{"id":"825e3402-5e2f-4892-a79f-7e96fbd6af72","arxiv_id":"2412.20084","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"STNMamba applies Mamba state space models to video anomaly detection, reporting marginal AUC gains over previous methods with drastically lower FLOPs and parameter counts.","lead":"STNMamba is a video anomaly detection network based on Mamba state space models, with dual encoders and memory banks that predict the next frame and flag unusual events. It reports slightly higher accuracy than prior methods on three benchmarks while using a fraction of the compute, making it appealing for real-time surveillance if the efficiency claims hold.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set hyperparameter selection undermines the SOTA claim: tau and k are chosen by maximizing AUC on the test sets themselves, so the reported 98.0/89.0/74.9 are not independent test-set numbers.","rationale":"Reading in good faith, the paper proposes a Mamba-based VAD architecture and claims state-of-the-art accuracy on three benchmarks at low cost. For that claim to hold, the reported AUCs must be genuine out-of-sample values. The weakest link in that chain is explicit test-set hyperparameter selection: the sensitivity study in Sec. IV-C4 picks tau and k using the test datasets and then reports those test AUCs as the method's performance. This is an internal validity problem, not a disagreement with community consensus, and it can be fixed by a validation split. I do not see independent reason to doubt the architecture's plausibility or the low parameter count; the ablations show consistent small gains. But the 0.3-0.7 point margins over prior methods are too small to rule out selection bias. The reader's formal weakest assumption was efficiency comparability; I find that real but secondary to the performance claim. The verdict should remain CONDITIONAL: the method may work, but the SOTA claim needs re-validation before acceptance. Since this is the same verdict the reader reached, no adjustment is needed.","tokens_in":796,"tokens_out":1313,"duration_ms":70705,"concrete_test":"Carve a validation split from each dataset's training videos (e.g., 20% of training videos, or leave-one-training-video-out on Ped2), tune tau in {0.2, ..., 1.0} and k in {20, ..., 100} on that split, fix the selected values, then evaluate on the test set exactly once; repeat with three random seeds and report mean plus or minus std. If the best test AUCs no longer exceed PDM-Net on Ped2/Avenue, or the margins shrink to noise, the SOTA claim is not established. This also yields an honest ShanghaiTech number and the selected hyperparameters for reproduction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is the frame-level SOTA performance. The paper's own sensitivity analysis (Sec. IV-C4, Fig. 5) selects tau and k by maximizing test AUC on UCSD Ped2 and CUHK Avenue. tau enters the final anomaly score in Eq. (18); k controls the top-k memory retrieval in Sec. III-D2 (Eqs. 8-9). Both are free hyperparameters, and the selection is performed on the same videos that later appear in the reported 98.0% and 89.0%. The gains over the second-best methods are only 0.3 and 0.5 AUC points, exactly the scale that such selection can create. No held-out validation, repeated-seed statistics, or released code is provided, so the 74.9% on ShanghaiTech is also not demonstrably independent of model choices. The efficiency comparison in Table I is additionally under-specified (input resolution, frame count, and measurement protocol for the 1.5G FLOPs), but it is secondary: the headline is state-of-the-art performance with fewer parameters.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes STNMamba, a Mamba-based network for unsupervised video anomaly detection. It uses a dual-encoder design (MS-VSSB for multi-scale spatial features, CA-VSSB for temporal motion cues), a Spatial-Temporal Interaction Module with fusion blocks and memory banks, and a decoder that predicts the next frame. Anomaly scores combine prediction PSNR and memory-retrieval distance. On UCSD Ped2, CUHK Avenue, and ShanghaiTech the authors report frame-level AUCs of 98.0%, 89.0%, and 74.9%, respectively, and claim state-of-the-art performance with 1.5G FLOPs, 7.2M parameters, and 40 FPS.","tokens_in":16305,"tokens_out":3748,"duration_ms":36923,"significance":"The architecture is a reasonable and potentially useful first step toward Mamba-based VAD: the dual-encoder with multi-scale and channel-aware Mamba blocks, the multi-level spatial-temporal interaction, and the memory-augmented score are clearly described, and the component ablations (Tables II-IV) show that each proposed block contributes positively. If the reported results are reproducible under a clean evaluation protocol, the efficiency advantage would be practically relevant for edge deployment. The paper is well written and the method is novel relative to the cited CNN/transformer baselines. However, the state-of-the-art claim is currently not supported because the key hyperparameters are tuned on the test sets and the efficiency table is not measured under a single comparable protocol.","major_comments":[{"comment":"The hyperparameters