{"id":"541037fd-18a5-4209-bb0e-9892ae8a7361","arxiv_id":"2411.11003","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TeG fuses three temporal scales of Video Swin Transformer features with cross/self-attention, reporting 87.16% AUC on UCF-Crime and 84.57% AP on XD-Violence.","lead":"This paper introduces TeG, a surveillance video model that combines short, medium, and long video features through attention to spot abnormal events. It reports competitive scores on public crime and violence benchmarks and a small real-city field test.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No ablation isolates temporal-granularity fusion; gains could come from the VST backbone or inherited RTFM loss.","rationale":"The reader's weakest assumption is exactly that the reported gains are not ablated against a single-scale VST backbone or an attention-free fusion baseline. This is the single most load-bearing concern because the paper's main novelty is the attention-based temporal-granularity fusion; if the performance is driven by the pretrained VST features or the inherited RTFM loss, the central methodological contribution is unsupported. The paper's comparison to prior methods is confounded by different backbones and losses. The field-lab evaluation is too small and lacks true-negative samples to independently validate the claim of detecting unseen anomalies. However, the lack of ablations is an addressable omission rather than a demonstrated error, so the CONDITIONAL verdict remains appropriate. The concrete test I propose would directly isolate the effect of the attention fusion and settle the concern.","tokens_in":8008,"tokens_out":4487,"duration_ms":100054,"concrete_test":"Run the identical training setup (VST features, RTFM loss, same splits and hyperparameters) with four configurations: (a) single granularity G=32 only, no attention; (b) concatenation of FS, FM, FL without attention, followed by the same 3-layer FCN; (c) TeG without MCA (MSA only); (d) TeG as described. If (a) or (b) achieves within ~1 AUC point of TeG on UCF-Crime, the temporal-granularity attention mechanism is not the driver of the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that fusing short/medium/long temporal-granularity features via MCA/MSA improves anomaly detection. Table I compares TeG against prior works, but none share TeG's exact backbone (VST pretrained on Kinetics-400) or loss (RTFM's robust feature magnitude and MIL ranking loss). TeG uses the same loss as RTFM and reports 87.16 vs 84.30 AUC on UCF-Crime and 84.57 vs 77.81 AP on XD-Violence; however, without a single-scale VST baseline trained with the same loss, the improvement cannot be attributed to the attention-based granularity fusion. Likewise, there is no comparison against simply concatenating FS, FM, FL without attention, or against removing the residual connection. Without these ablations, the load-bearing assumption that temporal-granularity attention is the active ingredient remains untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TeG, a weakly-supervised video anomaly detection method that extracts spatio-temporal features from a Video Swin Transformer backbone at three temporal granularities (8, 32, and 64 frames), fuses them with multi-head cross-attention and multi-head self-attention blocks plus a residual connection, and trains with the RTFM feature-magnitude and MIL ranking losses. The method is evaluated on UCF-Crime, ShanghaiTech, and XD-Violence, reporting 87.16% AUC, 95.32% AUC, and 84.57% AP, respectively. The paper also extends UCF-Crime with three Smart-City anomaly classes and reports a field-lab experiment on 109 staged samples, including a real-time deployment with anomaly alerts to control-room operators.","tokens_in":8237,"tokens_out":3513,"duration_ms":41362,"significance":"If the reported results are reproducible, TeG is competitive on two of the three standard benchmarks, and the deployment component is a useful industrial validation. The dataset extension is a modest but useful contribution, and the execution-time analysis is a strength. However, the central methodological claim — that multi-granularity temporal fusion via attention is the active ingredient behind the reported gains — is not supported by any controlled comparison in the current manuscript. The field-lab evidence is also too small and coarsely aggregated to substantiate generalization to unseen anomaly classes.","major_comments":[{"comment":"The comparison in Table I contrasts TeG with prior methods that differ simultaneously in backbone, fusion architecture, and loss function. Because TeG uses the same RTFM loss, a minimal control is RTFM with the same VST backbone and single-scale features, or TeG with the