Benchmarking Compact VLMs for Clip-Level Surveillance Anomaly Detection Under Weak Supervision

Kirill Borodin; Kirill Kondrashov; Nikita Vasiliev; Ksenia Gladkova; Inna Larina; Mikhail Gorodnichev; Grach Mkrtchian

cs.CVMar 3, 2026

Benchmarking Compact VLMs for Clip-Level Surveillance Anomaly Detection Under Weak Supervision

Authors: Kirill Borodin, Kirill Kondrashov, Nikita Vasiliev, Ksenia Gladkova, Inna Larina, Mikhail Gorodnichev, Grach Mkrtchian

Organizations: Faculty of Information Technology, Moscow Technical University of Communication and Informatics, Moscow 111024, Russia

Abstract

CCTV safety monitoring demands anomaly detectors combine reliable clip-level accuracy with predictable per-clip latency despite weak supervision. This work investigates compact vision-language models (VLMs) as practical detectors for this regime. A unified evaluation protocol standardizes preprocessing, prompting, dataset splits, metrics, and runtime settings to compare parameter-efficiently adapted compact VLMs against training-free VLM pipelines and weakly supervised baselines. Evaluation spans accuracy, precision, recall, F1, ROC-AUC, and average per-clip latency to jointly quantify detection quality and efficiency. With parameter-efficient adaptation, compact VLMs achieve performance on par with, and in several cases exceeding, established approaches while retaining competitive per-clip latency. Adaptation further reduces prompt sensitivity, producing more consistent behavior across prompt regimes under the shared protocol. These results show that parameter-efficient fine-tuning enables compact VLMs to serve as dependable clip-level anomaly detectors, yielding a favorable accuracy-efficiency trade-off within a transparent and consistent experimental setup.

Figures & tables

Figure 1: Overview of the proposed multimodal inference pipeline for video anomaly detection, integrating LoRA-fine-tuned VLMs with prompting strategies for enhanced accuracy and efficiency.

Model	Prompt	Accuracy	Precision	Recall	F1	ROC-AUC	Latency
gemma-3-4B-it Gemma Team (2025)	Basic	0.7619	0.7105	0.9643	0.8182	0.7366	9.1008 ± 0.3205
	Zero Shot	0.6825	0.6402	0.9786	0.7740	0.6455	7.5379 ± 0.9962
	Few Shot	0.8056	0.8473	0.7929	0.8192	0.8071	8.1017 ± 0.3094
	CoT	0.7061	0.8211	0.5865	0.6842	0.7173	15.1717 ± 0.3021
InternVL3-2B Zhu et al. (2025)	Basic prompt	0.5992	0.8421	0.3429	0.4873	0.6312	2.6972 ± 0.0452
	Zero Shot	0.5020	1.0000	0.1071	0.1935	0.5536	2.7692 ± 0.0471

Table 1: Tiny models comparison with different prompt types on UCF-crime dataset.

Figure 2: Confusion matrix for tiny models with different prompt types on UCF-crime dataset.

Model	Prompt	Accuracy	Precision	Recall	F1	ROC-AUC	Latency
InternVL3-8B Zhu et al. (2025)	Basic	0.8532	0.9558	0.7714	0.8538	0.863	4.8936 $\pm$ 0.1034
	Zero Shot	0.7627	0.9595	0.5726	0.7172	0.7729	5.0435 ± 0.0978
	Few Shot	0.6270	1.0000	0.3286	0.4946	0.6643	4.4191 ± 0.1228
	CoT	0.6360	0.9123	0.3768	0.5333	0.6661	15.9177 ± 0.6653
Video-LLaVA-7B-hf Lin et al. (2024)	Basic	0.5794	0.6232	0.6143	0.6187	0.5750	2.5132 ± 0.1236
	Zero Shot	0.1315	0.2238	0.2302	0.2270	0.1196	3.3413 ± 0.2293

Table 2: Small models comparison with different prompt types on UCF-crime dataset.

Figure 3: Confusion matrix for small models with different prompt types on UCF-crime dataset.

Figure 4: Confusion matrix for tiny models finetuned with LoRA with different prompt types on UCF-crime dataset.

Model	Prompt	Accuracy	Precision	Recall	F1	ROC-AUC	Latency
gemma-3-4b-it Gemma Team (2025)	Basic	0.90 (+17.65%)	0.87 (+23.04%)	0.95 ( $-$ 1.52%)	0.91 (+11.26%)	0.89 (+20.83%)	8.10 $\pm$ 1.04 ( $-$ 12.35%)
	Zero Shot	0.89 (+30.75%)	0.87 ( +36.41% )	0.94 ( $-$ 3.70%)	0.91 (+17.13%)	0.89 (+37.32%)	8.10 $\pm$ 1.01 (+6.93%)
	Few Shot	0.91 (+13.26%)	0.90 (+5.98%)	0.95 (+19.76%)	0.92 (+12.68%)	0.91 (+12.49%)	8.37 $\pm$ 0.91 (+3.23%)
	CoT	0.90 (+26.95%)	0.87 (+6.47%)	0.95 (+61.91%)	0.91 (+33.05%)	0.89 (+24.08%)	8.77 $\pm$ 1.00 ( $-$ 73.58%)
InternVL3-2B Zhu et al. (2025)	Basic prompt	0.89 (+32.59%)	0.89 (+5.85%)	0.91 (+62.2%)	0.90 (+45.56%)	0.89 (+28.81%)	2.91 ± 0.08 (+7.45%)
	Zero Shot	0.91 (+45%)	0.98 ( $-$ 1.67%)	0.86 (+87.49%)	0.92 (+78.88%)	0.92 (+39.8%)	2.77 ± 0.04 (+0.12%)

