Research map / Label-efficient and robust learning / OOD and anomaly detection
SeaS: Few-shot Industrial Anomaly Image Generation with Separation and Sharing Fine-tuning
Zhewei Dai, Shilei Zeng, Haotian Liu, Xurui Li, Feng Xue, Yu Zhou
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 |
|---|---|---|---|
| use unbalanced abnormal text prompt with learnable tokens instead of predefined typical prompt (ablation) | SeaS · MVTec AD (BiSeNet V2) | pixel-level IoU: 50.46 -> 55.28 | Improved |
| mix different anomaly types of a product during fine-tuning (ablation) | SeaS · MVTec AD (BiSeNet V2) | pixel-level IoU: 53.11 -> 55.28 | Improved |
| add NA loss (normal alignment) (ablation) | SeaS · MVTec AD (BiSeNet V2) | IS: 1.67 -> 1.88 | Improved |
| add second term of DA loss (ablation) | SeaS · MVTec AD (BiSeNet V2) | pixel-level IoU: 54.99 -> 55.28 | Improved |
| add Mask Refinement Module with progressive refinement alongside coarse mask supervision (ablation) | SeaS RMP branch · MVTec AD (BiSeNet V2) | pixel-level AP: 65.28 -> 69.21 | Improved |
| add coarse mask supervision to MRM-based refinement (ablation) | SeaS RMP branch · MVTec AD (BiSeNet V2) | pixel-level AP: 62.04 -> 69.21 | Improved |
| train one unified segmentation model instead of per-product models (ablation) | BiSeNet V2 · MVTec AD | pixel-level IoU: 54.11 -> 55.28 | Improved |
| use 4 learnable defect tokens and one object token instead of 1 and 1 (ablation) | SeaS prompt · MVTec AD (BiSeNet V2) | pixel-level AP: 63.69 -> 69.21 | Improved |
| sample abnormal and normal images together instead of sequentially (Abnormal-Normal) (ablation) | SeaS fine-tuning · MVTec AD | IS: 1.53 -> 1.88 | Improved |
| use VAE decoder features instead of VAE encoder features for mask refinement (ablation) | SeaS RMP · MVTec AD (BiSeNet V2) | pixel-level AP: 66.26 -> 69.21 | Improved |
| add normal image supervision for mask prediction (ablation) | SeaS RMP · MVTec AD (BiSeNet V2) | pixel-level IoU: 53.97 -> 55.28 | Improved |
| use cross-attention maps l=2,3 instead of l=2,3,4 (ablation) | SeaS DA loss · MVTec AD (BiSeNet V2) | pixel-level IoU: 52.46 -> 55.28 | Improved |
Closest papers by meaning
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.90)
- Text-Guided Variational Image Generation for Industrial Anomaly Detection and Segmentation (CVPR 2024 · similarity 0.89)
- AnomalyDiffusion: Few-Shot Anomaly Image Generation with Diffusion Model (AAAI 2024 · similarity 0.89)
- Dual-Interrelated Diffusion Model for Few-Shot Anomaly Image Generation (CVPR 2025 · similarity 0.89)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.89)
- Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection (CVPR 2025 · 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.