attention fusion removed. Without such ablations, the observed differences (87.16 vs 84.30 AUC on UCF-Crime, 84.57 vs 77.81 AP on XD-Violence) cannot be attributed to temporal-granularity fusion rather than to the VST backbone or the inherited loss. This directly affects the paper's central claim and should be addressed with controlled experiments.","section":"Section IV.B, Table I"},{"comment":"The field-lab validation uses only 109 staged samples from one hour of recording, with no confidence intervals and no per-class breakdown. The 'unseen' category pools 66 improper-zone, 7 unlawful-stop, and 18 improper-turn samples into a single binary decision, and the reported 79.10% accuracy and 0.88 F1 are suggestive but not statistically robust evidence that the model generalizes to unseen classes. Please report per-class performance, confidence intervals, and ideally frame-level temporal localization for the field-lab data.","section":"Section IV.C, Table III"},{"comment":"The paper states that TeG demonstrates 'better capability for identifying complex anomalies' and outperforms RTFM, but on ShanghaiTech TeG's 95.32% AUC is lower than RTFM's 97.21% and also lower than several other cited methods. The attribution of this gap to 'simpler anomalies' in ShanghaiTech is a post-hoc explanation without supporting analysis. The authors should either temper the superiority claim or provide an analysis, such as class-wise AUC, that explains the trade-off.","section":"Section IV.B, ShanghaiTech results"}],"minor_comments":[{"comment":"The notation in Eq. (1) is unclear: 'SL_i = P c_j' uses an undefined operator P, and the outer union over 32 segments appears to mix set notation with concatenation. Please rewrite the equation with explicit concatenation and indexing.","section":"Section III, Eq. (1)"},{"comment":"There are several typos: 'real-word surveillance' should be 'real-world surveillance', 'future fusion' should be 'feature fusion', and Section IV.C contains 'and and'.","section":"Section I and Section IV.D"},{"comment":"The temporal granularity values (8, 32, 64), number of attention heads (4), and loss hyperparameters (m=100, k=3) are stated as empirically set, but no sensitivity analysis or ablation is provided. At minimum, please justify these choices or provide a short study.","section":"Section IV.A and III"},{"comment":"The reference formatting is inconsistent: entries [12]–[14] and [15]–[16] use a different citation style from the rest of the bibliography. Please unify the formatting.","section":"References"},{"comment":"It is not stated explicitly whether the Table I results use the original UCF-Crime dataset or the extended 17-class version. Since Section IV.C states that the model is trained on the extended dataset, the protocol for Table I should be clarified to avoid ambiguity.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the missing ablation isolating the temporal-granularity attention fusion. If the authors add single-scale and attention-free baselines with the same backbone and loss, and the reported gains persist, the paper would be much stronger. The field-lab sample size is also small, but that component could be reframed as a demonstration rather than a statistical evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: TeG is a typical weakly-supervised video anomaly detector that gets competitive scores on UCF-Crime and XD-Violence, extends UCF-Crime with four new classes, and reports a small but real field-lab deployment. The paper is worth a serious referee, but the central methodological claim—that fusing short/medium/long temporal-granularity features via MCA/MSA is what drives the gains—is not actually tested, because there is no ablation against a single-scale VST backbone or an attention-free fusion.\n\nWhat's genuinely useful: the 17-class extension of UCF-Crime is a practical contribution for surveillance projects; the field-lab validation with seen and unseen anomalies, reporting 79-83% accuracy on 109 samples, is a nice reality check; and the execution-time numbers (317 ms per 10-sec segment) show the system could run near real time on a 2080Ti. The benchmark numbers are plausible and the paper is clearly written.\n\nThe soft spots are real and structural. The loss is exactly RTFM's, the backbone is VST, and the paper never shows what happens if you remove the multi-granularity attention or just concatenate features. So we genuinely don't know whether the improvement over RTFM comes from the backbone, the loss, or the proposed fusion. ShanghaiTech is also below several cited methods (95.32 vs RTFM's 97.21), which the authors wave away as 'simpler anomalies'—that needs defense, not a hand-wave. The field test has only 109 samples and reports binary sample-level accuracy/F1, not frame-level AUC, so it's suggestive, not conclusive. Code and data are not released, which makes independent verification harder.