{"id":"ae909417-d84f-4883-a8d5-1e8c2a8b3e9d","arxiv_id":"2506.03162","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual-branch VideoMamba with gated class-token fusion achieves 95.85% accuracy on a newly combined violence-detection benchmark and 74.13% on DVD, with about half the parameters and FLOPs of the CUE-Net baseline.","lead":"This paper introduces a dual-branch state-space model for detecting violence in surveillance video, fusing spatial and temporal class tokens through learned gates, and reports top accuracy on a new combined benchmark and on the DVD dataset. It matters for generalists because it is evidence that efficient Mamba-style architectures can match or beat heavier CNN and Transformer baselines at lower compute.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation-driven architecture selection in §4.5 is reported on the combined dataset without stating whether a held-out validation split was used; if these numbers are test-set accuracies, the headline 95.85% is a selected maximum and the SOTA claim is unverified.","rationale":"A good-faith reading is that the paper is an applied architecture paper whose central claim is empirical: Dual Branch VideoMamba with GCTF achieves state-of-the-art accuracy with lower cost than CUE-Net. For that claim to be trustworthy, the reported test accuracies must be unbiased estimates of generalization. The most load-bearing condition is therefore that architecture selection was not performed on the same test set used for final reporting. This condition is least secure because §4.5 is silent about the split: it reports accuracy on the combined dataset for every design choice without stating whether those numbers come from validation or test. The reader's weakest_assumption identifies exactly this gap, and I agree with it. I considered alternative concerns, such as unfair baseline tuning or a confounding cropping module. Those are secondary: the ablation table in §4.5.1 shows that cropped VideoMamba baselines still lag the full model, and CUE-Net itself includes cropping. The split issue is more direct because it can bias the single headline number itself. The internal inconsistency between Tab. 5 and Supp. Tab. 13 for the same gated lateral-connection configuration reinforces the need for a precise protocol statement, but it is not the primary attack. This is a protocol omission rather than an observed fabrication; the released code may resolve it. Since the reader's verdict is already CONDITIONAL and my concern does not change that assessment, I recommend no verdict change.","tokens_in":15662,"tokens_out":9950,"duration_ms":105028,"concrete_test":"Check the released code (github.com/damith92/FuseMamba-VD) for the data-split logic used by the ablation scripts: specifically, whether the accuracy values in Tables 4-6, 12, and 13 are computed on a dedicated validation split or on the same test split used for Table 1. If the latter, the concern lands and the headline must be re-estimated on an untouched test split before the SOTA claim can be accepted; if a held-out validation split was used for all ablations, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Tab. 1: 95.85% on the combined dataset, 74.13% on DVD) requires that the final architecture was selected without fitting to the test set. Section 4.5 reports every design choice—cropping and skip connections (Tab. 4), lateral fusion mechanism (Tab. 5), placement of lateral connections (Tab. 6), frame counts (Supp. Tab. 12), and final fusion (Supp. Tab. 13)—as 'Accuracy' on 'the combined dataset,' but never states that these ablations used a held-out validation split rather than the training/test split described in §4.1. Section 4.3 calls the reported numbers 'test accuracies,' and Supp. §6 only says the best validation model was saved, not that the ablation tables were computed on validation. With roughly 38 configurations evaluated on 915 test clips, the headline 95.85% is naturally read as the best of many test-set selections; the margin over CUE-Net (94.97%) is only about 8 clips, so selection bias could plausibly erase or reverse the advantage. The concern is strengthened by an internal inconsistency in the fusion ablations: Tab. 5 reports Gated LCs (Branch-1→Branch-2) as 95.85%, while Supp. Tab. 13 reports the same configuration as 95.30%, so it is not clear which configuration produced the headline number.