VADMamba++: Efficient Video Anomaly Detection via Hybrid Modeling in Grayscale Space

Jihao Lyu; Minghua Zhao; Jing Hu; Yifei Chen; Shuangli Du; Cheng Shi

cs.CVApr 1, 2026

VADMamba++: Efficient Video Anomaly Detection via Hybrid Modeling in Grayscale Space

Authors: Jihao Lyu, Minghua Zhao, Jing Hu, Yifei Chen, Shuangli Du, Cheng Shi

Organizations: Shaanxi Key Laboratory for Network Computing and Security Technology, School of Computer Science and Engineering, Xi’an University of Technology, Xi’an, 710048, China.

Abstract

VADMamba pioneered the introduction of Mamba to Video Anomaly Detection (VAD), achieving high accuracy and fast inference through hybrid proxy tasks. Nevertheless, its heavy reliance on optical flow as auxiliary input and inter-task fusion scoring constrains its applicability to a single proxy task. In this paper, we introduce VADMamba++, an efficient VAD method based on the Gray-to-RGB paradigm that enforces a Single-Channel to Three-Channel reconstruction mapping, designed for a single proxy task and operating without auxiliary inputs. This paradigm compels inferring color appearances from grayscale structures, allowing anomalies to be more effectively revealed through dual inconsistencies between structure and chromatic cues. Specifically, VADMamba++ reconstructs grayscale frames into the RGB space to simultaneously discriminate structural geometry and chromatic fidelity, thereby enhancing sensitivity to explicit visual anomalies. We further design a hybrid modeling backbone that integrates Mamba, CNN, and Transformer modules to capture diverse normal patterns while suppressing the appearance of anomalies. Furthermore, an intra-task fusion scoring strategy integrates explicit future-frame prediction errors with implicit quantized feature errors, further improving accuracy under a single task setting. Extensive experiments on three benchmark datasets demonstrate that VADMamba++ outperforms state-of-the-art methods while meeting performance and efficiency, especially under a strict single-task setting with only frame-level inputs.

Figures & tables

Fig. 1 : Comparison of AUC and FPS on Ped2 [ 20 ] , where single-task methods are marked with black edges. The blue dashed line shows the typical tradeoff: higher accuracy often reduces speed, while our VADMamba++ notably breaks this trend.

Fig. 2 : Pipeline comparison between (a) VADMamba and (b) VADMamba++. We highlight three key evolutions: 1) Multi-input vs. Single-input ; 2) Multi-task vs. Single-task ; and 3) Inter-task fusion vs. Intra-task fusion .

Fig. 3 : Overview of the proposed VADMamba++, which employs an asymmetric Mamba-based encoder–decoder framework to colorize grayscale input frames into the next RGB frame. The TMC encoder combines Transformer, Mamba, and CNN blocks for spatiotemporal feature extraction, whereas the MC decoder reconstructs a colorized frame guided by the quantized representation from the VQ module.

Fig. 4 : The architectures of (a)BiSS2D and (b)DPA.

TABLE I : Comparison with state-of-the-art methods on three benchmark datasets. We list the feature of the methods, in which ’F’ denotes frame, ’OF’ denotes optical flow, ’FG’ denotes foreground, ’FD’ denotes frame differential, ’SA’ denotes synthetic anomaly. ∗ denotes the same GPU device. The gray marker denotes Mamba-based methods.

Fig. 5 : Examples of ground truth frames, predicted results, and error maps across five scenarios. The first four columns depict abnormal cases, whereas the last column presents a normal case. Brighter regions in the error maps indicate larger prediction errors.

Fig. 6 : Qualitative comparison of normal (green box) and anomaly (red box) results across different methods. (a) VADMamba, (b) VADMamba++ (w/o G2R), (c) VADMamba++, and (d) Ground Truth object.

Fig. 7 : Frame-level anomaly scores on (a) Ped2, (b) Avenue, and (c) SHT. Red boxes denote abnormal frames, while dotted, dashed, and solid lines indicate implicit, explicit, and fusion scores, respectively.

Fig. 8 : Sensitivity analysis of (a) $k$ and (b) $\lambda$ on three datasets.

w/o	G2R	T	BiSS2D	DPA	$S_{I}$
Ped2	98.5	98.4	96.2	98.2	98.1
Avenue	89.8	89.8	89.0	89.4	88.9
SHT	74.3	74.7	73.6	74.2	73.5

TABLE II : Ablation studies on the component.

Panel of Table 11 — TABLE III : Impact of encoder-decoder feature extraction on Ped2.

	TMC	TCM	MCT	CMT
Ped2	99.6	98.9	98.8	98.6
Avenue	91.4	90.7	90.5	89.8
SHT	76.2	75.4	74.9	74.3

TABLE IV : Impact of varying module ordering in the encoder.

