Research map / Video / Video segmentation and tracking
Learning Event Completeness for Weakly Supervised Video Anomaly Detection
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 LEC-VAD, a CLIP-based dual-branch weakly supervised video anomaly detector that adds memory-bank text prototypes and Gaussian-mixture-prior local consistency losses to produce more complete anomaly intervals.
Problem
Weakly supervised video anomaly detection with only video-level labels often yields fragmented, incomplete anomaly intervals because frame-wise classification training does not match segment-level localization at inference.
Approach
A dual category-agnostic and category-aware structure with CLIP text prototypes stored in a momentum-updated memory bank, plus learnable Gaussian-mixture masks that enforce local smoothness of anomaly scores so predicted events are more complete.
Key results
- On XD-Violence coarse-grained, LEC-VAD reaches 88.47 AP with I3D (vs. PE-MIL 88.05) and 86.56 AP with CLIP (vs. ReFLIP 85.81).
- On UCF-Crime coarse-grained AUC with CLIP, LEC-VAD reaches 89.97 vs. ITC 89.04.
- Fine-grained XD-Violence mAP AVG 35.14 vs. ReFLIP 27.36; UCF-Crime AVG 13.56 vs. ReFLIP 9.62.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LEC-VAD instead of the strongest listed baseline with I3D features | XD-Violence coarse-grained detection · XD-Violence | AP(%): 88.05 -> 88.47 | Helped |
| use LEC-VAD instead of ReFLIP with CLIP features | XD-Violence coarse-grained detection · XD-Violence | AP(%): 85.81 -> 86.56 | Helped |
| use LEC-VAD instead of GLAWS with C3D features | UCF-Crime coarse-grained detection · UCF-Crime | AUC(%): 83.03 -> 84.75 | Helped |
| use LEC-VAD instead of MGFN with I3D features | UCF-Crime coarse-grained detection · UCF-Crime | AUC(%): 86.98 -> 88.21 | Helped |
| use LEC-VAD instead of ITC with CLIP features | UCF-Crime coarse-grained detection · UCF-Crime | AUC(%): 89.04 -> 89.97 | Helped |
| use LEC-VAD instead of ReFLIP | XD-Violence fine-grained detection · XD-Violence | mAP@IoU AVG: 27.36 -> 35.14 | Helped |
| use LEC-VAD instead of ReFLIP | UCF-Crime fine-grained detection · UCF-Crime | mAP@IoU AVG: 9.62 -> 13.56 | Helped |
| add CMB to the VOB-only model (ablation) | LEC-VAD · XD-Violence fine-grained | mAP@IoU AVG: 30.17 -> 34.94 | Helped |
| add VOB to the CMB-only model (ablation) | LEC-VAD · XD-Violence fine-grained | mAP@IoU AVG: 33.23 -> 34.94 | Helped |
| add Lreg to a model without Lreg or Lgmm (ablation) | LEC-VAD · XD-Violence fine-grained | mAP@IoU AVG: 30.04 -> 32.49 | Helped |
| add Lgmm to a model without Lreg or Lgmm (ablation) | LEC-VAD · XD-Violence fine-grained | mAP@IoU AVG: 30.04 -> 33.51 | Helped |
| add VAP to a model without VAP or PMB (ablation) | LEC-VAD · XD-Violence fine-grained | mAP@IoU AVG: 28.90 -> 32.73 | Helped |
Limitations
- Coarse-grained gains over strongest baselines are small (e.g., under 1 AP/AUC point on several settings).
- Relies on fixed hyperparameters tuned on test-set sweeps; sensitivity shown only for a few parameters.
Experimental setup
Datasets: XD-Violence UCF-Crime
Benchmarks: XD-Violence UCF-Crime
Metrics: AP AUC mAP@IoU AVG mAP
Baselines: ReFLIP ITC VadCLIP VadCLIP-style methods (PE-MIL, DMU, PEL4VAD) CLIP-TSA OVVAD STPrompt RAD AVVD
Builds on: CLIP AutoLoc PEL4VAD ReFLIP VadCLIP
Closest papers by meaning
- VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly Detection (AAAI 2024 · similarity 0.91)
- Prompt-Enhanced Multiple Instance Learning for Weakly Supervised Video Anomaly Detection (CVPR 2024 · similarity 0.90)
- Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · similarity 0.90)
- VADTree: Explainable Training-Free Video Anomaly Detection via Hierarchical Granularity-Aware Tree (NeurIPS 2025 · similarity 0.89)
- Federated Weakly Supervised Video Anomaly Detection with Multimodal Prompt (AAAI 2025 · similarity 0.89)
- Text Prompt with Normality Guidance for Weakly Supervised 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.