Semantic Segmentation in Streetscape

Data Science Project 2020

Requirements

• PyTorch 1.x or 0.41

Installation

pip install -r requirements.txt

Training on Cityscapes Dataset

1. Download dataset from original image: here file annotation image: here to inputs/ and unzip. File structure:

inputs
└── cityscapes
    ├── images
    |   ├── 0a7e06.png
    │   ├── 0aab0a.png
    │   ├── ...
    |
    └── masks
        ├── 0
        |   ├── 0a7e06.png
        |   ├── ...
        |
        ├── 1
        |   ├── 0a7e06.png
        |   ├── ...
        |
        ├── ...
        | 
        └── 27
            ├── 0a7e06.png
            ├── ...

2. Preprocess

python preprocess.py

3. Train model

• with Deep Supervision

python train.py --dataset cityscapes --arch NestedUNet --num_classes 29 --deep_supervision True --epochs 200

• without Deep Supervision

python train.py --dataset cityscapes --arch NestedUNet --num_classes 29 --deep_supervision False --epochs 200

4. Evaluate and Visualize

• with Deep Supervision

python val.py --name cityscapes_NestedUNet_wDS

• without Deep Supervision

python val.py --name cityscapes_NestedUNet_woDS

Results

Cityscape Dataset (Image size: 512x256)

• Loss Function: BCEDiceLoss

Model	IoU	Dice Coeff
UNet++ without Deep Supervision	0.7934	0.8785
UNet++ with Deep Supervision	0.7767	0.8626
UNet	0.7920	0.8776

Cityscape Dataset

Model	IoU	Dice Coeff
UNet++ without Deep Supervision	0.7934	0.8785
PSPNet	0.8628	0.6771
HRNet-ORC	0.8886	0.7405
UNet	0.7920	0.8776