\n\nNone of this is fatal; these are standard weaknesses of many applied papers in this area. But the paper overclaims in the abstract when it says the model 'effectively captures both short and prolonged anomalies'—the experiments don't isolate that capability. A serious revision should add ablations, error bars across runs, and ideally release code and the extended dataset annotations.\n\nWho's this for? Researchers working on weakly-supervised video anomaly detection, and practitioners in smart-city surveillance who want a field-tested baseline. I'd give it a fair peer review, but I'd push for the ablations before accepting. I don't think I'd cite it in my own work until the attribution question is resolved.","headline":"Competitive numbers and a real field deployment, but the core temporal-granularity claim is untested because there are no ablations.","tokens_in":8713,"tokens_out":1757,"would_cite":false,"duration_ms":19086,"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":"Fusing short-, medium-, and long-term video features with attention achieves 87.16% AUC on UCF-Crime and detects unseen anomalies in a city field lab.","keywords":["video anomaly detection","temporal granularity","multi-head cross-attention","multi-head self-attention","weakly supervised learning","smart city surveillance","UCF-Crime","real-time deployment"],"falsifier":"Train the identical Video Swin Transformer and loss on UCF-Crime using only the 64-frame (long) granularity, with no cross-scale attention, and measure AUC. If the score matches or exceeds 87.16%, the temporal-granularity and attention-fusion mechanism is not what carries the result.","tokens_in":7847,"feed_emoji":"🚨","tokens_out":6272,"duration_ms":59512,"temperature":0.7,"pith_summary":"The paper argues that surveillance anomalies have very different durations, and a detector that looks at only one time scale will miss either abrupt actions or slow-developing events. It introduces TeG, which splits each video segment into chunks of 8, 32, and 64 frames, extracts features for each scale with a Video Swin Transformer, and fuses the three feature sets with multi-head cross-attention and self-attention before classification. On public benchmarks the model reports 87.16% area under the curve on UCF-Crime, 95.32% on ShanghaiTech, and 84.57% average precision on XD-Violence. The paper also extends UCF-Crime with additional smart-city anomaly types and validates the system in a field lab, where it flags both trained (seen) and untrained (unseen) anomaly classes at roughly 80% accuracy. The point is that multi-scale temporal fusion can make weakly supervised anomaly detection work for the full range of real-world event durations.","feed_headline":"Multi-scale timing lifts anomaly detection to 87% AUC on UCF-Crime","feed_subtitle":"TeG fuses 8-, 32-, and 64-frame video chunks and flags both trained and never-seen incidents in live field tests.","key_machinery":"The central mechanism is the temporal-granularity feature fusion built from multi-head cross-attention (MCA) and multi-head self-attention (MSA). MCA lets feature sets from different time scales attend to each other, forming pairwise fused matrices $F_{SM}$, $F_{ML}$, and $F_{SL}$; MSA re-weights each scale by global correlations across all 32 video segments, forming $F_{SML}$. The four matrices are concatenated with a residual connection, so the classifier sees both cross-scale correlations and whole-clip temporal context. The same machinery is what the paper credits for detecting anomalies of different durations.","core_discovery":"TeG's central claim is that explicitly modeling temporal granularity improves weakly supervised video anomaly detection. Each video is divided into 32 segments, and each segment is re-chunked at three granularities, 8, 32, and 64 frames, to capture short, medium, and long dynamics. The Video Swin Transformer produces a feature vector per chunk, averaged per segment into three representations $F_S$, $F_M$, and $F_L$. Multi-head cross-attention correlates pairs of scales and self-attention models dependencies across segments, producing four matrices that are concatenated with a residual connection and fed to a three-layer classifier trained with the top-k MIL ranking loss with feature-magnitude learning. The authors report that this design reaches 87.16% AUC on UCF-Crime, 95.32% AUC on ShanghaiTech, and 84.57% AP on XD-Violence, and that in a field lab with five cameras it detects both the anomaly classes it trained on and three untrained ones, with 79.81% accuracy and an