tau and k are selected by maximizing frame-level AUC on the test sets of UCSD Ped2 and CUHK Avenue, and the same test sets are then used for the reported 98.0% and 89.0% results; the gains over the second-best methods (0.3 and 0.5 points) are of the same scale as this selection bias, so the state-of-the-art claim is not established. Please either fix tau and k on a held-out validation split, report results without test-set selection, or provide repeated-seed statistics with error bars; as written, the 74.9% on ShanghaiTech is also not demonstrably independent of model choices.","section":"Section IV-C4, Fig. 5, Eq. (18)"},{"comment":"The efficiency comparison is under-specified. The text says some entries were collected from original papers and others reproduced, but it does not state the input resolution, number of frames, or exact GPU and measurement protocol for the reproduced runs, nor how the 1.5G FLOPs, 7.2M parameters, and 40 FPS for STNMamba were measured. Because the central claim of 'fewer parameters and lower computational costs' rests on this table, please report a uniform measurement setup (same resolution, frame count, hardware, and software versions) and ideally release the evaluation code.","section":"Section IV-B2, Table I"}],"minor_comments":[{"comment":"The term F_i_m is used in the definition of the memory-enhanced feature but is never defined; please state explicitly that it is the reshaped memory-reconstructed feature obtained from Eq. (7).","section":"Section III-D2, Eq. (11)"},{"comment":"The notation m1_i and m2_i is ambiguous because the subscript i simultaneously indexes the query and the rank of the nearest memory item; please define these quantities explicitly.","section":"Section III-F, Eq. (14)"},{"comment":"The PSNR formula should use the square of the maximum pixel value in the numerator; as written, the expression is dimensionally inconsistent.","section":"Section III-G, Eq. (16)"},{"comment":"The statement that MNAD 'achieves slightly faster inference' is inaccurate given Table I, which reports 65 FPS for MNAD versus 40 FPS for STNMamba; please revise the wording.","section":"Section IV-B2"},{"comment":"The comparison with ASTT cites reference [6], but ASTT is reference [12] in the bibliography; please correct the citation.","section":"Section IV-B2"}],"recommendation":"major_revision","confidential_remarks":"The main issue is experimental rigor rather than architectural novelty; I would encourage the authors to resubmit after redoing the hyperparameter selection and efficiency measurements under a transparent protocol. The architectural contribution is promising, but the current validation does not support the headline state-of-the-art claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a plausible, lightweight Mamba-based VAD with a real efficiency advantage, but the headline AUCs are not honest test-set numbers because tau and k were selected on the test sets themselves. The stress-test note is correct, and it is the main issue.\n\nWhat is actually new: applying Mamba to VAD is a sensible idea, and the paper is likely among the first to do it. The architecture is a clean combination of existing parts — VSSB from VMamba, depthwise convs, ECA, and memory banks from MemAE/MNAD — but the multi-level spatial-temporal fusion with hierarchical memory is a reasonable contribution. The efficiency numbers (1.5G FLOPs, 7.2M params, 40 FPS) are the strongest part of the paper, and the ablations are thorough enough to show each component does something. The writing is clear and the related work is fine.\n\nThe soft spots are serious. Section IV-C4 and Fig. 5 explicitly pick tau=0.8 and k=60 by maximizing AUC on the UCSD Ped2 and CUHK Avenue test sets, and those exact settings produce the 98.0% and 89.0% in Table I. The ShanghaiTech number is not shown in the sensitivity plots, but given there is no held-out validation anywhere, it is not demonstrably independent either. The margins over second-best are only 0.3, 0.5, and 0.7 AUC points — exactly the scale that test-set hyperparameter tuning can manufacture. No error bars, no repeated seeds, no code. The efficiency comparison in Table I is also under-specified: some numbers are quoted from papers, others reproduced, but the input resolution, frame count, and measurement protocol for the 1.5G FLOPs are not stated. That said, the efficiency advantage is large enough that I doubt it would vanish under fair measurement.\n\nNovelty is modest but real. The 'first Mamba-VAD' claim is plausible but should be checked against concurrent work; the authors do phrase it as 'to the best of our knowledge,' which is honest.\n\nWho gets value: VAD researchers working on efficient models will find the architecture useful, and the paper is a good case study in how test-set hyperparameter selection can corrupt benchmark claims. The proper fix is a re-run with a validation split before final numbers, plus a more careful efficiency benchmark.