Research map / Video / Video-language understanding
Text Prompt with Normality Guidance 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 TPWNG, a CLIP-based pseudo-label generation and self-training framework for weakly supervised video anomaly detection that uses learnable text prompts, a normality visual prompt, and temporal context self-adaptive learning.
Problem
Weakly supervised video anomaly detection needs accurate frame-level pseudo-labels, but existing pseudo-label self-training methods use only visual modality and ignore category text and temporal adaptivity.
Approach
Fine-tune CLIP with two ranking losses and a distributional inconsistency loss to align event description text with video frames, add learnable text prompts and a normality visual prompt, use normality guidance to infer frame pseudo-labels, and model temporal context with an adaptive attention span.
Key results
- Achieves 87.79% AUC on UCF-Crime, outperforming CLIP-TSA (87.58%) by 0.21%.
- Achieves 83.68% AP on XD-Violence, outperforming CLIP-TSA (82.17%) by 1.52%.
- Removing the normality visual prompt drops performance to 85.25% AUC on UCF-Crime and 81.58% AP on XD-Violence.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TPWNG instead of CLIP-TSA | WSVAD · UCF-Crime | AUC: 87.58% -> 87.79% | Helped |
| use TPWNG instead of CLIP-TSA | WSVAD · XD-Violence | AP: 82.17% -> 83.68% | Helped |
| add a normality visual prompt using match-similarity aggregation (ablation) | TPWNG · UCF-Crime | AUC: 85.25% -> 87.79% | Helped |
| add a normality visual prompt using match-similarity aggregation (ablation) | TPWNG · XD-Violence | AP: 81.58% -> 83.68% | Helped |
| use match-similarity aggregation instead of frame averaging for the normality visual prompt (ablation) | TPWNG · UCF-Crime | AUC: 87.32% -> 87.79% | Helped |
| use match-similarity aggregation instead of frame averaging for the normality visual prompt (ablation) | TPWNG · XD-Violence | AP: 83.13% -> 83.68% | Helped |
| add normality guidance to pseudo-label generation (ablation) | TPWNG · UCF-Crime | AUC: 85.83% -> 87.79% | Helped |
| add normality guidance to pseudo-label generation (ablation) | TPWNG · XD-Violence | AP: 81.32% -> 83.68% | Helped |
| use TCSAL instead of a Transformer encoder (ablation) | TPWNG · UCF-Crime | AUC: 85.12% -> 87.79% | Helped |
| use TCSAL instead of a Transformer encoder (ablation) | TPWNG · XD-Violence | AP: 80.02% -> 83.68% | Helped |
| add the normal-video ranking loss (ablation) | TPWNG with classification, sparsity, and smoothing losses · UCF-Crime | AUC: 77.12% -> 81.34% | Helped |
| add the abnormal-video ranking loss (ablation) | TPWNG with classification, sparsity, and smoothing losses · UCF-Crime | AUC: 77.12% -> 84.45% | Helped |
Limitations
- Relies on frozen CLIP image encoder with only the text encoder's final projection fine-tuned.
- Pseudo-label threshold and guidance weight are dataset-specific hyperparameters.
Experimental setup
Datasets: UCF-Crime XD-Violence
Benchmarks: UCF-Crime XD-Violence
Metrics: AUC AP
Baselines: Sultani et al. GCN HL-Net RTFM CRFD MSL MGFN UR-DMU CLIP-TSA MIST Zhang et al.
Builds on: CLIP CoOp RTFM MIST
Closest papers by meaning
- Prompt-Enhanced Multiple Instance Learning for Weakly Supervised Video Anomaly Detection (CVPR 2024 · similarity 0.90)
- Learning Event Completeness for Weakly Supervised Video Anomaly Detection (ICML 2025 · similarity 0.89)
- Federated Weakly Supervised Video Anomaly Detection with Multimodal Prompt (AAAI 2025 · similarity 0.88)
- VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly Detection (AAAI 2024 · similarity 0.87)
- FrameShield: Adversarially Robust Video Anomaly Detection (NeurIPS 2025 · similarity 0.86)
- Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · similarity 0.86)
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.