PaperlayerView exact revision on arXiv

Source-linked AI summary

PARSEE-VAD: Efficient Training-Free Online Video Anomaly Detection via Proposition-Aware Reasoning and Streaming Evidence Escalation

Ji Wang, Shuangqing Zhang, Guo-Sen Xie, Fang Zhao

arXiv:2609.33236v2cs.CV

TL;DR

Online VAD with frozen multimodal models must read reliable current-window semantics under causal and computational constraints while maintaining temporal continuity. PARSEE-VAD uses proposition-based current evidence and compact score-state updates; it reports competitive training-free online performance across four benchmarks. Its conclusion is a compact causal interface between present evidence and what persists across decisions.

  • Problem

    Training-free online VAD must convert current visual semantics into anomaly decisions under no-future and limited-budget constraints, while temporal continuity raises the question of whether high-dimensional history must be retransmitted.

    Source: §1 · pp. 1–2

  • Method

    PAR reads signed proposition evidence from a shared visual-prefix state and conditionally acquires specialist queries, while SEE converts that evidence into bounded score-derived state across decisions.

    Source: §1 · pp. 2–3

  • Results

    PARSEE-VAD demonstrates competitive training-free online performance on UCF-Crime, XD-Violence, MSAD, and UBnormal.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Across three primary real-world benchmarks and an additional UBnormal transfer evaluation, the separation provides a compact causal interface between present evidence and evidence allowed to persist.

    Source: §5 · p. 10

    Experiments use established public VAD benchmarks; real-world deployment requires human oversight, privacy safeguards, access control, and evaluation for the deployment domain.

    Source: ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

Training-free online video anomaly detection (VAD) with frozen multimodal language models faces two coupled challenges: extracting reliable current-window semantics under causal and computational constraints, and maintaining temporal continuity without repeatedly transmitting high-dimensional history. Encoding history through text can compress visual evidence and introduce semantic bias, whereas retaining visual history expands multimodal context. We introduce PARSEE-VAD, a two-module framework that separates semantic evidence acquisition from score-state evolution. Proposition-Aware Reasoning (PAR) extracts structured propositional evidence from the current causal window and conditionally activates more specific queries when coarse evidence warrants further refinement. By sharing a reusable causal visual prefix across queries, PAR reduces redundant computation through selective execution. Streaming Evidence Escalation (SEE) maps the acquired proposition evidence into a compact score-domain event state through current evidence escalation, then propagates only the resulting bounded state across decisions to support temporal continuity. Experiments on four benchmarks demonstrate strong training-free online performance while selective routing reduces specialist computation and score-state propagation remains sparse. These results support a current-first principle for streaming multimodal inference: resolve present semantics first, then use compact historical state only to repair residual continuity gaps.

1 INTRODUCTION

Training-free online VAD must infer anomalies from current causal observations under limited inference budgets, while transient evidence gaps complicate decisions. PARSEE-VAD separates current semantic evidence acquisition from compact score-state evolution.

1 INTRODUCTION

  • Online VAD produces incremental predictions from incoming frames without future information, and labeled-anomaly training is costly to scale across diverse scenes.Frozen vision-language and multimodal models offer semantic priors without fitting a target-domain detector.Source: §1 · p. 1
  • Scalar anomaly rubrics can make decisions depend on how anomaly semantics are partitioned and described, creating an actionability–generality tension.Detailed rubrics add task-specific semantic commitments; looser rubrics leave score mapping less constrained.Source: §1 · p. 2
  • Textual history can compress visual evidence and make retained state depend on generated wording, while visual history expands multimodal context.The paper asks whether continuity requires retransmitting high-dimensional history.Source: §1 · p. 2
  • PAR exposes signed proposition evidence and conditionally acquires specialist evidence, using a shared visual-prefix state to make independent probes efficient.This reduces dependence on generated scalar anomaly scoring.Source: §1 · p. 3
  • SEE applies current proposition evidence to a coarse state, then carries only bounded score-derived state across decisions.This avoids repeatedly reintroducing high-dimensional textual or visual history.Source: §1 · p. 3

2 RELATED WORK

Related work spans multiple VAD training regimes, training-free semantic approaches, and online systems with different memory and decision strategies. PARSEE-VAD distinguishes current evidence acquisition from compact temporal state propagation.

