Research map / Video / Video-language understanding

Uncovering What Why and How: A Comprehensive Benchmark for Causation Understanding of Video Anomaly

CVPR 2024 · Video › Video-language understanding · approach: Event and anomaly reasoning in video · 26 citations (OpenAlex)

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.

Introduces CUVA, a 1,000-video benchmark for "what/why/how" causation understanding of video anomalies, plus the MMEval LLM-based metric and the A-Guardian prompt-based baseline built on Video-ChatGPT.

Problem

Existing video anomaly understanding benchmarks focus on detection and localization and lack cause/effect explanations, suitable text-based evaluation metrics for multimodal outputs, and long videos.

Approach

A benchmark with human free-text annotations of anomaly type, moment, cause, and effect, plus an importance curve for severity; an MMEval metric built on a video LLM that scores, ranks, and explains answers; and A-Guardian, which combines multi-turn hard prompts with a MIST-style soft-prompt segment selector.

Key results

• A-Guardian on Video-ChatGPT raises MMEval cause score from 44.57 to 58.92 over the hard-prompt-free baseline, with the full method achieving 79.65 on description.
• On the human consistency test, MMEval agreed with human rankings at 89.1% for effect, 80.2% for cause, and 82.3% for description answer pools, above BLEU, ROUGE, BLEURT, MoverScore and UniEval.
• Timestamp localization was the hardest task; the best timestamp IOU among the tested VLMs was 9.0% (mPLUG-Owl).

Measured effects reported in the paper

Surprising findings

• Traditional NLG metrics such as BLEU and ROUGE scored similarly on answers with very different meanings, so they fail to separate reasoning quality.

Limitations

• Timestamp localization and fine-grained temporal reasoning remain very weak for current VLMs.
• MMEval relies on a video LLM as judge, whose reliability is validated only with a single human consistency study.

Experimental setup

Datasets: CUVA

Benchmarks: CUVA

Metrics: BLEU ROUGE BLEURT MoverScore UniEval MMEval accuracy IOU human consistency agreement

Baselines: mPLUG-Owl Video-LLaMA PandaGPT Otter Video-ChatGPT BLEU ROUGE BLEURT MoverScore UniEval

Builds on: Video-ChatGPT MIST LLaVA CLIP

Closest papers by meaning

• Anomize: Better Open Vocabulary Video Anomaly Detection (CVPR 2025 · similarity 0.86)
• A Unified Reasoning Framework for Holistic Zero-Shot Video Anomaly Analysis (NeurIPS 2025 · similarity 0.85)
• VADTree: Explainable Training-Free Video Anomaly Detection via Hierarchical Granularity-Aware Tree (NeurIPS 2025 · similarity 0.83)
• Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity (CVPR 2025 · similarity 0.83)
• Open-Vocabulary Video Anomaly Detection (CVPR 2024 · similarity 0.83)
• Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought (NeurIPS 2025 · similarity 0.82)