|
Download knowledge/dinomaly.md from Wizzas/DefectRAG: direct link, hf CLI and curl.
- Browser
- Download file 813 Bytes
-
https://huggingface.co/spaces/Wizzas/DefectRAG/resolve/808959b678752425ec94de718e34584ce1762dbb/knowledge/dinomaly.md
- Command line
-
hf download hf://spaces/Wizzas/DefectRAG@808959b678752425ec94de718e34584ce1762dbb/knowledge/dinomaly.md
-
curl -L -o dinomaly.md https://huggingface.co/spaces/Wizzas/DefectRAG/resolve/808959b678752425ec94de718e34584ce1762dbb/knowledge/dinomaly.md
813 Bytes
| # Reconstruction anomaly detection | |
| Anomalib Dinomaly uses frozen DINOv2 encoder representations and trains a bottleneck and decoder on normal images. Reconstruction differences support anomaly scoring and spatial localization. Its documented backbone interface uses DINOv2 variants. This does not establish an interchangeable DINOv3 adapter. | |
| # Interpretation | |
| A feature reconstruction discrepancy indicates departure from learned normal appearance. Evaluation on labelled images is required to measure its usefulness. A spatial anomaly map alone does not establish a named physical defect or a manufacturing root cause. DefectRAG's compact nearest-normal fallback uses cosine patch distances rather than Dinomaly's reconstruction decoder, so published Dinomaly benchmark numbers do not describe this application. | |