2.1 VIDEO ANOMALY DETECTION

  • Classical VAD includes supervised, weakly supervised, one-class, and unsupervised regimes; CLIP- and prompt-based methods add semantic modeling.PARSEE-VAD keeps the multimodal language model frozen for training-free online inference.Source: §2.1 · p. 3

2.2 TRAINING-FREE AND MLLM-BASED VAD

  • Training-free methods mediate current evidence through generated language, definition-guided judgments, or semantic scoring axes, whereas PARSEE-VAD directly reads signed proposition evidence and selectively adds specialist inference.Related approaches include hierarchical, agentic, frozen-representation, and conditionally invoked reasoning methods.Source: §2.2 · p. 3

2.3 ONLINE VAD AND TEMPORAL MEMORY

  • Online VAD excludes future frames, but methods use different streaming memory, reasoning, and conditional-computation strategies.PAR exposes and selectively acquires evidence from the current causal observation; SEE updates bounded score state, and published accuracy numbers are not latency-normalized comparisons.Source: §2.3 · p. 3

3 METHOD

PARSEE-VAD uses current causal observations to acquire proposition evidence, then converts that evidence into a score state for decisions and temporal continuity. PAR shares visual computation across queries and routes specialist probes selectively; SEE applies current corrections and bounded state carry.

3 METHOD

  • PARSEE-VAD combines a causal observation and compact prior state to emit a decision score, assigning completed-interval scores retrospectively for frame-level evaluation.The stream uses a 60-frame (2-s) decision stride.Source: §3 · pp. 3–4

3.1 PROPOSITION-AWARE REASONING

  • PA-readout evaluates the frozen visual encoder once and reuses its visual-prefix state across proposition-specific text queries.Q2 tests overall abnormality and Q3 visible physical dynamics; conditional P3 and P4 probe forceful human interaction and consequential object or environmental interaction.Source: §3.1 · pp. 4–5
  • Each proposition uses forward and reversed A/B order, averaging sign-aligned final-token margins to reduce option-position bias.PA-routing acquires P3 for anomaly-positive dynamic windows and requires stronger Q2 evidence not positively explained by P3 for P4; it ends with acquired margins for SEE.Source: §3.1 · p. 5

3.2 STREAMING EVIDENCE ESCALATION

  • SEE begins with coarse Q2 evidence, forms a current-evidence state from PAR propositions, and carries only bounded score-derived state across decisions.Negative Q3 evidence suppresses Q2-positive interpretations without visible dynamics; routed specialist evidence can reinforce or oppose them, while clipping bounds the correction and a positive Q2 base is required.Source: §3.2 · p. 5
  • When the current correction is negative, a one-step carry retains at most enough preceding positive correction to cancel it, never raising the state above coarse Q2.The next carry stores the raw correction, while the historical branch separately maintains recent current-evidence states.Source: §3.2 · p. 5
  • Temporal escalation activates only when recent state and current coarse evidence indicate continuity while the current score is a relative dip.The gate requires positive recent state and persistent coarse Q2 support; history stores the current-evidence state rather than the final score, preventing recursive escalation.Source: §3.2 · p. 6

4 EXPERIMENTS

Across four benchmarks, PARSEE-VAD reports strong accuracy under frozen-model, causal online constraints, while its analyses show selective proposition acquisition and sparse temporal correction. Its experiments also evaluate runtime, visual-budget trade-offs, and stage-wise and mechanistic behavior.

