, with a clear note to the authors to provide the dataset, code, and statistical context.\n\nBest.","headline":"HsVT is a coherent hybrid ANN-SNN detector with a real spiking temporal module and a useful simulated benchmark, but the headline gains over RVT are within run-to-run noise and the significance claim is not supported.","tokens_in":14806,"tokens_out":2499,"would_cite":true,"duration_ms":24362,"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":"The paper claims that a hybrid ANN-SNN vision Transformer, HsVT, matches or beats the RVT reference on event-based object detection while using fewer parameters, and introduces a privacy-preserving fall-detection event benchmark.","keywords":["event-based object detection","spiking neural networks","vision transformer","hybrid ANN-SNN","event camera","fall detection dataset","parameter efficiency","temporal feature extraction"],"falsifier":"Record the same fall scenarios with a real event camera and run HsVT and RVT under identical training and evaluation; if the real-data mAP gap does not reproduce the synthetic-data gap, the fall-detection advantage is an artifact of simulation. A simpler check is to corrupt or re-render the synthetic events with realistic sensor noise and see whether the claimed margin survives.","tokens_in":13824,"feed_emoji":"⚡","tokens_out":11385,"duration_ms":92529,"temperature":0.7,"pith_summary":"HsVT is a four-block hybrid network for event-camera object detection: ANN-style block and grid self-attention paired with SpikingMLP extract spatial features, while LSTM cells in the first three blocks and a spiking temporal-feature module in the last block carry information across time. The paper's claim is that this combination is competitive with the strong RVT baseline on GEN1—HsVT-B reaches 0.478 mAP 50:95 with 17.2M parameters against RVT's 0.472 with 18.5M—and better than RVT on the FALL and AIR datasets for every model size. The paper also releases the Fall Detection Dataset, an event-stream benchmark produced by converting Le2i fall videos through the ESIM simulator, arguing that event representation protects privacy and reduces storage. A sympathetic reader would see this as evidence that spiking temporal modules can replace heavier recurrent units without giving up accuracy, which matters for low-power, privacy-conscious monitoring.","feed_headline":"Event-camera detector HsVT outperforms RVT with fewer parameters","feed_subtitle":"On GEN1, HsVT-B reaches 0.478 mAP with 17.2M parameters; every size beats RVT on fall and air detection.","key_machinery":"The load-bearing mechanism is the four-block HsVT backbone with two complementary paths. Vertically, each block passes spatial features down the network; horizontally, each block receives the temporal output of the adjacent time step, letting information accumulate over time. Spatial extraction combines MaxViT's Block-SA and Grid-SA self-attention with SpikingMLP; temporal extraction uses LSTM in blocks 1 through 3 and STFE, a convolutional-batchnorm-spiking-neuron module with an LSTM-like recurrent section, in block 4. This layout is what lets the paper attribute spatial coverage to ANN attention and temporal modeling to cheap spiking units, and the ablation isolating STFE placement is the evidence that final-block placement carries the performance.","core_discovery":"The core discovery is that a deliberately hybrid computation—ANN self-attention for spatial structure, spiking neurons for temporal structure—yields a parameter-efficient event detector. On GEN1 the base HsVT reports 0.478 mAP 50:95 with 17.2M parameters, slightly above RVT's 0.472 with 18.5M; tiny and small variants reach 0.449 and 0.465. On the FALL and AIR datasets, HsVT reports higher mAP than RVT across all variants (for example, AIR tiny 0.641 versus 0.613, and FALL small 0.492 versus 0.466). The authors attribute the gain to replacing LSTM-only temporal encoding with a final-block spiking temporal feature extraction module, which in ablation reaches 0.640 mAP on AIR with 0.20M parameters and 101.19M FLOPs, together with horizontal propagation of temporal information between blocks.","pith_inferences":["An implication the authors leave implicit: if this hybrid pattern transfers, heavy LSTM temporal modules in other event-based tasks such as tracking, action recognition, or pose estimation could be replaced by spiking temporal modules, cutting parameters without necessarily losing accuracy.","The FALL benchmark's privacy and memory advantages come from the event representation itself, but the dataset is synthetic; capturing a real event-camera fall dataset with the same annotations would test whether the advantage survives real sensor noise.","The paper reports that larger variants do not consistently beat smaller ones on FALL and AIR and attributes this to dataset size and overfitting; a larger or more diverse real event-camera corpus could change the scaling behavior and model-size rankings.","The ablation finding that STFE