Project context
The team project is authored by Benjamin Förster, Ali Abul Hawa, Mahboubeh Nadaf, and Karim Khalifa. It progressed from EDA and prototype experiments toward a reproducible anomaly-detection framework with two principal evaluated model families: a Keras convolutional autoencoder and PatchCore with a ResNet-18 backbone.
Role in the project
My work included exploratory analysis and reporting, metric and threshold reasoning for highly imbalanced anomaly data, autoencoder experiments, ResNet representation work, PatchCore implementation and evaluation, and subsequent DINOv2 experiments using patch-level representations and nearest-neighbour anomaly scoring.
The later shared repository contains integration and engineering by the wider team, so this page separates that shared system from my own development path.
The data changes the evaluation problem
The dataset contains 3,629 normal training images, 467 normal test images, and 1,258 defective test images across 15 categories. Defects are often tiny: the median defective region occupies only 1.54% of an image, one quarter occupy less than 0.49%, and some are as small as 0.037%.
That makes pixel accuracy misleading. A detector can classify almost everything as normal and still report an apparently strong accuracy while finding no useful defects.
PatchCore, simplified
PatchCore builds a compact reference of normal local image features. At inference time, patches from a new image are compared with that normal memory. Larger distances indicate regions that differ more strongly from the normal reference.
Same flow in compact form
Qualitative PatchCore results
These are real outputs from evaluated PatchCore runs across four MVTec AD categories. Each result shows the input, ground-truth mask, continuous anomaly-map overlay, and the thresholded prediction.
Selected from regenerated PatchCore outputs used in the project evaluation. Open any result for a full-size view.
Evaluation choices
Pixel AUROC can remain high because normal pixels dominate the denominator of the false-positive rate. The project therefore uses AUPIMO as the main localization metric and F1 plus precision-recall behavior at image level, with threshold sensitivity treated explicitly.
The EDA also showed strong category heterogeneity. A Kruskal-Wallis test across defect-area distributions gave H = 626.5 with p = 1.2 × 10-124, while anomalous-pixel ratios ranged from roughly 0.34% for screw to 14.5% for metal nut.
Evaluated PatchCore baseline
Across the completed 15-category baseline, localization varied substantially by category: AUPIMO reached 0.982 for bottle and 0.972 for leather, while zipper fell to 0.251. Screw was the hardest image-level category at F1 0.796.
The weaker categories direct attention to representation scale, preprocessing, coreset construction, smoothing, threshold calibration, and evaluation configuration.
DINOv2 continuation
After the PatchCore path, the next branch explored DINOv2 patch representations with the same broad idea of learning a normal reference and scoring test patches by distance. That work is maintained separately on the shared repository's feature/dinov2-baseline branch.
The headline metrics above are deliberately limited to the completed PatchCore baseline. DINOv2 is shown here as a later experimental continuation rather than mixed into the evaluated baseline without a directly comparable completed result.
Integrated project pipeline
The shared repository moves the work beyond isolated notebooks into a reproducible framework with Pixi environments, task automation, tests and linting, FastAPI and Streamlit interfaces, model caching, documentation, category-adaptive hyperparameter sweeps, and explicit evaluation workflows.