Metric Learning Application Development Resource

This repository is a work in progress, currently focusing on Deep Mmetric Learning and contains notebooks for:

• Anomaly detection application example for MVTec Anomaly Detection dataset[1], and related small utilities.
• Benchmarks of various deep metric learnings by Anomaly Detection problem setting with MNIST/CIFAR-10 datasets.
• Visualizations such as grad-CAM, failure cases of the benchmarks, and so on.

Possible final goals:

• Test/benchmark resources summary for metric learning methods as well as various applications.
• Utilities and/or library for metric learning applications.
• Links for resources.

1. How to run examples

Examples in this repository depend on some external modules/source codes. Follow steps below to install them first.

Then just running each notebook will show you the results.

1.1 Install dependent modules first

• Install dl-cliche. dl-cliche is a general purpose utility module, my repo is always depending on it...

pip install dl-cliche

• Install fast.ai and other dependent modules.

pip install fastai
 :

1.2 Download external source codes

wget https://raw.githubusercontent.com/ronghuaiyang/arcface-pytorch/master/models/metrics.py
wget https://raw.githubusercontent.com/KaiyangZhou/pytorch-center-loss/master/center_loss.py

2. Examples

• MVTecAD contains anomaly detection application examples for MVTec Anomaly Detection dataset[1].
• MNIST contains benchmark examples for MNIST dataset.
• CIFAR-10 contains benchmark examples for CIFAR-10 dataset.

3. Published documents

• MVTec AD - Medium: Spotting Defects! - Deep Metric Learning Solution For MVTec Anomaly Detection Dataset
• MVTec AD - (Japanese) Qiita: 欠陥発見! MVTec異常検知データセットへの深層距離学習(Deep Metric Learning)応用
• CIFAR-10/MNIST - (Japanese) Qiita: 深層距離学習(Deep Metric Learning)各手法の定量評価 (MNIST/CIFAR10・異常検知)

4. Acknowledgements & references

Many thanks to the people who worked for MVTec AD dataset, it would greatly help future researches. And many thanks to fast.ai library[2] for minimizing time to develop experiments.

• [1] Paul Bergmann, Michael Fauser, 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. https://www.mvtec.com/fileadmin/Redaktion/mvtec.com/company/research/mvtec_ad.pdf
• [2] Jeremy Howard et al. (2018). The fast.ai deep learning library, https://github.com/fastai/fastai.