4.1 EXPERIMENTAL SETUP

  • Experiments use a single PAR/SEE configuration across four benchmarks, with nine-frame causal observations, frozen Qwen3.5-9B, and a 512sq visual budget by default.UCF-Crime, XD-Violence, and UBnormal use 60-frame decision cadence; some MSAD videos use shorter regular strides.Source: §4.1 · p. 6

4.2 COMPARISON WITH PRIOR METHODS

  • PARSEE-VAD improves over several online baselines on UCF-Crime and XD-Violence, reaches 90.55 AUC / 82.02 AP on MSAD, and obtains 70.18 AUC on UBnormal without a universal best score.The comparisons use each method’s reported causal protocol, not a latency-normalized systems ranking.Source: §4.2 · p. 6, Table 1 · p. 7
  • The canonical system processes a 2-s decision interval in 1.372 s on average (C/D = 0.686), below its own cadence but not the lowest published processing load.Hardware and timing protocols differ across methods.Source: §4.2 · pp. 6–7, Table 2 · p. 7

4.3 ABLATION AND ANALYSIS

  • At 512sq, most observed MSAD AUROC gain arrives before runtime rises sharply at the largest visual-token budget; shared-prefix reuse cuts repeated computation about threefold while preserving nearly all routing decisions.Reuse amortizes branch cost, but not construction of the shared prefix.Source: §4.3 · pp. 7–8, Figure 3 · p. 8
  • Replacing generated 0–100 scoring with Q2 raises UCF-Crime AUC from 80.115 to 84.872 and MSAD AUC/AP from 88.211/78.071 to 90.275/79.073; adding Q3 then raises MSAD AP from 79.073 to 79.978 with AUC essentially unchanged.Q3 selectively suppresses Q2-positive windows that lack visible dynamics.Source: §4.3 · p. 8, Table 3 · p. 8
  • On MSAD, PA-routing reaches 90.518/81.970 AUC/AP with 0.565 specialist queries per window, versus 90.245/81.155 with two non-routed specialist queries.Routing lowers semantic continuation cost relative to all-probe acquisition while AP increases.Source: §4.3 · pp. 8–9
  • Temporal maintenance activates on only 1.1–2.9% of decisions, versus 23.8–43.8% for current-evidence updates; its paired bootstrap is positive on UCF-Crime and XD-Violence but overlaps zero on MSAD.The temporal stage is a targeted continuity correction rather than a primary source of aggregate performance gain.Source: §4.3 · p. 9

4.4 MECHANISTIC ANALYSIS

  • Q3 suppression and routed specialist corrections are more frequent and locally aligned with benchmark labels than temporal persistence, but negative Q3 margins can both suppress scores and prevent specialist acquisition.This can penalize persistent low-dynamics hazards twice, while subtle interactions may remain unresolved when the frozen backbone exposes insufficient current evidence.Source: §4.4 · p. 9, Figure 4 · p. 9

5 CONCLUSION

PARSEE-VAD separates current-window semantic evidence acquisition from temporal score-state evolution, using routed proposition queries and compact state to support training-free online VAD. Its prompts distinguish coarse anomaly and physical-change evidence from specialist interaction evidence, while SEE escalates current evidence and maintains continuity across decisions.

5 CONCLUSION

  • PARSEE-VAD separates current-window semantic evidence acquisition from score-state evolution, carrying bounded score-derived state across causal decisions.The conclusion describes a compact causal interface between present evidence and what persists.Source: §5 · p. 10

B.1 PA-READOUT

  • The proposition prompts evaluate forward and reversed answer orders to reduce sensitivity to option position.PAR reads signed margins from a coarse-to-specific proposition hierarchy, with Q2 and Q3 always read and specialist P3/P4 logits acquired conditionally.Source: A FULL PROPOSITION PROMPTS · p. 12, §B.1 · p. 15

A.1 Q2: OVERALL ANOMALY EVIDENCE

  • The Q2 prompt classifies the current event using only visual evidence available up to the current time.Q3 instead asks whether ordered frames directly demonstrate clear physical change over time.Source: §A.2 · p. 13

A.3 P3: FORCEFUL HUMAN INTERACTION

  • P3 targets directly visible forceful bodily interaction, while P4 targets concrete physical events involving objects, property, vehicles, or the environment.Both prompts exclude inferences based only on ordinary movement or gestures.Source: §A.3 · pp. 13–14, §A.4 · p. 14

