Research map / Video / Video-language understanding
Pistachio: Towards Synthetic, Balanced, and Long-Form Video Anomaly Benchmarks
jie li, Hongyi Cai, Mingkang Dong, Muxin Pu, Shan You, Fei Wang, Tao Huang
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 on Pistachio plus 10% of UCF-Crime instead of full UCF-Crime only (ablation) | CLIP-TSA · UCF-Crime test | AUC / AP: 82.56/21.62 -> 83.96/24.25 | Improved |
| increase the UCF-Crime share mixed with Pistachio from 10% to 50% (ablation) | CLIP-TSA · UCF-Crime test | AP: 24.25 -> 33.04 | Improved |
| train on Pistachio only instead of UCF-Crime | CLIP-TSA · UCF-Crime test | AUC: 82.56 -> 61.58 | Hurt |
| train on Pistachio instead of UCF-Crime | CLIP-TSA · XD-Violence test | AP: 51.75 -> 62.48 | Improved |
| train on Pistachio instead of UCF-Crime | CLIP-TSA · Hailuo test | AUC: 68.71 -> 75.89 | Improved |
| joint training on all five datasets instead of Pistachio only (ablation) | CLIP-TSA · Cosmos test | AP: 58.74 -> 87.37 | Improved |
| use training-free VADTree instead of trained methods | VADTree · Pistachio VAD test | AP: 25.15 AP (72.52 AUC) vs Fed-WSVAD 71.9 | Hurt |
| use storyline-based last-frame chaining instead of direct one-shot long generation (ablation) | Wan generation pipeline · VLM quality score | Avg. quality: 2.75 -> 4.17 | Improved |
| add multi-step storyline to chained generation (ablation) | Wan generation pipeline · VLM quality score | Avg. quality: 3.45 -> 4.17 | Improved |
| use last-frame chaining instead of fixed first-last-frame looping (ablation) | Wan generation pipeline · VLM quality score | Avg. quality: 4.04 -> 4.17 | Improved |
| scale VAU model from InternVL3-8B to 14B (ablation) | InternVL3 · Pistachio-VAU | Avg F1: 29.11 -> 29.39 | No clear effect |
| add 4-step distillation and TeaCache to Wan generation (ablation) | Wan generation · Generation efficiency | time per video: 349.7s -> 30.0s | Improved |
Closest papers by meaning
- Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity (CVPR 2025 · similarity 0.90)
- Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
- PANDA: Towards Generalist Video Anomaly Detection via Agentic AI Engineer (NeurIPS 2025 · similarity 0.88)
- VADTree: Explainable Training-Free Video Anomaly Detection via Hierarchical Granularity-Aware Tree (NeurIPS 2025 · similarity 0.88)
- Open-Vocabulary Video Anomaly Detection (CVPR 2024 · similarity 0.88)
- Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · similarity 0.88)
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.