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FuseMamba-VD (also titled 'Dual Branch VideoMamba with Gated Class Token Fusion') for violence detection in surveillance video. The architecture uses two parallel VideoMamba branches, one with spatial-first scanning and one with temporal-first scanning, and continuously fuses their class tokens through a learnable gating mechanism. The authors also introduce a combined benchmark built from RWF-2000, RLVS, SURV, and VioPeru with a duplicate-removal step, and they evaluate on the recently introduced DVD dataset. They report state-of-the-art accuracy on both benchmarks (95.85% combined, 74.13% DVD) with substantially lower parameters and FLOPs than the strongest CNN+Transformer baseline CUE-Net.","tokens_in":15900,"tokens_out":5078,"duration_ms":43086,"significance":"If the reported results hold after proper validation, the paper would make a useful contribution: it demonstrates that a dual-branch SSM design with gated class-token fusion can match or beat heavier CNN/Transformer baselines for violence detection, with a large efficiency advantage. The ablation study is informative about the contribution of cropping, skip connections, lateral fusion direction, and continuous fusion. The authors also provide code and pretrained models, which supports reproducibility. However, the core 'state-of-the-art' claim currently rests on test-set selection details and an internal inconsistency in the fusion ablations, so the significance can only be assessed after those issues are resolved.","major_comments":[{"comment":"The ablations in §4.5 (Tabs 4–6) and Supplementary Tabs 12–13 report accuracies on the combined dataset but never state that these numbers come from a held-out validation split rather than the test set described in §4.1. Section 4.3 calls the reported numbers 'test accuracies,' and Supplementary §6 only says that the best validation model was saved. If the test set was used to choose among the many ablated configurations, the headline 95.85% is a selected maximum and the 0.88-point margin over CUE-Net (about 8 clips out of 915) could be due to selection bias. Please specify the validation protocol used for model selection, or re-report the ablations on a held-out validation split.","section":"§4.5 and §4.1"},{"comment":"Table 5 reports 'Gated LCs (Branch-1→Branch-2)' as 95.85%, while Supplementary Table 13 reports the same named configuration as 95.30%. The paper does not explain whether these rows refer to different components (lateral fusion vs. final fusion) or whether one of the numbers is an error. Because the headline accuracy is 95.85%, this inconsistency must be resolved before the state-of-the-art claim can be evaluated.","section":"§4.5.2 vs. Supplementary §7.1.2"},{"comment":"All numbers in Table 1 appear to be single-run accuracies with no variance, confidence intervals, or repeated-seed information. The central comparisons against CUE-Net (95.85 vs. 94.97 on the combined test set; 74.13 vs. 73.68 on the DVD test set) are not accompanied by any statistical test, and the only significance test reported (McNemar) is against VideoMamba-Large, not against the main competitor. To support the claim of state-of-the-art performance, report mean±std over multiple runs or an appropriate significance test for the CUE-Net comparison.","section":"§4.3, Table 1"}],"minor_comments":[{"comment":"In Eq. (2), the notation \\bar{B} = (\\Delta A)^{-1}(\\exp(\\Delta A)-I)\\cdot\\Delta B is confusing because \\Delta appears both as a timescale parameter and as a multiplier on B. Please clarify the discretization formula, for example by writing \\bar{B} = (\\Delta A)^{-1}(e^{\\Delta A}-I)\\Delta B with an explicit definition of \\Delta B.","section":"Equation (2)"},{"comment":"The header of Table 4 ('Cropping -× Cropping -✓') is difficult to parse. Please use explicit column labels such as 'Cropping: off/on' and 'Skip: off/on' so the four configurations are clear.","section":"Table 4"},{"comment":"The text says 'we compare four fusion mechanisms for lateral connections,' but Table 5 lists six configurations (full-hidden-state concatenation, CLS-token concatenation, additive, cross-attention, and gated in both directions). Please correct the count or restructure the table.","section":"§4.5.2"},{"comment":"The arXiv metadata title is 'FuseMamba-VD: Dual Branch VideoMamba with Gated Class Token Fusion for Violence Detection,' while the manuscript title is 'Dual Branch VideoMamba with Gated Class Token Fusion for Violence Detection.' Please align these titles.","section":"Title/arXiv metadata"},{"comment":"The claim of being 'the first state-space-based design for violence detection' needs stronger support. The related work mentions STNMamba for video anomaly detection; since anomaly detection and violence detection are closely related, the paper should explicitly discuss why prior SSM-based anomaly-detection work does not count, or soften the 'first' claim.","section":"§1, Contribution 1"},{"comment":"The duplicate-removal procedure is not fully specified: after manually inspecting pairs with 75% or higher cosine similarity, what criterion led to removing exactly one RLVS test video? Please state the threshold or decision rule used for removal so the benchmark construction is reproducible.","section":"§4.1.1"},{"comment":"The abbreviation 'FLOPS' should be 'FLOPs' throughout the manuscript.