\n\nRecommendation: send it to peer review. The efficiency direction deserves referee time, and the methodological flaw is fixable, not fatal to the core approach. But a serious referee should ask for the re-run before the numbers are taken at face value.","headline":"The efficiency story is real and the Mamba-VAD architecture is worth a look, but the reported SOTA numbers are compromised by test-set hyperparameter selection.","tokens_in":16990,"tokens_out":1848,"would_cite":false,"duration_ms":20788,"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":"Mamba model tops video anomaly detection at 7.2M params","keywords":["video anomaly detection","Mamba","state space models","normality learning","spatial-temporal consistency","memory network","efficient video surveillance","future frame prediction"],"falsifier":"Measure STNMamba and the top compared baselines (e.g., MNAD and ASTT) on the same GPU with the same input resolution (256×256) and the same number of input frames, and recompute FLOPs and parameters with a standard profiler; if STNMamba's numbers rise to or above the baselines' when measured uniformly, the central efficiency claim is falsified.","tokens_in":15874,"feed_emoji":"🎥","tokens_out":5199,"duration_ms":43866,"temperature":0.7,"pith_summary":"STNMamba is a new network for unsupervised video anomaly detection built on Mamba, the state-space sequence model with linear-time complexity. The paper claims that by adapting Mamba's selective scanning to both appearance and motion streams—multi-scale in space, channel-aware in time—and by fusing the two streams at multiple levels with memory-stored normal prototypes, a lightweight model can outperform CNN- and transformer-based detectors. On three benchmarks the method reports frame-level AUCs of 98.0%, 89.0%, and 74.9%, with only 7.2 million parameters and 1.5 GFLOPs per frame. The contribution matters because existing high-accuracy video anomaly detectors are computationally heavy, and the paper argues its design is the first to bring Mamba's efficiency to this task while also improving accuracy.","feed_headline":"Mamba model tops video anomaly detection at 7.2M params","feed_subtitle":"98.0% AUC on UCSD Ped2 and 89.0% on CUHK with only 1.5G FLOPs and 7.2M parameters.","key_machinery":"The load-bearing mechanism is the Vision Space State Block (VSSB) with 2D-Selective-Scan (SS2D)—the visual adaptation of Mamba's selective scan that flattens a feature map along four directions and processes it with linear complexity. Around this core the paper wraps three additions: MS-VSSB adds parallel depth-wise convolutions (kernels 1×1, 3×3, 5×5) before the VSSB for multi-scale appearance; CA-VSSB adds a channel-attention branch (average and max pooling) to emphasize motion-relevant channels; and the Spatial-Temporal Interaction Module (STIM) uses an STFB (linear + depth-wise conv + Hadamard interaction + another SS2D + ECA channel selection) to project spatial and temporal features into a shared space at each level, followed by a learnable memory bank with top-k reading and softmax writing to store prototypes of normal patterns. This multi-level fusion plus memory constraint is what the paper argues lets Mamba model spatial-temporal consistency without quadratic attention cost.","core_discovery":"The paper's central claim is that Mamba—a selective state-space model—can be made the backbone of a prediction-based video anomaly detector that learns spatial-temporal normality better and far more cheaply than CNN or transformer alternatives. The authors introduce STNMamba, which uses a spatial encoder of Multi-Scale Vision Space State Blocks (MS-VSSB) to aggregate appearance features at several kernel scales, a temporal encoder of Channel-Aware Vision Space State Blocks (CA-VSSB) to emphasize motion channels from RGB frame differences, and a Spatial-Temporal Interaction Module (STIM) that fuses the two streams at every level using Spatial-Temporal Fusion Blocks (STFB) and reads/writes a memory bank of normal prototypes. Anomalies are scored by combining future-frame prediction error (PSNR) with feature distance to the nearest memory item. According to the paper, this achieves state-of-the-art frame-level AUCs of 98.0% on UCSD Ped2, 89.0% on CUHK Avenue, and 74.9% on ShanghaiTech, while running at 40 FPS with 7.2M parameters and 1.5G FLOPs, and it is the first Mamba-based method reported for video anomaly detection.","pith_inferences":["The efficiency comparison in Table I mixes numbers the authors reproduced with numbers taken from original papers; if those were measured at different input resolutions or with different frame counts, the 1.5G FLOPs/7.2M params advantage over MNAD and ASTT could shrink, so a standardized measurement protocol would settle this.","Mamba's linear complexity should make the gap in FLOPs grow as input resolution increases; a natural test is to measure STNMamba at 640×360 or 856×480 (the native dataset resolutions) rather than the reported 256×256.","The normality-score curve in Fig. 6 suggests detection latency is short (scores drop sharply at anomaly onset); a frame-level latency analysis over many videos would quantify whether this holds for rare, gradual anomalies such as loitering.","Because the temporal encoder uses only RGB frame differences instead of optical flow, the method may be less sensitive to small, slow motions; evaluating on scenes with subtle anomalies (e.g., a purse