kbalde / cityscapes-segmentation-with-unet

Multiclass image segmentation in Keras

This repository is no longer maintained, Please visit https://github.com/srihari-humbarwadi/DeepLabV3_Plus-Tensorflow2.0

Readme

The model was trained on cityscape-dataset with classes merged into 8 main classes. The model was trained for 300k iterations with lr=0.001, and no wd. Results are expected to improve drastically after hyperparameter tuning You can download the pretrained weights from here Pretrained weights

List of classes

• void
• flat
• construction
• object
• nature
• sky
• human
• vehicle

Results

The following color map has been used to display the results

color_map = {
 '0': [0, 0, 0],
 '1': [153, 153, 0],
 '2': [255, 204, 204],
 '3': [255, 0, 127],
 '4': [0, 255, 0],
 '5': [0, 204, 204],
 '6': [255, 0, 0],
 '7': [0, 0, 255]
}

Video results

Part 1 Part 2 Part 3 Part 4

Image results

To-do

• Hyperparameter tuning
• Extend to cover all the original classes
• Try with Conv2DTranspose