F1 score of 0.88 over all 109 recorded anomalies.","pith_inferences":["The paper does not isolate the contribution of the attention fusion; a reader could train the same backbone and loss at a single granularity and compare. If a single scale matches the reported numbers, then the multi-scale attention design is not the decisive factor.","Granularity settings of 8, 32, and 64 frames are fixed empirically; sweeping these values would show how sensitive the result is and whether adaptive scales could improve it.","Extending the field-lab protocol into a standardized open-set benchmark would let other systems measure how well they generalize to anomaly classes absent from training.","Inspecting the attention weights could test the paper's underlying duration hypothesis directly: the model should concentrate on short scales for fast events and long scales for slow ones."],"forward_implications":["A single model can flag both brief anomalies like thrown objects or accidents and longer ones like fights, because it pools evidence from 8-frame, 32-frame, and 64-frame chunks.","On the UCF-Crime and XD-Violence benchmarks, the multi-scale approach exceeds the reported scores of the same-loss baseline, suggesting the fusion adds detection power for complex real-world scenes.","A weakly supervised system trained only on video-level labels can generalize to anomaly classes it never saw, as the field-lab results on unseen classes show.","The system can run in near real time: processing one 10-second segment takes about 1.5 seconds on a GTX-2080Ti, so control-room operators can receive alerts with type, time, camera, and a GIF of the event.","Adding new anomaly classes to UCF-Crime (from 13 to 17) supports evaluation of smart-city scenarios such as dangerous throwing, littering, and traffic accidents."],"supporting_citations":[{"why":"Supplies the Video Swin Transformer backbone that extracts spatio-temporal features at each temporal granularity.","marker":"[34]"},{"why":"Provides the UCF-Crime weakly labeled dataset and the 13 original anomaly classes that the paper extends.","marker":"[17]"},{"why":"Supplies the top-k MIL ranking loss and feature-magnitude training objective used to train TeG.","marker":"[18]"},{"why":"Provides the multi-head cross-attention principle used to correlate features across temporal scales.","marker":"[25]"},{"why":"Provides the self-attention mechanism used to model dependencies across video segments.","marker":"[35]"}],"fun_headline_variants":["TeG: Multi-scale temporal attention for smart-city anomaly detection","Temporal granularity method TeG reaches 87% AUC on UCF-Crime","Multi-scale video chunks help TeG spot anomalies in live city feeds","Cross-attention on 8/32/64-frame chunks lifts anomaly AUC to 87%","TeG's granular time scales track anomalies from train to field"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results are credited to multi-scale temporal fusion, but the paper does not compare against the same backbone and loss without that fusion, so the key assumption is that the granularity design, not the pretrained video model or the training loss, is what produces the reported accuracy.","fun_headline_variants_meta":{"raw":{"variants":["TeG: Multi-scale temporal attention for smart-city anomaly detection","Temporal granularity method TeG reaches 87% AUC on UCF-Crime","Multi-scale video chunks help TeG spot anomalies in live city feeds","Cross-attention on 8/32/64-frame chunks lifts anomaly AUC to 87%","TeG's granular time scales track anomalies from train to field"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1450,"prompt_tokens":903,"completion_tokens":547,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":448}},"tokens_in":519,"tokens_out":547,"duration_ms":6496,"temperature":1.0,"reasoning_tokens":448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:01:05.310667+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical Video Swin Transformer and loss on UCF-Crime using only the 64-frame (long) granularity, with no cross-scale attention, and measure AUC. If the score matches or exceeds 87.16%, the temporal-granularity and attention-fusion mechanism is not what carries the result.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF international conference on computer vision","cited_arxiv_id":null,"evidence_quote":"Supplies the top-k MIL ranking loss and feature-magnitude training objective used to train TeG."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the multi-head cross-attention principle used to correlate features across temporal scales."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"Provides the self-attention mechanism used to model dependencies across video segments."}],"review_version":1}