Reproduce “Towards Fine-Grained Vision-Language Alignment for Few-Shot Anomaly Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuanting Fan, Jun Liu, Xiaochen Chen, Bin-Bin Gao, Jian Li, Yong Liu, Jinlong Peng, Chengjie Wang
The FineGrainedAD framework from Tencent Youtu Lab improves few-shot anomaly detection by introducing Multi-Level Fine-Grained Semantic Captions (MFSC) and a fine-grained vision-language alignment strategy. This approach achieves 96.71% pixel-level AUROC on MVTec-AD and 97.54% on VisA in 4-shot settings, providing more precise anomaly localization with high inference efficiency.
Key insights
- Automatically generated multi-level fine-grained semantic captions effectively bridge the gap between high-level language understanding and precise, component-level visual anomaly localization.
- Explicit alignment between multi-level learnable prompts and corresponding visual regions, guided by enhanced clustering and tailored loss functions, significantly improves anomaly perception and localization accuracy.
- A lightweight framework built on a frozen VLM backbone, with optimized prompt learning and alignment, can achieve state-of-the-art few-shot anomaly detection performance without requiring auxiliary data or heavy external models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.