An inspection model can perform well on one component and fail after being adapted to another. A static benchmark sees the two good checkpoints separately and misses the production problem.
InCLAD's October 10 paper organises industrial anomaly detection as a sequence of component-level tasks. It includes different task orders and cross-dataset settings, with separate image-level detection and pixel-level localisation assessments. Read the research paper.
Factory inspection is not a sequence of independent launches
An industrial detector may begin with one component and later be extended to another. The practical requirement is that the operating model remains useful on both. Static benchmarks can show strong performance for separately trained checkpoints without revealing whether adaptation destroys earlier inspection capabilities. Manufacturing change therefore introduces a continual-learning problem that ordinary train-and-test evaluation can miss.
InCLAD organises inspection as task-incremental learning, with each component defining a task and a semi-supervised learning objective. It asks what happens after new components enter the stream, rather than assuming that success on each isolated dataset implies a maintainable inspection system.
Order and domain become controlled variables
The benchmark includes random, easy-to-hard and hard-to-easy sequences, as well as sequences spanning datasets. Task order can affect which representations are established early and which later updates must preserve. Comparing several orders helps reveal a method that succeeds only when its initial tasks happen to supply the right foundation.
The study draws on public industrial datasets with different visual demands, including DAGM's subtle defects against textured backgrounds, MVTec AD's structural anomalies and VisA's surface and structural defects. The benchmark's purpose is not to declare these datasets interchangeable. Their heterogeneity creates a way to test retention and adaptation when component distributions change.
Detection and localisation can regress differently
Image-level detection asks whether an item is anomalous. Pixel-level localisation asks where the anomaly lies. A model may retain the ability to flag an item while losing precision about the defect region, which matters when the output guides inspection or diagnosis. InCLAD evaluates both perspectives rather than compressing continual performance into one classification score.
The paper compares methods on matched task streams and splits, including continual variants, often using replay. Some original anomaly methods require method-specific adaptations because they do not directly support replay updates. Those adaptations are part of what is being tested; a method name alone does not define an identical continual-learning procedure across all detectors.
Continual adaptations must be named alongside the base detector
The experiments compare non-continual baselines with continual variants under identical streams and ordering metadata. Replay is the usual continual mechanism, but PatchCore, PaDiM and CFA need method-specific adaptations because their original formulations do not directly accept the same update process. This means the benchmark compares defined implementations, not abstract detector names detached from how they learn over time.
Compute and memory comparisons must also follow the adaptation. A replay buffer can preserve performance while adding storage, and an efficient static detector may become expensive to update. The paper lists stronger standardised resource budgets as future work. Until those are established, quality results are informative about retention and adaptation but incomplete as a total-cost account for deployment.
A benchmark stream is still an approximation of a factory
The authors explicitly note that sequencing public components and datasets does not reproduce every temporal or operational change in production. Different sensors, evolving quality standards and long-term label drift remain outside much of the benchmark's coverage. Compute and memory budgets also need standardisation before retention gains can be compared as complete operational solutions.
Test all previous components after every addition, retaining both detection and localisation results. Adaptability has to survive the lifetime of the operating inspection system. A collection of excellent isolated detectors does not establish that a single continually updated detector can preserve them. InCLAD makes that distinction measurable and gives industrial AI research a more realistic target than repeated fresh starts on static data.
References
Questions, news tips or corrections? Contact [email protected].
