Project B: Semantic Segmentation for Road Detection

Project Overview

This project explores semantic segmentation techniques for road detection in aerial imagery. Using various architectures, including U-Net, FPN, and SPINFCN8, the project evaluates their performance on a road segmentation dataset. The results demonstrate the effectiveness of advanced segmentation models in achieving high accuracy and F1 scores.

---

Table of Contents

1. Dataset
2. Implemented Architectures
3. How to Run
4. Results
5. Ethical Considerations
6. Acknowledgments

---

Dataset

• Source: AIcrowd Road Segmentation Challenge.
• Details: The dataset consists of:
  • 100 training images with masks.
  • 50 test images without labels.
• Image Specifications:
  • Size: 400x400 pixels.
  • Channels: RGB.

Download Instructions

To download the dataset, follow these steps:

1. Navigate to the AIcrowd Road Segmentation Challenge.
2. Log in with your credentials. If you do not have an account, create one using your institutional email.
3. Download the dataset files:
   • train.zip (contains training images and masks).
   • test.zip (contains test images).
4. Extract the downloaded files:

   unzip train.zip -d data/training
   unzip test.zip -d data/test_set_images

Creating the Validation Set

The dataset does not include a predefined validation set. To create one:

1. Split the train directory into:

   • A training subset containing 90% of the images and masks.
   • A validation subset containing the remaining 10%.

2. Manually move the selected files into the following structure:

data/
├── training/
│   ├── images/  # Training images
│   └── groundtruth/   # Corresponding segmentation masks
├── validation/
│   ├── images/  # Validation images
│   └── groundtruth/   # Corresponding segmentation masks
└── test/
    └── test_set_images/  # Test images

Preprocessing Steps

• Normalization of pixel values.
• Data augmentation techniques such as flipping and rotation.
• Splitting the training set into 90% training and 10% validation subsets.

---

Implemented Architectures

1. Custom U-Net

• Baseline model implemented from scratch.
• Encoder-decoder structure with skip connections.

2. Advanced Architectures (via smp Library)

• FPN: Multiscale feature aggregation.
• DeepLabV3+: Atrous Spatial Pyramid Pooling (ASPP).
• Encoder: ResNet50, some with pretrained weights.

3. SPINFCN8

• Optimized for efficient and accurate segmentation.
• Combines spatial pyramids with fully convolutional layers.

---

How to Run

Prerequisites

• Python 3.8 or later.
• Required libraries: numpy, pytorch, torchvision, segmentation_models_pytorch.

Steps

1. Clone the repository:

   git clone https://github.com/CS-433/ml-project-2-morocco

2. Install dependencies:

   pip install -r requirements.txt

3. Create a directory structure for storing the results of the predictions and model checkpoints (model.pth). Run the following commands in your terminal:

   mkdir -p results/current/predictions

4. For each model, you can adjust the model’s hyperparameters and select the backbone in config.py. For optimal performance as demonstrated in our tests, use the configurations specified in best_config.py (copy paste in config.py) for the SPIN model.

5. In src/SPIN/config.py you can change the variable MODEL which by default is SPINRoadMapperFCN8() but you can also use SPINRoadMapper() with a speicfic backbone and weight (example : MODEL = SPINRoadMapper(model_func=segmentation.deeplabv3_resnet101, weights=segmentation.DeepLabV3_ResNet101_Weights))

6. To test a new model change the model name in the run.py file (in the case of SMP change model_name too):

   model = 'SPIN' // 'UNET' // 'SMP'
   model_name = 'FPN' // 'UNET' // 'UNET_PRETRAINED'

7. Run the training script:

   python run.py

---

Results

Quantitative Metrics

Model	F1 score (%)	Pixel Accuracy (%)
Custom U-Net	74.5	81.6
SMP U-Net (pretrained weights)	78.3	87.8
SPIN + DeepLabV3+	87.2	93.0
SPIN + FCN8	87.9	93.3

---

Ethical Considerations

Potential Risks

• Misuse in surveillance and military applications.

Mitigation Steps

• Limited the scope to civilian road segmentation applications.
• Ensured the dataset contains no personally identifiable information.

---

Acknowledgments

• Course: Machine Learning (CS-433), EPFL.
• Libraries: segmentation_models_pytorch, PyTorch.
• Dataset: AIcrowd Road Segmentation Challenge.

---

License

This project is licensed under the MIT License. See the LICENSE file for details.