MaskArchitectureAnomaly_CourseProject

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

MaskArchitectureAnomaly_CourseProject

★ 10 Open GitHub ↗

No description available.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Mask Architecture Anomaly Segmentation for Road Scenes

Course project for **Fundamentals of Artificial Intelligence, Machine and Deep
Learning** at **Politecnico di Torino** — Group 49.

We compare a pixel-based CNN (**ERFNet**) and a mask-classification transformer
(**EoMT**) on
two complementary problems for road scenes:

1. **Closed-set semantic segmentation** on Cityscapes, studying how the
   pretraining domain (Cityscapes vs. COCO-panoptic) affects performance and
   how much lightweight fine-tuning recovers.
2. **Anomaly segmentation** on five benchmarks (SMIYC RA-21 / RO-21,
   Fishyscapes Lost&Found / Static, Road Anomaly) using training-free post-hoc
   scores: MSP, MaxLogit, MaxEntropy and RbA, plus a temperature-scaling sweep.

## Key results

| Model | mIoU (Cityscapes val) |
|---|---|
| EoMT – Cityscapes (provided) | **0.817** |
| EoMT – COCO (no fine-tuning) | 0.546 |
| EoMT – Finetuned (ours, 5 epochs) | 0.771 |

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

* [eval](eval) contains tools for evaluating/visualizing an ERFNet model's output and performing anomaly segmentation.
* [trained_models](trained_models) Contains the ERFNet trained models for the baseline eval. 
* [eomt](eomt) It is almost the original folder of the EoMT project. Inside it you will find code to train and pretrained checkpoints for EoMT.
* [notebooks](notebooks) Our project work, one notebook per task: **Task4** (EoMT-Cityscapes vs. EoMT-COCO, qualitative comparison + mIoU with COCO→Cityscapes remapping), **Task5** (fine-tuning EoMT-COCO on Cityscapes), **Task7** (ERFNet anomaly-segmentation baselines), **Task8** (EoMT anomaly baselines, RbA, temperature scaling).
* [results](results) Saved outputs per task: mIoU summaries, per-class IoU tables, anomaly AuPRC/FPR95 results (JSON/CSV) and sample predictions.

## Running the notebooks

The notebooks are written for **Google Colab** (GPU runtime). Each one mounts
Google Drive (for the Cityscapes data, checkpoints and results), clones this
repository, and installs dependencies from `eomt/requirements.txt`. Update the
paths in the first cells to match your own Drive layout.

MaskArchitectureAnomaly_CourseProject

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README