snatch) would test whether channel attention compensates for the lack of flow."],"forward_implications":["If the reported efficiency numbers hold under comparable measurement, Mamba-based detectors can serve as a real-time option for video surveillance on edge devices, where transformer-based models are too heavy.","The dual-encoder plus multi-level fusion design gives a template for applying state-space models to other video understanding tasks that need spatial-temporal consistency, such as video prediction and action recognition.","The memory-bank mechanism in STIM suggests that normal-prototype constraints can be attached at every level of a Mamba encoder, not just the bottleneck, and that this hierarchical memory is what drives the reported gains.","The paper's ablations indicate that each component (multi-scale spatial, channel-aware temporal, fusion block, memory, multi-level fusion) contributes additively, so the architecture can be pruned to fit even tighter compute budgets with a predictable drop in AUC.","On ShanghaiTech, the largest and most complex benchmark, the reported 74.9% AUC would make Mamba the best prediction-based unsupervised result, which is the strongest evidence for the method's scalability."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space model that the whole architecture builds on for linear-complexity long-range modeling.","marker":"[17]"},{"why":"Provides the Vision Space State Block and 2D-Selective-Scan (SS2D) that the encoders and fusion blocks are constructed from.","marker":"[19]"},{"why":"Introduces memory-augmented autoencoders for normality memorization, the conceptual basis of the memory banks used in STIM.","marker":"[4]"},{"why":"Memory-guided normality baseline that STNMamba is compared against for both frame-level AUC and efficiency (1/2 params, 1/40 FLOPs claimed).","marker":"[13]"},{"why":"Establishes future-frame prediction as the VAD baseline and defines the PSNR-based anomaly scoring that STNMamba adopts.","marker":"[5]"},{"why":"Motivates the use of RGB frame differences to capture short-term motion without optical flow, which the temporal encoder relies on.","marker":"[8]"},{"why":"Appearance-motion dual-stream baseline with bottleneck fusion, the architecture STIM improves upon with multi-level interaction.","marker":"[7]"}],"fun_headline_variants":["Mamba video anomaly detector: 40 FPS, 7.2M params","STNMamba: Mamba-based VAD with 98% AUC on Ped2","Lightweight Mamba detects video anomalies at 40 FPS","First Mamba-based video anomaly detection network"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the paper's reported FLOPs, parameter counts, and frames-per-second for all compared methods are measured under comparable conditions, since the efficiency advantage over CNN and transformer baselines rests entirely on those numbers.","fun_headline_variants_meta":{"raw":{"variants":["Mamba video anomaly detector: 40 FPS, 7.2M params","STNMamba: Mamba-based VAD with 98% AUC on Ped2","Lightweight Mamba detects video anomalies at 40 FPS","First Mamba-based video anomaly detection network"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1487,"prompt_tokens":1071,"completion_tokens":416,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":687,"completion_tokens_details":{"reasoning_tokens":339}},"tokens_in":687,"tokens_out":416,"duration_ms":4156,"temperature":1.0,"reasoning_tokens":339,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:34:41.365366+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure STNMamba and the top compared baselines (e.g., MNAD and ASTT) on the same GPU with the same input resolution (256×256) and the same number of input frames, and recompute FLOPs and parameters with a standard profiler; if STNMamba's numbers rise to or above the baselines' when measured uniformly, the central efficiency claim is falsified.","supporting_citations":[{"cited_title":"Memorizing normality to detect anomaly: Memory- augmented deep autoencoder for unsupervised anomaly detec tion,","cited_arxiv_id":null,"evidence_quote":"Introduces memory-augmented autoencoders for normality memorization, the conceptual basis of the memory banks used in STIM."},{"cited_title":"Learning memory-guided nor mality for anomaly detection,","cited_arxiv_id":null,"evidence_quote":"Memory-guided normality baseline that STNMamba is compared against for both frame-level AUC and efficiency (1/2 params, 1/40 FLOPs claimed)."},{"cited_title":"Future frame predicti on for anomaly detection - A new baseline,","cited_arxiv_id":null,"evidence_quote":"Establishes future-frame prediction as the VAD baseline and defines the PSNR-based anomaly scoring that STNMamba adopts."},{"cited_title":"Video anomaly detection with spatio-temporal dissociation,","cited_arxiv_id":null,"evidence_quote":"Motivates the use of RGB frame differences to capture short-term motion without optical flow, which the temporal encoder relies on."},{"cited_title":"Appearance-m otion memory consistency network for video anomaly detection,","cited_arxiv_id":null,"evidence_quote":"Appearance-motion dual-stream baseline with bottleneck fusion, the architecture STIM improves upon with multi-level interaction."}],"review_version":1}