README

1 Overview

This assignment demonstrates two segmentation approaches described in the accompanying paper:

1. Triple-scale model training – three networks (Big, Medium, Small) trained jointly.
2. Multi-Head model – a single backbone with three parallel heads (Big, Medium, Small).

Everything you need to reproduce the experiments is already in the repository; just make sure the Cityscapes dataset is in the expected folders.

---

2 Repository layout

Path	Purpose
final_assignment/train.py	Launches triple-scale training.
final_assignment/models/	Pre-defined model architectures: big.pth, medium.pth, small.pth.
final_assignment/Multiheadmodel/	Code for the Multi-Head model.
final_assignment/data/cityscapes/gtFine/	Cityscapes ground-truth masks (download separately).
final_assignment/data/cityscapes/leftImg8bit/	Cityscapes RGB images (download separately).
dice_loss.py	Shared Dice-loss implementation used by all models.

---

3 Running the experiments

3.1 Triple-scale training

python final_assignment/train.py

### 3.2 Multi-Head model

python final_assignment/Multiheadmodel/train_multihead.py
Builds one backbone and attaches the three heads (also loaded from models/).

## 4 Prerequisites
Cityscapes dataset – place it exactly like this:

final_assignment/data/cityscapes/
  ├── gtFine/
  └── leftImg8bit/
Python ≥3.8 with torch, torchvision, numpy, tqdm, etc.

## 5 Notes
Modify dice_loss.py to experiment with other loss functions.