A.5 GENERATED-SCORE COMPARATOR

  • On UCF-Crime, raw generated scalar scoring before historical workflow propagation yields 80.115 AUC; on MSAD, it yields 88.211 AUC and 78.071 AP, using the same completed-interval frame mapping as Q2.Source: §A.5 · p. 15

B.2 PA-ROUTING

  • PA-routing decides whether specialist P3/P4 logits are acquired and ends with the available proposition margins, without using temporal state.The P4 branch requires stronger coarse Q2 support than the zero-threshold gates.Source: §B.2 · p. 16

B.3 SEE: CURRENT EVIDENCE ESCALATION

  • SEE maps routed signed evidence into current-evidence state Lt, applying a bounded proposition correction to the coarse Q2 evidence.The operator is memoryless at this stage; PAR determines which specialist margins are available.Source: §B.3 · p. 16

B.4 SEE: RECENT CORRECTION CARRY

  • SEE carries the previous raw current-evidence correction and uses two previous event states with coarse Q2 margins to assess event continuity.The history stores the raw current correction rather than an inherited rescue, preventing an intervention from recursively creating evidence for a later intervention.Source: §B.4 · p. 16, §B.5 · pp. 16–17

C ADDITIONAL IMPLEMENTATION DETAILS

Shared-prefix reuse branches proposition-specific text prefills from one reusable multimodal state for the same current observation. The measured approximation reduces synchronized model time by about threefold while preserving nearly all routing paths.

C.1 SHARED PREFIX STATE

  • A reusable visual-prefix state lets Q2, Q3, P3, and P4 branch from the same current observation without inheriting another query’s text.In the hybrid Qwen3.5 decoder, the reused state includes standard K/V states and Gated-DeltaNet convolutional/recurrent state.Source: §C.1 · p. 17
  • Shared-prefix reuse is an implementation approximation, and exact equivalence to an independent full-prompt forward is not guaranteed.The authors measure proposition-logit and routing deviation rather than attributing drift to an unverified positional mechanism.Source: §C.1 · p. 17
  • Native reuse reduces synchronized model time by about threefold while preserving almost all routing paths.On matched traces, the approximation mainly perturbs absolute score geometry rather than the direction of temporal updates.Source: §C.1 · p. 17

C.2 HARDWARE AND RUNTIME MEASUREMENT

Runtime measurements distinguish synchronized model-section timing from end-to-end system timing, which includes host-side work and may contain contention. The reported dense-to-routed timing reductions are separate from the reduction in specialist-query count.

C.2 HARDWARE AND RUNTIME MEASUREMENT

  • Each variant processes a full 3,607-window MSAD pass, with CUDA-synchronized measurements around visual-prefix and proposition-tail model sections.End-to-end timing also includes frame decoding, resizing, shared prefill, proposition continuations, processing, and orchestration.Source: §C.2 · p. 17
  • End-to-end timing is a system measurement, not a hardware-invariant comparison, because CPU, NUMA, and storage affinity were not isolated.The synchronized model-section timing is identified as the cleaner algorithmic compute comparison; Table 4 uses a separate controlled harness.Source: §C.2 · pp. 17–18, Table 4 · p. 18, Table 5 · p. 18
  • Dense-to-routed execution reduces end-to-end mean time by 13.2% and synchronized model-section mean time by 17.6%.These timing differences are reported separately from the 71.7% reduction in specialist-query count.Source: §C.2 · p. 17

C.3 LOW-LATENCY OPERATING POINT

A four-frame, 30-frame-stride operating point reduces processing time while changing observation budget and cadence together. The reported canonical stream also has complete decision coverage, fixed inference parameters across datasets, and gradual performance variation across nearby settings.

