Reproduce “ReplayCAD: Generative Diffusion Replay for Continual 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
Lei Hu, Zhiyong Gan, Ling Deng, Jinglin Liang, Lingyu Liang, Shuangping Huang, Tianshui Chen
ReplayCAD introduces a diffusion-driven generative replay framework for continual anomaly detection, which mitigates catastrophic forgetting by generating high-quality historical data with preserved pixel-level details. The method achieves an 11.5% improvement in Pixel-AP on the VisA dataset and an 8.1% improvement on MVTec compared to existing state-of-the-art approaches, demonstrating enhanced anomaly segmentation.
Key insights
- Preserving pixel-level details through high-fidelity generative replay is essential for accurate anomaly segmentation in continual learning scenarios.
- The combination of both semantic-aware and spatially-enhanced guidance is critical for generating diverse and high-quality synthetic data for historical classes.
- Utilizing a pre-trained Latent Diffusion Model (LDM) for conditional feature-guided generation offers an efficient mechanism for data replay, minimizing computational overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.