The solutions for the dacon competition (1st place).

Implementation of anomaly detection detection on MVTec 3D-AD dataset using point-voxel diffusion.

Toolbox for Unsupervised Anomaly Detection on MVTec AD

A Curated List of Awesome Unsupervised Anomaly Detection on MVTec AD Dataset

Context In the context of textile fabric, rare anomaly can occurs, hence compromising the quality of the tissues. In order to avoid that in some scenario, it is crucial to detect the defect. This dataset is for educational purposes Content Image size: 32x32 or 64x64 classes: ['good', 'color', 'cut', 'hole', 'thread', 'metal contamination'] rotations: 8 different rotations in [0, 20, 40, 60, 80, 100, 120, 140] Given an image size, a train and test dataset are available with randomly generated patches. Source images from the train and test are non-overlapping different tasks are possible: classification of the classes type classification of angles using only "good" images and testing of other classes texture representation learning / self-supervised learning Acknowledgements Based on the public dataset by the MVTec company Paul Bergmann, Michael Fuser, David Sattlegger, Carsten Steger. MVTec AD - A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection; in: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019 Inspiration the main goal of this dataset is to explore self-supervised learning on texture images in order to solve anomaly detection problems and learn a robust representation of texture in lieu of traditional image processing features (e.g. glcm, gabor,….)

Using SPADE for anomaly detection on MVTEC LOCO AD dataset

Zero and few-shot industrial image anomaly detection framework comparing AnomalyDINO & MuSc models across MVTec AD, BTAD, and ViSA datasets with MLflow tracking and flexible configuration.

Comprehensive benchmark for anomaly detection models on MVTec AD dataset using Anomalib. Includes PatchCore, EfficientAD, FastFlow, STFPM, PaDiM with Gradio demo.

No description available

Anomaly detection on MVTec AD using VQ-VAE-2

Zero-Shot Anomaly Detection and Segmentation using DINOv3 without additional training or fine-tuning

Goal: Apply pre-trained foundation models (like DINOv3, Mirroring DINO, or SAM) to detect surface defects and irregular textures in industrial images — specifically using the MVTec Anomaly Detection (AD) dataset).

VAE-based visual anomaly detection system built on the MVTec AD (Hazelnut) dataset. Trained using only normal images, the model detects defects via reconstruction error. Includes image-level anomaly scoring, pixel-level defect localization with heatmaps, and statistically derived thresholds for unsupervised industrial inspection tasks.

Anomaly detection on MVTec AD dataset

Anomaly Detection on MVTecAD with PatchCore

Unsupervised Industrial Anomaly Detection using MVTec AD

No description available

Unsupervised Anomaly Detection and Localization using Convolutional Autoencoder (CAE) on MVTec LOCO AD Dataset.

Hybrid Anomaly Detection with Dimensional Constraints (HAD-DC) detects both visual and dimensional anomalies in MVTec AD metal nuts. It combines an autoencoder-based visual detector with camera calibration–based dimensional checks, producing a hybrid anomaly score for robust quality inspection.

unsupervised anomaly detection on the mvtec ad dataset using vision transformer embeddings (dino/mae). includes feature extraction, knn/mahalanobis scoring, heatmaps, evaluation metrics (auroc/pro), and a streamlit demo for interactive visualization.

Implementation and analysis of PatchCore and DRAEM for industrial anomaly detection. Includes code, evaluation results, and visualizations for MVTec AD and Severstal Steel datasets. See notebooks and report for details and reproducibility.

An AI-driven industrial anomaly detection system. Utilizes a fine-tuned YOLOv8 model trained on the MVTec AD dataset to inspect surfaces, sending real-time stop/start signals directly to Siemens S7 PLCs via Snap7.

Anomaly detection system for industrial quality inspection using DINOv2 Vision Transformers for feature extraction and PatchCore with K-NN for scoring. Augments training data with Stable Diffusion-generated synthetic defects. Deployed for aerospace and manufacturing QA, achieving robust defect detection on MVTec AD benchmark dataset.

Real-time unsupervised anomaly detection for industrial QA using CAE/VAE on MVTec AD. Trained on defect-free images, evaluated via Image AUROC & Pixel AUPRO, with smoothed L1 error for robust localization. Targets <100ms inference and PLC-integrable Pass/Fail signals.

InspAIct is a Streamlit-powered anomaly detection app that uses the PatchCore algorithm to identify manufacturing defects in bottles from the MVTec AD dataset. It detects issues like cracks, contamination, and broken seals — with heatmap visualizations to make defects easy to spot.

This project uses a convolutional autoencoder to detect surface defects in hazelnuts from the MVTec AD dataset. The model learns to reconstruct normal (non-defective) hazelnut images, and any reconstruction error helps identify defective regions. It's a lightweight, unsupervised approach for visual anomaly detection and segmentation.

Unsupervised Industrial Anomaly Detection using MVTec AD

No description available

Unsupervised Industrial Anomaly Detection using MVTec AD

Unsupervised Industrial Anomaly Detection using MVTec AD