Research map / Label-efficient and robust learning / OOD and anomaly detection
Towards Multi-Domain Learning for Generalizable 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 multi-domain learning for weakly-supervised video anomaly detection (MDVAD), with a six-dataset benchmark, four protocols, and multi-head Null/NullAng-MIL baselines plus an abnormal conflict classifier to handle conflicting abnormality definitions across datasets.
Problem
Video anomaly detection models trained on one dataset generalize poorly because abnormality definitions conflict across datasets (e.g., a pedestrian on road is normal in UCFC but abnormal in TAD), so naive multi-domain training suffers from abnormal conflicts.
Approach
Use domain-specific abnormal heads where only the head of the input's domain receives gradients (Null-MIL, and an angular-margin NullAng-MIL variant with per-domain normal heads), plus an auxiliary abnormal conflict classifier trained on score variance across heads to make shared features conflict-aware.
Key results
- Held-in (E1) average AUC of 86.86 for Null-MIL and 86.26 for NullAng-MIL+AC, versus 86.52 for in-domain single-domain models averaged.
- Leave-one-out (E2) NullAng-MIL+AC-style MIL(Ours) reaches 75.96 average AUC, outperforming MMIL, ARNet, WSAL and COMO in the WVAD comparison.
- Low-shot adaptation (E3) MIL(Ours) with AC classifier reaches 79.78 average AUC, the best among compared WVAD models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multi-domain Null(Ang)-MIL with AC classifier instead of the WVAD baselines | MDVAD models · MDVAD held-in (E1) | AUC: 86.26 vs 85.85 | Helped |
| use multi-domain Null(Ang)-MIL with AC classifier instead of the WVAD baselines | MDVAD models · MDVAD leave-one-out (E2) | AUC: 75.96 vs 74.92 | Helped |
| use multi-domain Null(Ang)-MIL with AC classifier instead of the WVAD baselines | MDVAD models · MDVAD low-shot adaptation (E3) | AUC: 79.78 vs 76.51 | Helped |
| use multi-domain Null(Ang)-MIL with AC classifier instead of the strongest WVAD baseline | MDVAD models · MDVAD full fine-tuning (E4) | AUC: 87.31 vs 87.70 | Hurt |
| use Null-MIL instead of single-head MIL (ablation) | MDVAD model · MDVAD held-in (E1) | AUC: 85.56 -> 86.86 | Helped |
| use NullAng-MIL instead of single-head MIL (ablation) | MDVAD model · MDVAD held-in (E1) | AUC: 85.56 -> 86.26 | Helped |
| add the AC classifier (ablation) | NullAng-MIL · MDVAD leave-one-out (E2) | AUC: 72.93 -> 75.96 | Helped |
| add the AC classifier (ablation) | NullAng-MIL · MDVAD low-shot adaptation (E3) | AUC: 76.19 -> 79.78 | Helped |
| replace NullAng-MIL with Null-MIL (ablation) | multi-domain VAD model · MDVAD held-in (E1) | AUC: 86.26 -> 86.86 | Helped |
| use NullAng-MIL instead of Null-MIL (ablation) | multi-domain VAD model · MDVAD leave-one-out (E2) | AUC: Null-MIL 55.42–74.65; NullAng-MIL 60.41–82.53 | Helped |
| add the AC classifier (ablation) | MIL baseline · UBN open-set | AUC: 67.95 -> 70.56 | Helped |
| add the AC classifier (ablation) | Null-MIL · UBN open-set | AUC: 72.14 -> 70.94 | Hurt |
Surprising findings
- Single-domain VAD models degrade substantially cross-domain (e.g., TAD->UCFC 50.83 AUC vs 82.32 in-domain), and naive multi-domain MIL can underperform single-domain on some targets.
- Null-MIL is strong in held-in but weak in leave-one-out, while NullAng-MIL is better for unseen targets.
Limitations
- Balanced sampling reduces each dataset to 386 videos, discarding training data from larger datasets such as XD.
- Architecture is not novel; relies on a single I3D backbone and only weakly-supervised MIL-style models.
- Unseen-domain scoring uses max over heads, which the authors note is suboptimal for unknown target domains with Null-MIL.
Experimental setup
Datasets: UCF-Crime XD-Violence LAD UBI-Fights TAD ShanghaiTech Campus CADP NWPU UBnormal Kinetics
Benchmarks: MDVAD benchmark (UCFC, XD, LAD, UBIF, TAD, ST) UBnormal (open-set VAD)
Metrics: AUC AP
Baselines: MIL (single-head) MMIL ARNet WSAL COMO Null-MIL NullAng-MIL
Builds on: RTFM MMIL CoMo Null-class strategy for multi-domain learning
Closest papers by meaning
- Prompt-Enhanced Multiple Instance Learning for Weakly Supervised Video Anomaly Detection (CVPR 2024 · similarity 0.89)
- Open-Vocabulary Video Anomaly Detection (CVPR 2024 · similarity 0.87)
- VADTree: Explainable Training-Free Video Anomaly Detection via Hierarchical Granularity-Aware Tree (NeurIPS 2025 · similarity 0.87)
- Advancing Video Anomaly Detection: A Concise Review and a New Dataset (NeurIPS 2024 · similarity 0.87)
- Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · similarity 0.87)
- Learning Event Completeness for Weakly Supervised Video Anomaly Detection (ICML 2025 · 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.