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The DVD dataset is introduced in a companion paper [22] by overlapping authors. The manuscript should disclose this connection and confirm that the splits and annotations used here are identical to those in [22]. The combined benchmark is a useful contribution, but the exact split manifests and duplicate-removal code should be released alongside the model code to make the benchmark reproducible. The comparisons lean heavily on the authors' own CUE-Net baseline; independent re-implementations or additional baselines would strengthen the evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper is a competent applied-ML contribution that deserves serious refereeing, but the headline SOTA claim needs a clarification and a fairer comparison before I'd trust the numbers.\n\nWhat's actually new: it's the first SSM-based architecture for violence detection, and the dual-branch VideoMamba with gated class-token fusion is a genuine architectural variant, not a trivial repackaging. The ablations are thorough (cropping, skip connections, fusion mechanism, placement, frame counts), the efficiency gain over CUE-Net is real (154M vs 354M params, 1830 vs 5826 GFLOPs), and the combined benchmark of RWF-2000/RLVS/SURV/VioPeru is a useful resource. Code and pretrained models are promised. Those are real strengths.\n\nThe main soft spot is exactly what the stress-test flags: Section 4.5 reports every ablation accuracy 'on the combined dataset' without stating whether those numbers came from a held-out validation split or the test set. The implementation section says the best validation model was saved, but that doesn't tell us which split produced the ablation tables. With roughly 38 configurations and a margin of only 8 clips over CUE-Net (95.85% vs 94.97%), test-set selection could easily erase that advantage. This is a missing methodological sentence, not necessarily a fatal flaw, but the authors need to state it explicitly.\n\nA second soft spot: the main table compares against VideoMamba-M/L without cropping, while the dual branch uses cropping. The ablation in Table 4 shows cropped VideoMamba-L with skip reaches 93.55%, still below the dual branch's 95.85%, so the architecture's contribution survives that confound, but the main table should include the cropped baseline for an apples-to-apples comparison. The DVD dataset results come from a benchmark introduced by the same group in a companion paper; that's self-referential but not disqualifying.\n\nThe apparent inconsistency between Tab. 5 (95.85%) and Supp. Tab. 13 (95.30%) for 'Gated LCs (Branch-1→Branch-2)' looks like a contradiction, but reading carefully, they are different ablations: one varies the lateral fusion mechanism with concatenation in the final block, the other varies the final fusion block with gated LCs as the final merger. Not the same configuration, though the paper could have made that clearer.\n\nBottom line: worth refereeing. The architecture is plausible, the benchmark is useful, and the efficiency story is solid. I'd ask the authors to state the ablation split, add the cropped VideoMamba baseline, and report variance or confidence intervals. If the numbers hold, this is a solid subfield-level contribution.","headline":"Solid applied-ML paper with a plausible architecture and a useful benchmark, but the headline SOTA needs a clearer statement of the ablation split and a fairer baseline before the numbers can be trusted.","tokens_in":16503,"tokens_out":4656,"would_cite":true,"duration_ms":40558,"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":"A dual-branch state-space model with gated class-token fusion claims 95.85% test accuracy on violence detection while using less than half the compute of the previous best model.","keywords":["violence detection","state-space models","VideoMamba","gated class token fusion","surveillance video","dual-branch architecture","video classification","benchmark dataset"],"falsifier":"Re-run the Section 4.5 ablations on a held-out validation split of the combined dataset, choose the configuration with the best validation accuracy, and then report its test accuracy; if that test accuracy is substantially below 95.85%, the headline result is an artifact of model selection on the test set.","tokens_in":15406,"feed_emoji":"🎥","tokens_out":7895,"duration_ms":52236,"temperature":0.7,"pith_summary":"This paper aims to show that a state-space model (SSM) backbone—specifically a dual-branch VideoMamba—can