Research map / Label-efficient and robust learning / OOD and anomaly detection
Debiasing Trace Guidance: Top-down Trace Distillation and Bottom-up Velocity Alignment for Unsupervised Anomaly Detection
Xingjian Wang, Li Chai, Jiming Chen
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 |
|---|---|---|---|
| add trace distillation and velocity alignment to baseline InstaFlow finetuning (ablation) | InstaFlow baseline · MVTec-AD (multi-class) | AUROCcls / AUROCseg: 97.8 / 97.6 -> 99.0 / 98.4 | Improved |
| add learnable prompting (ablation) | InstaFlow baseline · MVTec-AD (multi-class) | AUROCcls / AUROCseg: 97.8 / 97.6 -> 98.2 / 97.8 | Improved |
| replace average pooling with learned wavelet trace distillation (ablation) | DTG · MVTec-AD (multi-class) | AUROCcls / AUROCseg: 98.5 / 98.1 -> 99.0 / 98.4 | Improved |
| add average pooling low-dimension guidance (ablation) | DTG · MVTec-AD (multi-class) | AUROCcls: 0.3 increment in AUROCcls | Improved |
| add LRec and LCoe constraints to the distillation filter (ablation) | DTG · MVTec-AD (multi-class) | AUROCcls / AUROCseg: 97.1 / 96.8 -> 99.0 / 98.4 | Improved |
| decomposition level l=2 instead of l=1 (ablation) | DTG · MVTec-AD (multi-class) | AUROCcls / AUROCseg: 98.8 / 98.4 -> 99.0 / 98.4 | No clear effect |
| decomposition level l=2 instead of l=4 (ablation) | DTG · MVTec-AD (multi-class) | AUROCcls / AUROCseg: 98.5 / 98.2 -> 99.0 / 98.4 | Improved |
| mask ratio 0.4 instead of 0.2 (ablation) | DTG · MVTec-AD (multi-class) | AUROCcls / AUROCseg: 98.8 / 98.3 -> 99.0 / 98.4 | Improved |
| mask ratio 0.4 instead of 0.6 (ablation) | DTG · MVTec-AD (multi-class) | AUROCcls / AUROCseg: 98.3 / 97.9 -> 99.0 / 98.4 | Improved |
| DTG vs prior best DiAD | DTG · MVTec-AD (multi-class) | AUROCcls / AUROCseg: 0.3 increment in AUROCcls and 0.1 increment in AUROCseg | Improved |
| DTG vs VPOT on VisA | DTG · VisA | AUROCcls: 2.4 increment in AUROCcls | Improved |
| DTG vs SpoofTrace on CeFA | DTG · CASIA-SURF CeFA | ACER: 52.0%/54.8% in ACER for Protocol 1/3 | Improved |
Closest papers by meaning
- InvAD: Inversion-based Reconstruction-Free Anomaly Detection with Diffusion Models (CVPR 2026 · similarity 0.86)
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.84)
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.84)
- Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.83)
- DZAD: Diffusion-based Zero-shot Anomaly Detection (AAAI 2025 · similarity 0.83)
- Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection (AAAI 2024 · similarity 0.83)
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.
