Research map / Video / Video segmentation and tracking
Collaborative Learning of Anomalies with Privacy (CLAP) for Unsupervised Video Anomaly Detection: A New Baseline
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 CLAP, a federated, fully unsupervised video anomaly detection method that generates pseudo-labels via entropy-based clustering and Gaussian hypothesis testing, refined through server aggregation and local feedback, evaluated on UCF-Crime and XD-Violence under new collaborative data splits.
Problem
Unsupervised video anomaly detection in surveillance when training data is distributed across privacy-sensitive participants who cannot share raw videos.
Approach
Combines common-knowledge-based data segregation (GMM clustering on video feature statistics including von Neumann entropy, then collaborative Gaussian hypothesis testing for segment pseudo-labels) with FedAvg server aggregation and a local pseudo-label refinement loop.
Key results
- Centralized AUC of 80.9% on UCF-Crime and 81.71% on XD-Violence, outperforming C2FPL (80.65%, 80.09%).
- Collaborative setting with five participants reaches 78.02% AUC on UCF-Crime and 77.65% on XD-Violence, versus 67.12% and 68.19% for GCL.
- Ablation on UCF-Crime: FedAvg alone 76.2%, adding SKA 77.1%, adding PLR 78.02%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use collaborative CLAP instead of C2FPL | unsupervised VAD · UCF-Crime, collaborative | AUC: 75.20 -> 78.02 | Helped |
| use collaborative CLAP instead of C2FPL | unsupervised VAD · XD-Violence, collaborative | AUC: 74.36 -> 77.65 | Helped |
| use CLAP instead of the strongest listed baseline | weakly supervised VAD · UCF-Crime, centralized | AUC: 85.30 -> 85.50 | Helped |
| use CLAP instead of C2FPL | weakly supervised VAD · XD-Violence, collaborative | AUC: 76.98 -> 85.67 | Helped |
| increase participant count from 5 to 50 | CLAP · UCF-Crime, random split | AUC: 78.02 -> 67.92 | Hurt |
| use scene-based rather than random splitting | CLAP · UCF-Crime | AUC: 78.02 -> 73.99 | Hurt |
| add server knowledge accumulation (SKA) (ablation) | CLAP · UCF-Crime | AUC: 76.2 -> 77.1 | Helped |
| add pseudo-label refinement (PLR) (ablation) | CLAP · UCF-Crime | AUC: 77.1 -> 78.02 | Helped |
| provide video-level labels to more participants | CLAP · UCF-Crime, 25 participants | AUC: 76.54 -> 83.23 | Helped |
Surprising findings
- Performance drops notably when participants increase to 50 on UCF-Crime (67.92% AUC), attributed to fewer videos per participant.
Limitations
- Performance drops as the number of participants increases.
- Requires many videos per participant; the weakly supervised extension is outside the core unsupervised scope.
Experimental setup
Datasets: UCF-Crime XD-Violence
Benchmarks: UCF-Crime XD-Violence
Metrics: AUC
Baselines: GCL C2FPL PRV Sultani et al. RTFM MSL S3R CLAWS+
Builds on: GCL C2FPL FedAvg
Closest papers by meaning
- Advancing Video Anomaly Detection: A Concise Review and a New Dataset (NeurIPS 2024 · similarity 0.86)
- Federated Weakly Supervised Video Anomaly Detection with Multimodal Prompt (AAAI 2025 · similarity 0.85)
- Open-Vocabulary Video Anomaly Detection (CVPR 2024 · similarity 0.85)
- Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · similarity 0.85)
- MoniTor: Exploiting Large Language Models with Instruction for Online Video Anomaly Detection (NeurIPS 2025 · similarity 0.85)
- Learning Event Completeness for Weakly Supervised Video Anomaly Detection (ICML 2025 · similarity 0.84)
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.