C.3 LOW-LATENCY OPERATING POINT

  • The 4F/30F operating point changes temporal span and decision cadence together, so it is not an isolated context-length or overlap ablation.It retains the same model, preprocessing budget, PA-routing, SEE, and inference-rule constants as the main system.Source: §C.3 · pp. 18–19, Table 6 · p. 19
  • On MSAD, the 4F/30F setting cuts mean/P95 time from 1.372/1.741 s to 0.753/0.931 s, with AUC/AP changing from 90.551/82.024 to 90.423/83.059.On UCF-Crime it reaches 84.567 AUC at 0.485/0.658 s mean/P95, versus 85.146 AUC for canonical settings.Source: §C.3 · pp. 18–19, Table 6 · p. 19

C.4 FRAME SAMPLING, SCORE AVAILABILITY, AND INTEGRITY

  • The canonical 512sq runs contain all expected decisions with no duplicate video–anchor pairs, failure statuses, or decision-index gaps.Independent reconstruction found zero label-anchor mismatches for UCF-Crime, XD-Violence, and MSAD; SEE’s H=2 gate waits for two preceding states.Source: §C.4 · p. 19

C.5 FIXED INFERENCE PARAMETERS

  • All reported evaluations use the same scalar configuration across datasets, including α = 0.75, ρc = 0.5, and horizon H = 2.The parameters encode bounded current correction and short-horizon continuity over preceding decisions.Source: §C.5 · p. 19

C.6 CURRENT ESCALATION AND ROUTING SENSITIVITY

  • Nearby escalation strengths produce a broad performance plateau rather than a sharp optimum at the deployed α = 0.75.α = 0.50 gives slightly higher MSAD AUC, while α = 0.625 gives slightly higher AP; measured finite P4 thresholds also show a broad operating region.Source: §C.6 · p. 19

D.1 BACKBONE ROBUSTNESS

Backbone robustness varies by proposition pathway: all three backbones improve AP, but their stage contributions differ, while SEE temporal maintenance remains substantially sparser than current-evidence correction.

D.1 BACKBONE ROBUSTNESS

  • All three backbones improve AP through different pathways: VideoLLaMA3-7B gains mostly from Q3, while Qwen3.5-9B benefits most from specialist-driven escalation.The final AP improvement does not mean each proposition contributes with the same sign or magnitude.Source: §D.1 · p. 20
  • Qwen3.5-2B experiences aggressive Q3 suppression, and its AUC drop indicates that fixed signed-margin thresholds are not automatically calibrated across model capacities.The cross-backbone diagnostic is limited to MSAD; backbone-specific calibration or quantile/temperature normalization is suggested as an extension.Source: §D.1 · pp. 19–20

D.2 INTERVENTION DENSITY

  • Current Q3/P3/P4 corrections are comparatively frequent, whereas historical maintenance activates only under recent-correction or continuity-valley gates.The final valley gate is the rarest temporal intervention.Source: §D.2 · p. 20

E ADDITIONAL EVALUATION ANALYSES

The additional analyses examine whether routing allocation is semantic rather than merely computational and document the integrity of the evaluated decision streams.

E ADDITIONAL EVALUATION ANALYSES

  • The routing controls separate the cost of computing specialist evidence from the effect of admitting it into the score state, then compare allocation policies under a matched specialist budget.PAR allocates by the Q2/Q3/P3 proposition structure, unlike random, periodic, magnitude, or generic-uncertainty policies; Table 11 tests the allocation principle on the archived UCF-Crime dense-probe bank.Source: E ADDITIONAL EVALUATION ANALYSES · p. 20
  • The routing audit tracks how often current Q3/P3/P4 evidence changes the coarse Q2 state and distinguishes those updates from historical maintenance and valley-gate activations.All observed valley-gate activations are positive-score recoveries.Source: E ADDITIONAL EVALUATION ANALYSES · p. 21, Table 10 · p. 21, Table 9 · p. 21
  • The three primary 512sq evaluations contain 18,442 UCF-Crime, 39,418 XD-Violence, and 3,607 MSAD causal decisions, with no reported duplicate pairs, failure statuses, decision-index gaps, or anchor-label mismatches.The UBnormal transfer run contains 1,649 decisions over all 211 official test videos.Source: E ADDITIONAL EVALUATION ANALYSES · p. 21

E.1 UBNORMAL COMPARISON PROVENANCE

The additional evaluation reports a frame-level UBnormal result with explicit comparison provenance and analyzes score availability, score-state increments, proposition readout, specialist semantics, and temporal maintenance.

