RECORDED DEMO
A walkthrough of the working inspection application: inspect a bottle, review its anomaly score, and explore where the model found unusual visual features.
Recorded from genuine local inference on October 3, 2026. This page plays a recording and does not run online inference.
The application inspects a damaged rim, a normal reference bottle, and a contamination sample. Each inspection runs the actual fitted model. The walkthrough switches between the original image, candidate hotspot boxes, and a relative heatmap.
An anomaly score measures feature distance; it is not a probability. The threshold creates a flagged or within-threshold decision. Heatmaps show relative anomaly intensity, and hotspot boxes outline candidate regions rather than detected object classes.
The interface connects to a fitted PyTorch model and FAISS nearest-neighbor scoring. The displayed threshold comes from a limited normal-image calibration pilot. These three samples demonstrate the application and do not establish accuracy or a production operating point.
The application repository implements the inspection workflow, inference integration, visualization, input handling, exports, and calibration investigation. The fitted checkpoint and its pretrained feature extractor are credited separately.
Sample images: MVTec AD, MVTec Software GmbH. Recording and screenshot: CC BY-NC-SA 4.0, for noncommercial use. Full attribution · MIT software license