Research map / Video / Video segmentation and tracking
TD-VAD: Breaking Visual Dependence in Video Anomaly Detection with Text-Driven Learning
Shuangqing Zhang, Lei-Lei Ma, Zhao Wang, Wen Dong, Xinyi Xu, Guo-Sen Xie, Caifeng Shan, Fang Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train VAD on LLM-generated video-like text with frozen CLIP, test on video frames | TD-VAD · XD-Violence | AP (%): 38.21 -> 75.83 | Improved |
| add both ECC-Attn and EFC-Attn event-evolution causal attention (ablation) | TD-VAD · XD-Violence | AP (%): 30.48 -> 75.83 | Improved |
| add EFC-Attn on top of ECC-Attn (ablation) | TD-VAD · XD-Violence | AP (%): 71.29 -> 75.83 | Improved |
| add ECC-Attn on top of EFC-Attn (ablation) | TD-VAD · XD-Violence | AP (%): 73.09 -> 75.83 | Improved |
| add binary anomaly head to the multi-class head (hierarchical branch) (ablation) | TD-VAD · XD-Violence | AP (%): 74.83 -> 75.83 | Improved |
| add multi-class head to the binary head (ablation) | TD-VAD · XD-Violence | AP (%): 68.22 -> 75.83 | Improved |
| increase timestamps per text description from 1 to 4 (ablation) | TD-VAD · XD-Violence | AP (%): 71.79 -> 75.83 | Improved |
| increase timestamps from 4 to 8 (ablation) | TD-VAD · XD-Violence | AP (%): 75.83 -> 74.67 | Hurt |
| vary EFC-Attn window length (2, 4, 8, 16) (ablation) | TD-VAD · XD-Violence | AP (%): 75.24 / 75.83 / 74.98 / 75.12 | No clear effect |
| change LLM prompt style (law enforcement, educational, domain expert) (ablation) | text generation · XD-Violence | AUC change (%): -0.50 / +0.02 / -0.97 | No clear effect |
| randomly drop 10-30% of temporal text segments in training (ablation) | TD-VAD · XD-Violence | AP (%): 75.83 -> 76.19 (10% removed) | No clear effect |
| generate training text with Kimi-k2.5 or GPT-3 instead of DeepSeek-V3 (ablation) | text generation · XD-Violence | AP (%): 75.83 -> 74.50 (Kimi-k2.5); 75.54 (GPT-3) | Hurt |
Closest papers by meaning
- Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · similarity 0.92)
- Ex-VAD: Explainable Fine-grained Video Anomaly Detection Based on Visual-Language Models (ICML 2025 · similarity 0.91)
- MoniTor: Exploiting Large Language Models with Instruction for Online Video Anomaly Detection (NeurIPS 2025 · similarity 0.90)
- Learning Event Completeness for Weakly Supervised Video Anomaly Detection (ICML 2025 · similarity 0.89)
- VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly Detection (AAAI 2024 · 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.