E.1 UBNORMAL COMPARISON PROVENANCE

  • UBnormal comparison values are drawn from identified prior reports rather than treated uniformly as reruns; for example, VADTree 65.80 is a secondary/reproduced value, while SphereVAD (online) 71.62 reflects its online deployment mode.The passage identifies the comparison-table sources and distinguishes the fixed-constant URF-HVAA value 68.98 from its adaptive value 69.02.Source: §E.1 · p. 21
  • 70.18% frame-level micro AUROC and 0.759735912 AP are reported for PARSEE-VAD on UBnormal under completed-interval mapping.The evaluation uses a dedicated 512sq run on the official 211-video test split with the same frozen Qwen3.5-9B and operator as the main experiments; entries with unverified source or protocol are left unfilled.Source: §E.1 · pp. 21–22
  • The Q2-to-full score changes are already visible at causal decision anchors, and current proposition escalation dominates the AP change while temporal updating is intentionally sparse.Paired video-level resampling finds the additional temporal increment smaller and overlapping zero on MSAD; wide intervals indicate that several small AUC changes are dominated by inter-video variation.Source: §E.1 · p. 22, Table 14 · p. 23, Table 13 · p. 23
  • Forward and reversed A/B margins remain strongly correlated but differ in sign on a non-negligible subset, motivating sign-aligned averaging rather than reliance on option position.The option-order analysis concerns the Q2 proposition readout.Source: §E.1 · p. 22, Table 15 · p. 23
  • Signed specialist evidence allows support or rejection of a semantic hypothesis, but the deployed signed rule can offset positive environmental evidence after a non-positive P3 result.The reported trade-off favors signed suppression on UCF-Crime decision anchors, while positive-only scoring is slightly better on XD-Violence and MSAD.Source: §E.1 · pp. 22–23, Table 16 · p. 23
  • Generic smoothing can exploit persistence on some datasets but can also broadly reorder scores; SEE instead applies temporal maintenance only when short-horizon conditions are met.Nearby H, τ, and η values leave the valley gate stable, while ρc produces only small decision-anchor changes without changing PAR acquisition or current-window escalation.Source: §E.1 · pp. 23–24, Table 17 · p. 24
  • The current hierarchy is better matched to active, interactive anomalies than to all persistent surveillance anomalies, because weak Q3 dynamics can suppress local scores and close specialist routing.The passage reports high negative-Q3 rates for persistent MSAD Fire and UCF-Crime Arson, but low rates for visibly dynamic Fighting.Source: §E.1 · p. 24

F LIMITATIONS

PARSEE-VAD inherits frozen-backbone perceptual limits and also has design and state constraints that bound the anomaly types and evidence gaps it can handle.

F LIMITATIONS

  • The frozen backbone's perceptual limitations coexist with an inductive bias in the proposition hierarchy, which uses Q3 active physical change to gate both specialists.Persistent hazards may therefore be suppressed even when Q2 remains positive; the hierarchy is described as better matched to active/interactive anomalies than to every static or diffuse condition.Source: F LIMITATIONS · p. 24
  • Errors in proposition geometry can propagate to Lt, while short-horizon temporal memory can maintain only evidence established by earlier windows.Long evidence gaps, severe occlusions that erase coarse Q2 evidence, and anomalies absent from current and stored scalar evidence remain unresolved.Source: F LIMITATIONS · pp. 24–25
  • PARSEE-VAD operates at decision anchors and has bounded decision delay; its measured runtime reflects a specific RTX 5880 Ada setup, and resource-limited deployment needs separate throughput, memory, and energy characterization.End-to-end timing is system-specific because host CPU/storage affinity was not isolated.Source: F LIMITATIONS · p. 25

Related papers

Loading 2609.33236v2…