Table 3: Tiny models finetuned with LoRA comparison with different prompt types on UCF-crime dataset.

Model	Accuracy	Precision	Recall	F1	ROC-AUC	Latency
InternVL3-2B-LoRA Zhu et al. (2025)	0.9127	0.9836	0.8571	0.9160	0.9196	2.7726 $\pm$ 0.0364
LAVAD ( Zanella et al., 2024 )	0.7724	0.7681	0.7571	0.7626	0.7719	4.8936 $\pm$ 0.1034
VERA ( Ye et al., 2025 )	0.6448	0.5781	0.9786	0.7268	0.6560	3.7055 $\pm$ 0.9577
PEL4VAD ( Pu et al., 2024 )	0.8724	0.8503	0.8929	0.8711	0.8731	2.5132 $\pm$ 0.1236

Table 4: Comparison with other models on UCF-crime dataset.

Figure 5: Confusion matrix for different models on UCF-crime dataset.

Figure 6: Qualitative clip‑level comparisons across models, prompts, and LoRA adaptation on representative surveillance scenarios.

Vision-Language Models

Vision-language models connect visual inputs with text for retrieval, question answering, generation, and grounded multimodal reasoning.

See all 736 papers

Parser-Free VLM Verification for Federated Weakly Supervised Video Anomaly Detection
Sep 7, 2026Sébastien Thuau, Amira Gran, Siba Haidar +1Multimodal Anomaly Detection Weakly Supervised Video Anomaly Detection
esieaLab, ESIEA, Paris, France · ETIS Laboratory, CNRS, UMR8051, CY Cergy Paris University, Paris, France
MuST-VAD: Mutual Structured Learning for Video Anomaly Detection
Aug 7, 2026Satoshi Hashimoto, Hitoshi Nishimura, Mori KurokawaVLM Adaptation Weakly Supervised Video Anomaly Detection
KDDI Research, Inc. Fujimino, Saitama, Japan
COPRA: Conditional Parameter Adaptation with Reinforcement Learning for Video Anomaly Detection
May 14, 2026Darryl Cherian Jacob, Xinyu Liu, Kai Wang +1VLM Adaptation Video Anomaly Detection
Auburn University · Tencent Hunyuan

Sep 7, 2026J/K move · Enter open · S save

Parser-Free VLM Verification for Federated Weakly Supervised Video Anomaly Detection

Sébastien Thuau, Amira Gran, Siba Haidar, Rachid Chelouah

esieaLab, ESIEA, Paris, France · ETIS Laboratory, CNRS, UMR8051, CY Cergy Paris University, Paris, France

Multimodal Anomaly Detection Weakly Supervised Video Anomaly Detection Federated Learning Video Surveillance

Open paper AskPDF arXiv

How can vision-language models help video anomaly detection (VAD) when surveillance data remain distributed, weakly labeled, and resource-constrained? Most weakly supervised VAD methods assume centralized training; recent VLM-based extensions further rely on dense inference, generated explanations, or additional adaptation. We introduce a lightweight federated MIL-VLM cascade in which only a compact MIL scorer is trained across clients, while a frozen VLM verifies high-scoring suspect segments post hoc. We study two VLM feedback interfaces: parsed text-generation decisions and a logit-based interface that extracts a continuous anomaly score from next-token Yes/No probabilities. Experiments on UCF-Crime with InternVL3.5-2B and Qwen3-VL-2B-Instruct show that text-generation verification can improve frame-level AUC after diagnostic temporal post-processing, but remains sensitive to prompts, parsers, model choice, and smoothing. In contrast, the logit interface provides a fixed parser-free signal that improves both frame-level AUC and frame-level AP over the MIL baseline across both VLMs, without temporal post-processing in its main configuration. Since suspect segments are updated independently once available, next-token logit feedback provides a simple segment-local alternative to text-generation verification.

Abstract

Vision-Language Models

Explore similar work

Parser-Free VLM Verification for Federated Weakly Supervised Video Anomaly Detection

MuST-VAD: Mutual Structured Learning for Video Anomaly Detection

COPRA: Conditional Parameter Adaptation with Reinforcement Learning for Video Anomaly Detection