Mask Architecture Anomaly Segmentation for Road Scenes [Course Project]

This repository provides a starter-code setup for the Real-Time Anomaly Segmentation project of the Machine Learning Course. It consists of the code base for training/testing ERFNet on the Cityscapes dataset and perform anomaly segmentation. It also contains some code referring to EoMT.

---

Synthetic Dataset Generation

This repository includes code for synthetic anomaly dataset generation. The generation pipeline and related scripts are available at:

Link: https://github.com/Giacomo-FMJ/MaskArchitectureAnomaly/tree/main/cityscapes_coco_anomaly

---

Available Branches

This repository contains multiple branches, each implementing different anomaly detection approaches:

Main Extensions (Proposed Architectures)

1. main - Current branch containing the base project structure and ERFNet baseline
2. fine-tuning_anomaly_head_revised - First proposed extension architecture implementing a Statistical Uncertainty-based Pixel-wise Anomaly Head (stats-only approach)
3. fine-tuning_anomaly_head_revised_features - Second proposed extension architecture implementing a Hybrid Pixel-wise Anomaly Head (statistical uncertainty + visual features from backbone)

Additional Experimental Branches (Less Relevant)

4. fine-tuning_anomaly_scores - Alternative baseline approach with anomaly scoring (not extensively covered in the report)
5. fine-tuning_only_queries - Query-based anomaly detection baseline (not extensively covered in the report)

Note: The core contributions and detailed documentation are available in branches 2 and 3.

---

Evaluation & Inference

• eomt/inference -COLAB.ipynb - Notebook for evaluating the baseline EoMT model on anomaly detection datasets. Includes inference pipeline with traditional uncertainty-based methods (MSP, MaxLogit, Entropy) and visualization tools.

---

Packages

For instructions, please refer to the README in each folder:

• eval contains tools for evaluating/visualizing the an ERFNet model's output and performing anomaly segmentation.
• trained_models Contains the ERFNet trained models for the baseline eval.
• eomt It is almost the original folder of the EoMT project. Inside it you will find code to train and pretrained checkpoints for EoMT.