Research map / Label-efficient and robust learning / OOD and anomaly detection
DART: Domain-Agnostic Residual Transfer for Generalist Anomaly Detection
Muhammad Aqeel, Maham Nazir, Marco Cristani, Francesco Setti
Read the paperCheck your idea against 137,217 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 DART, a domain-agnostic residual transfer approach for generalist anomaly detection that separates transferable anomaly cues from domain-specific appearance residuals using normal-abnormal reference samples.
Problem
Generalist anomaly detection must detect anomalies in unseen domains from only a few reference samples; existing normal-abnormal guided residuals mix transferable anomaly cues with domain-specific appearance factors (texture, geometry, acquisition conditions), limiting cross-domain transfer.
Approach
Targets residual domain entanglement in normal-abnormal guided GAD by making residuals domain-agnostic so that anomaly cues transfer across domains rather than carrying domain-specific appearance factors. Full method details are not available from the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DART residual disentanglement to NAGL | NAGL · VisA N^2+A^1 | image-level AUROC: 87.90 -> 89.72 | Improved |
| naively add abnormal reference without disentanglement (+Inv) | ResAD · VisA N^1 | image-level AUROC: -6.5 I-AUROC | Hurt |
| add DART on top of naive abnormal integration | InCTRL+Inv · VisA N^1 | image-level AUROC: 79.87 -> 86.06 | Improved |
| add DART to InCTRL under industrial-to-medical shift | InCTRL · MVTec AD -> BraTS N^1 | image-level AUROC: 65.20 -> 78.48 | Improved |
| add DART to NAGL at N^1 on BraTS | NAGL · MVTec AD -> BraTS N^1 | pixel-level AUROC: 93.00 -> 90.33 | Hurt |
| add sufficiency loss L_suf (ablation) | DART (InCTRL) · MVTec AD k=2 | image-level AUROC: 93.57 -> 96.21 | Improved |
| add adversarial invariance loss L_adv (ablation) | DART (InCTRL) · MVTec AD k=2 | image-level AUROC: 95.57 -> 96.21 | Improved |
| add reconstruction loss L_recon (ablation) | DART (InCTRL) · MVTec AD k=2 | image-level AUROC: 95.48 -> 96.21 | Improved |
| add cross-domain meta-training on top of RDM (ablation) | DART (InCTRL) · MVTec AD k=2 | image-level AUROC: 95.54 -> 96.21 | Improved |
| add CDMT without RDM (ablation) | InCTRL · MVTec AD and VisA k=2 | image-level AUROC: +0.00 | No clear effect |
| use transformer encoder instead of linear projection (ablation) | DART encoder (InCTRL) · VisA k=2 | image-level AUROC: 81.29 -> 88.64 | Improved |
| add LayerNorm in anomaly encoder (ablation) | DART encoder (InCTRL) · VisA k=2 | image-level AUROC: 84.65 -> 88.64 | Improved |
Closest papers by meaning
- Normal-Abnormal Guided Generalist Anomaly Detection (NeurIPS 2025 · similarity 0.91)
- Toward Generalist Anomaly Detection via In-context Residual Learning with Few-shot Sample Prompts (CVPR 2024 · similarity 0.88)
- Beyond Normal References: Discriminative Few-Shot Anomaly Detection (NeurIPS 2026 · similarity 0.87)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.86)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.86)
- Towards Multi-Domain Learning for Generalizable Video Anomaly Detection (NeurIPS 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.