$\mathcal{L}_{pred}$	$\mathcal{L}_{vq}$	$\mathcal{L}_{grad}$	$\mathcal{L}_{color}$	Ped2	Avenue	SHT
✓				98.2	88.9	73.1
✓	✓			98.4	89.2	73.6
✓	✓	✓		98.6	89.8	74.4
✓	✓	✓	✓	99.6	91.4	76.2

TABLE V : Effectiveness of loss functions.

$T$	4	8	12	16	20
AUC(%) $\uparrow$	98.7	99.1	99.3	99.6	99.1

TABLE VI : The AUC for different values of

T

on the Ped2 dataset.

	Ped2	Avenue	SHT
VADMamba (FP) [ 2 ]	98.1	86.2	74.4
+ G2R	98.2 (+0.1)	86.8 (+0.6)	74.7 (+0.3)
+ Implicit	98.8 (+0.6)	87.3 (+1.1)	74.9 (+0.5)
+ G2R & Implicit	98.9 (+0.7)	88.0 (+1.8)	75.3 (+0.9)

TABLE VII : Incremental evaluation of VADMamba with the G2R paradigm and the Implicit score.

Explore similar work

A Unified Score Matching Paradigm for Video Anomaly Detection and Anticipation
Oct 8, 2026Congqi Cao, Zhenhe Liang, Hanwen Zhang +4Video Anomaly Detection Score-Based Generative Modeling
National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, School of Computer Science, Northwestern Polytechnical University, Xi’an 710129, China · School of Electronics and Information, Northwestern Polytechnical University, Xi’an 710129, China
Geometry-Aware Semantic Reasoning for Training Free Video Anomaly Detection
Mar 10, 2026Ali Zia, Usman Ali, Muhammad Umer Ramzan +3Video Anomaly Detection Multimodal Anomaly Detection
School of Computing, Engineering & Mathematical Sciences, La Trobe University, Melbourne, Australia · School of Engineering and Applied Sciences, GIFT University, Gujranwala 52250, Pakistan
Steering and Rectifying Latent Representation Manifolds in Frozen Multi-modal LLMs for Video Anomaly Detection
Feb 27, 2026Zhaolin Cai, Fan Li, Huiyu Duan +2Video Anomaly Detection Multimodal Anomaly Detection
Xinjiang University · Xi’an Jiao Tong University · Shanghai Jiao Tong University +1

Oct 8, 2026J/K move · Enter open · S save

A Unified Score Matching Paradigm for Video Anomaly Detection and Anticipation

Congqi Cao, Zhenhe Liang, Hanwen Zhang, Yifan Zhao+3

National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, School of Computer Science, Northwestern Polytechnical University, Xi’an 710129, China · School of Electronics and Information, Northwestern Polytechnical University, Xi’an 710129, China

Video Anomaly Detection Score-Based Generative Modeling

Open paper AskPDF arXiv

Video anomaly detection (VAD) is a fundamental and safety-critical task in computer vision. Recent generative approaches detect anomalies from a distributional perspective, but remain limited by local anomaly modes. Meanwhile, video anomaly anticipation (VAA), as a proactive extension beyond post-hoc detection, introduces additional challenges. In particular, the contrastive inference paradigm in VAD, which relies on ground-truth frames, is not applicable to VAA, hindering its development. To address these challenges, we propose a unified score-driven framework, termed Uni-DSM, based on denoising score matching (DSM), which models anomaly patterns through likelihood estimation and score functions over the learned data distribution. Within this unified framework, we adopt a shared noise-conditioned score transformer backbone with scene-dependent embeddings and motion-aware weighting for distribution-level modeling. Instead of introducing separate architectures, Uni-DSM unifies VAD and VAA through different inference and supervision paradigms built upon the same score-based formulation. For VAD, we instantiate an autoregressive denoising score matching (ADSM) mechanism, which progressively accumulates anomalous evidence via autoregressive denoising, enabling enhanced perception of local modes beyond visual cues. For VAA, we extend the same architecture by incorporating a lightweight auxiliary decoder and a novel self-distilled denoising score matching (SDSM) mechanism. By constructing supervision from output discrepancies instead of relying on unavailable future ground truth, our method achieves efficient training suitable or early anomaly anticipation. Extensive experiments on multiple benchmark datasets demonstrate state-of-the-art performance in both VAD and VAA while maintaining high efficiency, establishing a unified and scalable pipeline from anomaly detection to anticipation.

Abstract

Explore similar work

A Unified Score Matching Paradigm for Video Anomaly Detection and Anticipation

Geometry-Aware Semantic Reasoning for Training Free Video Anomaly Detection

Steering and Rectifying Latent Representation Manifolds in Frozen Multi-modal LLMs for Video Anomaly Detection