Repository navigation
Proposal: Explainability utilities for visual anomaly detection #3631
Replies: 1 comment 1 reply
|
Hello! Documentation page explaining anomaly-map interpretation.
Phase 2: Anomaly Map Evaluation -> you should check out the already existing metrics and methods in anomalib for evaluation Phase 3: Optional Explainability API -> as the models dont predict learned anomalies because they are only trained on normal images i'm not sure how you could use gradient based methods for explanation. however i'm not completely into gradient flows in reconstruction based methods so maybe you can elaborate a bit further. maybe you can also explain a certain use case for your XAI methods that go beyond the current anomaly maps. |
Uh oh!
There was an error while loading. Please reload this page.
Proposal: Explainability Utilities and Evaluation for Visual Anomaly Detection
Hi Anomalib team,
First, thank you for maintaining Anomalib. While exploring the library and contributing documentation improvements, I started thinking about a possible direction that could complement the existing anomaly map generation, visualization, and localization metrics.
As anomaly detection models become increasingly deployed in industrial and safety-critical environments, users often ask not only:
but also:
Anomalib already provides anomaly maps and localization metrics, which are important building blocks for interpretability. I would like to discuss whether additional explainability-oriented utilities and evaluation tools would fit within the scope of the project.
Motivation
Many anomaly detection methods, such as PatchCore, EfficientAD, DFM, and reconstruction-based approaches, generate anomaly maps that are often interpreted as explanations.
However, users currently have limited tooling to:
Potential Roadmap
Phase 1: Interpretation and Visualization
Provide examples and utilities focused on understanding anomaly maps.
Possible additions:
This phase would mainly focus on education and usability.
Phase 2: Anomaly Map Evaluation
Introduce metrics that evaluate anomaly maps beyond localization performance.
Potential directions:
The goal would be to assess whether anomaly maps are actually informative and faithful to model behavior.
Phase 3: Optional Explainability API
Depending on community interest, a lightweight model-agnostic explainability interface could be explored.
Potential examples:
The objective would be to complement existing anomaly maps rather than replace them.
Related Work
My interest comes from research in explainable anomaly detection, including work on reconstruction-based visual anomaly detection and explanation evaluation.
Example repository:
https://github.com/rashidrao-pk/anomaly_detection_trust_case_study
Before investing effort into a prototype implementation, I would be interested in hearing whether the community sees value in this direction and which parts would best fit Anomalib's long-term roadmap.
Thank you for your feedback.
All reactions