TinyGLASS: Real-Time Self-Supervised In-Sensor Anomaly Detection**

Pietro Bonazzi, Rafael Sutter, Luigi Capogrosso, Mischa Buob, Michele Magno

ArXiv Preprint Link & Dataset Link

Table of Contents

• 📖 Introduction
• 🔧 Environments
• 📊 Data Preparation
• 🚀 Run Experiments
• 📂 Dataset Release
• 🔗 Citation
• 🙏 Acknowledgements
• 📜 License

Introduction

This repository contains source code for TinyGLASS implemented with PyTorch. TinyGLASS, a lightweight adaptation of the GLASS framework designed for real-time in-sensor anomaly detection on the Sony IMX500.

Environments

Create a new conda environment and install required packages.

conda create -n glass_env python=3.9.15
conda activate glass_env
pip install -r requirements.txt

Experiments are conducted on 3x NVIDIA A6000 (48GB). Same GPU and package version are recommended.

Data Preparation

The public datasets employed in the paper are listed below.

• MMS (Download link)
• MVTec AD (Download link)

Other valuable datasets:

• VisA (Download link)
• MPDD (Download link)

Run Experiments

To reproduce the results in the paper please run :

bash run_all.sh

Dataset Release

1.MMS Dataset (Download link)

The MMS Dataset comprises four defect categories for M&Ms candies, i.e., crack-hole, scratch, half, and normal, covering structural and surface-level anomalies. It was collected using an high-resolution microscope camera (mms_stretch) and the IMX500 camera (mms_rpi)

Citation

Please cite the following paper if the code and dataset help your project:

@misc{bonazzi2026tinyglassrealtimeselfsupervisedinsensor,
      title={TinyGLASS: Real-Time Self-Supervised In-Sensor Anomaly Detection}, 
      author={Pietro Bonazzi and Rafael Sutter and Luigi Capogrosso and Mischa Buob and Michele Magno},
      year={2026},
      eprint={2603.16451},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.16451}, 
}

Acknowledgements

Thanks for the great inspiration from GLASS.

License

The code and dataset in this repository are licensed under the MIT license.