padim

Pipeline training and inference Anomalib models UI in Anomaly Detection

Production-ready visual anomaly detection from normal images, featuring PaDiM, ultra-light PatchCore, automated edge deployment, and optimized ONNX, OpenVINO, TensorRT INT8, Hailo, and KV260 support.

Training-free deep learning for unsupervised industrial defect detection & localization on MVTec AD, PatchCore & PaDiM, with image and pixel level AUROC and a one-click free-Colab demo.

兼容 Anomalib 生态的通用异常检测推理库：一张图进去，缺陷 mask + 分数 + OK/NG 出来，一个 DLL + 一个头文件，零 OpenCV 依赖。

PaDiM-inspired anomaly detection for AWS infrastructure metrics (CloudWatch + Cost Explorer)

Anomaly detection pipeline for industrial surface inspection using the MVTec AD dataset. Includes classical computer vision baselines, autoencoder models, and PaDiM-based feature embeddings for detecting subtle manufacturing defects.

A PyTorch-based framework for unsupervised anomaly detection on the MVTec AD dataset. Implements PaDiM, PatchCore, FastFlow, and AutoEncoder models with utilities for dataset preparation, training, evaluation, and reporting.

detect and segment anomalies on wood surfaces using unsupervised deep learning models

Governed, reproducible visual-inspection runtime using PaDiM, ONNX Runtime, deterministic replay, and hash-anchored evidence.

Edge AI visual-inspection reference implementation: real anomaly detection, automatic rollback, air-gapped updates -- measured, not assumed.

Mahalanobis scoring stage of PaDiM anomaly detection, implemented four ways - PyTorch, C++17 CPU, hand-written CUDA (naive + PBCR), cuBLAS - behind a fixed artifact boundary, with a 1e-4 numerical agreement gate on every reported timing.

MVTec AD 上的非監督式表面瑕疵檢測（AOI）：PatchCore 與 PaDiM 在純 CPU 上跑三個類別的對照 benchmark，附解析度取捨表，並以漏檢／誤殺的成本分析回答閾值該怎麼定——零漏檢政策在各類別實際要付多少代價。

Anomaly Detection with Localization developed for the Machine Learning for Vision and Multimedia course @polito. Tech stack: PyTorch, Python.

C++ real-time anomaly detection: motion + scene-change + PaDiM, ROI-aware, severity-laddered alerts, TensorRT for the edge

Evidence-first industrial visual anomaly detection and pixel localization across all 15 MVTec AD categories.