outperform both CNN- and Transformer-based architectures for automated violence detection in surveillance video while being much cheaper to run. The authors report 95.85% test accuracy with F1-scores of 95.89% and 95.81% on a newly amalgamated benchmark built from RWF-2000, RLVS, SURV, and VioPeru, and 74.13% accuracy on the DVD dataset, achieved with 154M parameters and 1830 GFLOPS, which is more than a 50% reduction compared with the prior best model, CUE-Net. The work also contributes a leakage-checked combination of four strongly labeled datasets as a new benchmark. If these results hold, SSMs become a practical and scalable alternative to attention-heavy video models for real-world surveillance.","feed_headline":"State-space model hits 95.85% accuracy for violence detection","feed_subtitle":"A gated dual-branch VideoMamba beats CNN and Transformer hybrids while using 154M parameters, not 354M.","key_machinery":"The central object is the Dual Branch VideoMamba with Gated Class Token Fusion (GCTF): two parallel VideoMamba encoders—one scanning patch tokens spatial-first, the other temporal-first—each with its own learnable class token and initialized from Kinetics-400 pretrained weights. At every block, a learnable sigmoid gate combines the spatial branch's class token with the temporal branch's class token according to $\\sigma'_l \\odot \\mathrm{CLS}^2_l + (1-\\sigma'_l) \\odot \\mathrm{CLS}^1_l$, and the fused CLS token is passed into the temporal branch. A cropping module (YOLOv8-based) extracts the maximum bounding box around all detected people, and a final block concatenates the two branches' final CLS tokens before classification. The gating is what carries the argument: it performs cheap, layer-wise semantic fusion without the token-space misalignment that ablations show when full hidden states are concatenated.","core_discovery":"The paper's central claim is that continuous fusion of class tokens between two parallel scanning branches is what unlocks the performance: a spatial-first VideoMamba branch captures fine-grained spatial cues, a temporal-first branch captures motion dynamics, and a learnable sigmoid gate at every layer injects the spatial branch's CLS token into the temporal branch's CLS token, avoiding premature commitment to either stream. On top of this, a YOLOv8-based cropping module focuses the network on detected people, and a final concatenation of the two class tokens feeds a classification head. The authors report that this design achieves state-of-the-art accuracy on the combined benchmark and on DVD, and that the improvement over the comparable-size VideoMamba-Large is statistically significant under McNemar's test.","pith_inferences":["Inference: The paper does not explicitly state that the ablation studies used a held-out validation split; if they used the combined test set for model selection, the 95.85% headline is a selected maximum and likely overestimates generalization. A validation-based re-run would settle this.","Inference: The asymmetric benefit of spatial-to-temporal gating (95.85%) over temporal-to-spatial gating (93.22%) suggests spatial context acts as a prior that guides motion interpretation; this asymmetry could be tested in other SSM-based video models.","Inference: The VideoMAE cosine-similarity leakage protocol is itself a reusable contribution; other benchmarks that merge datasets from overlapping sources could adopt it to avoid hidden duplicate videos.","Inference: The larger gap on DVD (74.13%) relative to the combined benchmark hints that in-the-wild distribution shift, not architecture capacity, is the current bottleneck; multimodal signals or surveillance-specific pretraining might close it."],"forward_implications":["State-space models become a credible backbone for violence detection, offering a linear-complexity alternative to quadratic-attention transformers.","The reported efficiency (154M parameters, 1830 GFLOPS versus CUE-Net's 354M and 5826) supports near real-time deployment of surveillance analytics on modest hardware.","Gated class-token fusion from a spatial branch into a temporal branch may transfer to other video-understanding tasks that need long-range dependencies, such as anomaly detection.","The leakage-checked amalgamated benchmark provides the field with a standardized testbed for cross-dataset generalization in violence detection.","Per-dataset improvements over previously reported bests (94.50% on RWF-2000, 99.75% on RLVS, 96.67% on SURV) indicate the architecture captures complementary cues across surveillance conditions."],"supporting_citations":[{"why":"Supplies the VideoMamba state-space backbone and the Kinetics-400 pretrained weights used to initialize both branches.","marker":"[26]"},{"why":"Defines the