Research map / Video / Video-language understanding
VADTree: Explainable Training-Free Video Anomaly Detection via Hierarchical Granularity-Aware Tree
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.
Proposes VADTree, a training-free video anomaly detection framework that builds a hierarchical granularity-aware tree from pre-trained GEBD boundaries and scores nodes with a VLM (LLaVA-Video-7B) and LLM (DeepSeek-R1-Distill-Qwen-14B), reaching SOTA among training-free methods on UCF-Crime and XD-Violence.
Problem
Training-free explainable video anomaly detection struggles because fixed-length temporal window sampling cannot match anomalies of varying durations, causing noisy semantics and VLM hallucination.
Approach
Replace fixed sliding windows with an adaptive binary tree built from GEBD boundary confidences, stratified into coarse and fine clusters via K-means; node-wise VLM captioning with injected priors, LLM scoring, intra-cluster similarity-weighted refinement, and variance-based inter-cluster fusion.
Key results
- VADTree achieves 84.74% AUC on UCF-Crime, 4.5 points above LAVAD (80.28%) and 2.7 above EventVAD (82.03%) among training-free methods.
- On XD-Violence, VADTree reaches 90.44% AUC and 67.82% AP, surpassing LAVAD by 5.1 AUC points.
- On MSAD, VADTree achieves 89.32% AUC and 71.41% AP, exceeding the weakly supervised π-VAD on AUC (88.68%) and AP (71.26%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VADTree instead of SUVAD | Training-free VAD on UCF-Crime · UCF-Crime | AUC (%): 83.90 -> 84.74 | Helped |
| Use VADTree instead of EventVAD | Training-free VAD on XD-Violence · XD-Violence | AUC (%): 87.51 -> 90.44 | Helped |
| Use VADTree instead of π-VAD | VAD on MSAD · MSAD | AUC (%): 88.68 -> 89.32 | Helped |
| Use VADTree instead of π-VAD | VAD on MSAD · MSAD | AUCₐ (%): 71.25 -> 67.85 | Hurt |
| Add audio captions from Kimi-Audio-7B-Instruct | VADTree on XD-Violence · XD-Violence | AUC (%): 90.44 -> 90.55 | Helped |
| Add prior-infused node scoring (ablation) | HGTree fine cluster on UCF-Crime · UCF-Crime | AUC (%): 71.57 -> 75.67 | Helped |
| Add intra-cluster node refinement (ablation) | Prior-infused node scoring on UCF-Crime · UCF-Crime | AUC (%): 75.67 -> 83.05 | Helped |
| Add inter-cluster node correlation (ablation) | VADTree on UCF-Crime · UCF-Crime | AUC (%): 83.05 -> 84.74 | Helped |
| Use coarse and fine clusters instead of fine cluster only (ablation) | HGTree on UCF-Crime · UCF-Crime | AUC (%): 82.81 -> 84.74 | Helped |
| Replace K-Means with K-Medoids (ablation) | VADTree HGTree construction on UCF-Crime · UCF-Crime | AUC (%): 84.74 -> 85.24 | Helped |
| Increase refinement neighborhood size from K=0 to K=10 (ablation) | VADTree-Fine on UCF-Crime · UCF-Crime | AUC (%): 73.17 -> 83.05 | Helped |
| Increase refinement neighborhood size from K=0 to K=10 (ablation) | VADTree-Coarse on UCF-Crime · UCF-Crime | AUC (%): 75.67 -> 82.81 | Helped |
Surprising findings
- On MSAD, the training-free VADTree outperforms several weakly supervised methods on overall AUC and AP, though it trails π-VAD on anomaly-specific AUC_a and AP_a.
- Replacing K-Means with K-Medoids for clustering gave a further gain (85.24% vs 84.74% AUC), and using only leaf nodes (single cluster) lowered AUC by 1.9 points.
Limitations
- The LLM 'think' mode adds inference overhead, which the authors retain deliberately.
- Performance depends on GEBD model quality and on the VLM/LLM combination, with fluctuations across configurations.
- Scores from coarse and fine nodes can diverge due to lack of uniform scoring standard across independent VLM/LLM inferences.
Experimental setup
Datasets: UCF-Crime XD-Violence MSAD Kinetics-GEBD
Benchmarks: UCF-Crime XD-Violence MSAD
Metrics: AUC AP AUC_a AP_a NoS mIoU mIF
Baselines: LAVAD EventVAD VERA SUVAD MCANet LLaVA-1.5 Video-Llama2 Holmes-VAU Holmes-VAD VADor π-VAD GS-MoE
Builds on: LAVAD EfficientGEBD Kinetics-GEBD VideoTree
Closest papers by meaning
- Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity (CVPR 2025 · similarity 0.91)
- Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · similarity 0.90)
- Ex-VAD: Explainable Fine-grained Video Anomaly Detection Based on Visual-Language Models (ICML 2025 · similarity 0.90)
- Learning Event Completeness for Weakly Supervised Video Anomaly Detection (ICML 2025 · similarity 0.89)
- Track Any Anomalous Object:A Granular Video Anomaly Detection Pipeline (CVPR 2025 · similarity 0.89)
- Open-Vocabulary Video Anomaly Detection (CVPR 2024 · similarity 0.89)
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.