AD Robotics Hazards

--- license: mit language: - en tags: - anomaly-detection - computer-vision - vae-gan - one-class-classification - robotics - industrial-ai - visual-anomaly-detection - pytorch library_name: pytorch pipeline_tag: image-classification datasets: - hazards-robots-corridor --- # VAE-GAN for Corridor Hazard Anomaly Detection <img src='docs/Dataset.svg'> This repository contains trained **VAE-GAN model checkpoints** for visual anomaly detection in robotic environments. The models were trained using a one-class learning paradigm, where only normal operating conditions are observed during training and anomalous situations are detected during inference through reconstruction-based anomaly scoring. ## Dataset The models were trained on the **Corridor** scenario from the *Hazards&Robots* dataset. ### Original Dataset **Hazards&Robots: A Dataset for Visual Anomaly Detection in Robotics** Authors: - Dario Mantegazza - Alind Xhyra - Luca M. Gambardella - Alessandro Giusti - Jérôme Guzzi Original resources: - Zenodo: https://zenodo.org/records/7859211 - GitHub: https://github.com/idsia-robotics/hazard-detection - IDSIA Robotics Lab: https://idsia-robotics.github.io/ ### Preprocessed Dataset A preprocessed version of the Corridor dataset prepared for one-class anomaly detection experiments is available on Kaggle: https://www.kaggle.com/datasets/rashidrao/robotics-hazards The dataset has been reorganized into train, validation, and test splits following the standard anomaly detection protocol: - Training set: normal samples only - Validation set: normal samples for threshold calibration - Test set: normal and anomalous samples All credit for the original data belongs to the Hazards&Robots authors. ## Model Architecture The checkpoint is based on a VAE-GAN architecture consisting of: - Encoder network - Variational latent representation - Decoder / Generator - Adversarial discriminator The model is trained to learn the distribution of normal robotic corridor scenes and identify anomalies through reconstruction discrepancies. ## Source Code The complete training and evaluation framework used to produce these checkpoints is available at: https://github.com/rashidrao-pk/AD_MultiPointThreshold The repository includes: - VAE-GAN training pipeline - Multi-point thresholding framework - Threshold calibration utilities - Reconstruction-based anomaly scori...