performs best only in the final block suggests a design principle not stated by the paper: spiking temporal extraction may be most useful after spatial features are already abstract, rather than in early layers."],"forward_implications":["On GEN1, HsVT-B reaches 0.478 mAP 50:95 with 17.2M parameters, slightly above RVT's 0.472 with 18.5M, while HsVT-T and HsVT-S stay close to RVT at 4.1M and 9.1M parameters.","On FALL and AIR, every HsVT variant beats the corresponding RVT variant, with examples including AIR tiny 0.641 versus 0.613 and FALL small 0.492 versus 0.466.","The ablation supports replacing LSTM-only temporal encoding with the STFE module in the final block, which reaches the highest AIR mAP with fewer parameters than LSTM.","The released Fall Detection Dataset gives the community a public event-stream benchmark for fall detection that avoids facial exposure and stores only contrast-change events, lowering memory use.","The paper's energy estimates place HsVT-B at 134.5 mJ on GEN1, above pure-SNN detectors such as SFOD at 7.26 mJ and EAS-SNN-M at 28.10 mJ, with the added cost attributed to the ANN backbone and FPN plus head."],"supporting_citations":[{"why":"Supplies the RVT baseline and the multi-stage backbone with local and global attention and LSTM recurrence that HsVT modifies, and sets the GEN1 time interval.","marker":"(Gehrig & Scaramuzza, 2023)"},{"why":"Provides the ESIM event camera simulator used to convert Le2i frame videos into the synthetic event streams of the new Fall Detection Dataset.","marker":"(Rebecq et al., 2018)"},{"why":"Supplies the MaxViT block and grid self-attention scheme that the spatial feature extraction module is built on.","marker":"(Tu et al., 2022)"},{"why":"Defines the LSTM cells used in the first three blocks to carry temporal information between time steps.","marker":"(Hochreiter & Schmidhuber, 1997)"},{"why":"Defines the GEN1 automotive event dataset and its label schema, which the paper adopts for evaluation and for the FALL dataset format.","marker":"(De Tournemire et al., 2020)"},{"why":"Provides the Le2i frame-based fall detection dataset whose videos are converted into event streams for the new benchmark.","marker":"(Charfi et al., 2013)"},{"why":"Supplies the YOLOX detection framework that the RVT backbone and HsVT are combined with for training and evaluation.","marker":"(Ge, 2021)"}],"fun_headline_variants":["HsVT: hybrid spiking transformer outperforms RVT on event cams","Hybrid spiking and attention model HsVT tops RVT, cuts parameters","Event-based detection: HsVT beats RVT with fewer parameters","HsVT: spiking temporal, attention spatial, bests RVT on GEN1","Hybrid spiking transformer HsVT edges out RVT in event detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the synthetic event streams generated by the ESIM simulator from Le2i frame videos are faithful enough to real event-camera data, including sensor noise, dynamic range, and temporal correlations, that the reported fall-detection numbers predict real-world performance.","fun_headline_variants_meta":{"raw":{"variants":["HsVT: hybrid spiking transformer outperforms RVT on event cams","Hybrid spiking and attention model HsVT tops RVT, cuts parameters","Event-based detection: HsVT beats RVT with fewer parameters","HsVT: spiking temporal, attention spatial, bests RVT on GEN1","Hybrid spiking transformer HsVT edges out RVT in event detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000282,"raw_usage":{"total_tokens":1671,"prompt_tokens":953,"completion_tokens":718,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":619}},"tokens_in":569,"tokens_out":718,"duration_ms":6521,"temperature":1.0,"reasoning_tokens":619,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:09:45.545755+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the same fall scenarios with a real event camera and run HsVT and RVT under identical training and evaluation; if the real-data mAP gap does not reproduce the synthetic-data gap, the fall-detection advantage is an artifact of simulation. A simpler check is to corrupt or re-render the synthetic events with realistic sensor noise and see whether the claimed margin survives.","supporting_citations":[{"cited_title":"and Scaramuzza, D","cited_arxiv_id":null,"evidence_quote":"Supplies the RVT baseline and the multi-stage backbone with local and global attention and LSTM recurrence that HsVT modifies, and sets the GEN1 time interval."},{"cited_title":"ESIM : an open event camera simulator","cited_arxiv_id":null,"evidence_quote":"Provides the ESIM event camera simulator used to convert Le2i frame videos into the synthetic event streams of the new Fall Detection Dataset."},{"cited_title":"Optimised spatio-temporal descriptors for real-time fall detection: comparison of svm and adaboost based classification","cited_arxiv_id":null,"evidence_quote":"Provides the Le2i frame-based fall detection dataset whose videos are converted into event streams for the new benchmark."}],"review_version":1}