Research map / Video / Video segmentation and tracking
Advancing Video Anomaly Detection: A Concise Review and a New Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents a concise review of video anomaly detection (VAD) and introduces the MSAD dataset (14 scenarios, 720 videos) plus a few-shot scenario-adaptive model SA^2D, benchmarking weakly supervised and few-shot methods on it.
Problem
Existing VAD datasets and methods are limited to single scenarios, mostly human-related anomalies, and few camera views, so models generalize poorly across scenarios and viewpoints; concise reviews are also lacking.
Approach
Introduces a multi-scenario dataset (MSAD) with human and non-human anomalies under varied lighting and weather, and a scenario-adaptive few-shot future-frame-prediction model (SA^2D) that uses scenario grouping in N-way K-shot meta-learning.
Key results
- On CUHK Avenue, SA^2D trained on MSAD outperforms the view-adaptive FSAD baseline by 9.6% (Micro) and 6.2% (Macro) AUC in cross-scenario evaluation.
- On MSAD with I3D features, RTFM reaches 86.65 Micro AUC and TEVAD 86.82, with no single best method across backbones and datasets.
- Weakly supervised methods show backbone sensitivity on MSAD; e.g., UR-DMU drops from 85.02 (I3D) to 72.36 (SwinT).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train SA²D on MSAD instead of ShanghaiTech | SA²D · Ped2 | Micro AUC: 57.38 -> 70.35 | Helped |
| Train SA²D on MSAD instead of ShanghaiTech | SA²D · CUHK Avenue | Micro AUC: 69.98 -> 79.57 | Helped |
| Train SA²D on MSAD instead of ShanghaiTech | SA²D · CUHK Avenue | Macro AUC: 78.32 -> 84.49 | Helped |
| Use scenario information in few-shot sampling (ablation) | SA²D · ShanghaiTech v1 | Micro AUC: 63.74 -> 68.96 | Helped |
| Use scenario information in few-shot sampling (ablation) | SA²D · ShanghaiTech v3 | Micro AUC: 64.39 -> 67.59 | Helped |
| Use scenario information in few-shot sampling (ablation) | SA²D · ShanghaiTech v5 | Micro AUC: 55.04 -> 55.74 | Helped |
| Use scenario information in few-shot sampling (ablation) | SA²D · ShanghaiTech v6 | Micro AUC: 70.26 -> 75.47 | Helped |
| Use scenario information in few-shot sampling (ablation) | SA²D · ShanghaiTech v8 | Micro AUC: 59.97 -> 60.85 | Helped |
| Use SwinTransformer instead of I3D as the MSL backbone | MSL · UCF-Crime | AUC: 82.85 -> 85.62 | Helped |
| Use SwinTransformer instead of I3D as the MSL backbone | MSL · MSAD | AUC: 85.30 -> - | Hurt |
Surprising findings
- Models trained on MSAD perform poorly on ShanghaiTech but better on CUHK and UCSD Ped2, which the authors attribute to unrealistic anomaly categories such as biking and driving in ShanghaiTech.
- Weakly supervised methods are sensitive to backbone choice on MSAD, with I3D generally outperforming SwinTransformer on this dataset while the opposite holds on UCF-Crime.
Limitations
- Only static-camera surveillance videos are included, and the dataset covers only 14 scenarios.
- Anomaly definitions may be unrealistic in some categories (e.g., biking and driving), and the paper does not report error bars.
- Privacy is handled by face and license-plate blurring, and deepfake-based anonymization was not completed.
Experimental setup
Datasets: MSAD ShanghaiTech UCF-Crime UCSD Ped2 CUHK Avenue
Benchmarks: MSAD ShanghaiTech UCF-Crime CUHK Avenue UCSD Ped2
Metrics: Micro AUC Macro AUC frame-level AUC
Baselines: FSAD MIST RTFM MSL UR-DMU MGFN TEVAD
Builds on: FSAD MAML-style meta-learning RTFM U-Net ConvLSTM
Closest papers by meaning
- Open-Vocabulary Video Anomaly Detection (CVPR 2024 · similarity 0.90)
- Track Any Anomalous Object:A Granular Video Anomaly Detection Pipeline (CVPR 2025 · similarity 0.89)
- MoniTor: Exploiting Large Language Models with Instruction for Online Video Anomaly Detection (NeurIPS 2025 · similarity 0.89)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.88)
- Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · similarity 0.88)
- Towards Multi-Domain Learning for Generalizable Video Anomaly Detection (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.