Industrial visual anomaly detection has seen rapid progress in recent years, but the way systems are evaluated may not match how they are actually used. According to a new preprint on arXiv, most existing benchmarks are static: they assume a fixed set of components and defect types. In practice, however, inspection systems often need to be extended to new parts as production lines change.
The paper introduces InCLAD, a continual learning benchmark specifically for this setting. Rather than testing a model once on a frozen dataset, InCLAD is designed to measure how well an anomaly detection system can learn new industrial components without forgetting previously learned ones. The authors argue that this better represents deployment scenarios where factories add or swap components over time.
Because this is a single preprint, the claims have not yet been peer-reviewed or independently validated. The abstract does not include experimental results or comparisons with existing benchmarks, so the main contribution at this stage is the framing and proposed benchmark design. Still, the work points to an important limitation in current evaluation practices for industrial AI inspection systems.