bidirectional Vision Mamba block that each branch's encoder repeats.","marker":"[48]"},{"why":"Introduces the selective state-space model (S6) giving the architecture linear-complexity long-range modeling.","marker":"[14]"},{"why":"CUE-Net, the previous state-of-the-art baseline whose accuracy and compute the paper claims to beat, and source of the cropping-module practice.","marker":"[35]"},{"why":"RWF-2000, one of the four datasets merged into the combined benchmark and a standard real-world fight-detection test.","marker":"[9]"},{"why":"RLVS, one of the four merged datasets and the source of the duplicate removed from the test set.","marker":"[36]"},{"why":"SURV, one of the four merged surveillance datasets used in the combined benchmark.","marker":"[2]"},{"why":"VioPeru, one of the four merged datasets and the benchmark where the model matches the previous best accuracy.","marker":"[16]"},{"why":"DVD, the second evaluation dataset used to demonstrate generalization to a large, challenging in-the-wild benchmark.","marker":"[22]"},{"why":"VideoMAE embeddings used in the cosine-similarity check that guards against data leakage when merging datasets.","marker":"[42]"}],"fun_headline_variants":["Gated class token fusion: 95.85% violence detection","FuseMamba-VD: efficient dual-branch VideoMamba for violence detection","State-space model hits 95.85% violence detection, SOTA on combined benchmark","Dual-branch SSM with gated fusion: SOTA violence detection in 154M params","Violence detection: gated VideoMamba outperforms CNN-Transformers at half size"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the architectural choices picked in the ablation studies were selected on a held-out validation split; if the combined test set was reused for that selection, the reported 95.85% test accuracy is a selected maximum and not an unbiased estimate of generalization.","fun_headline_variants_meta":{"raw":{"variants":["Gated class token fusion: 95.85% violence detection","FuseMamba-VD: efficient dual-branch VideoMamba for violence detection","State-space model hits 95.85% violence detection, SOTA on combined benchmark","Dual-branch SSM with gated fusion: SOTA violence detection in 154M params","Violence detection: gated VideoMamba outperforms CNN-Transformers at half size"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000371,"raw_usage":{"total_tokens":1979,"prompt_tokens":930,"completion_tokens":1049,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":952}},"tokens_in":546,"tokens_out":1049,"duration_ms":7616,"temperature":1.0,"reasoning_tokens":952,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:41:42.553702+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Section 4.5 ablations on a held-out validation split of the combined dataset, choose the configuration with the best validation accuracy, and then report its test accuracy; if that test accuracy is substantially below 95.85%, the headline result is an artifact of model selection on the test set.","supporting_citations":[{"cited_title":"Videomamba: State space model for efficient video understanding","cited_arxiv_id":null,"evidence_quote":"Supplies the VideoMamba state-space backbone and the Kinetics-400 pretrained weights used to initialize both branches."},{"cited_title":"Vision mamba: efficient visual representation learning with bidirectional state space model","cited_arxiv_id":null,"evidence_quote":"Defines the bidirectional Vision Mamba block that each branch's encoder repeats."},{"cited_title":"Cue-net: Violence detection video analytics with spatial cropping, enhanced uniformerv2 and modified efficient additive attention","cited_arxiv_id":null,"evidence_quote":"CUE-Net, the previous state-of-the-art baseline whose accuracy and compute the paper claims to beat, and source of the cropping-module practice."},{"cited_title":"Rwf-2000: an open large scale video database for violence detection","cited_arxiv_id":null,"evidence_quote":"RWF-2000, one of the four datasets merged into the combined benchmark and a standard real-world fight-detection test."},{"cited_title":"Violence recog- nition from videos using deep learning techniques","cited_arxiv_id":null,"evidence_quote":"RLVS, one of the four merged datasets and the source of the duplicate removed from the test set."},{"cited_title":"Vision-based fight detection from surveillance cam- eras","cited_arxiv_id":null,"evidence_quote":"SURV, one of the four merged surveillance datasets used in the combined benchmark."},{"cited_title":"Efficient human violence recognition for surveillance in real time.Sensors, 24(2):668, 2024","cited_arxiv_id":null,"evidence_quote":"VioPeru, one of the four merged datasets and the benchmark where the model matches the previous best accuracy."}],"review_version":1}