This repository is no longer maintained, Please visit https://github.com/srihari-humbarwadi/DeepLabV3_Plus-Tensorflow2.0
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
- void
- flat
- construction
- object
- nature
- sky
- human
- vehicle
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]
}
- Hyperparameter tuning
- Extend to cover all the original classes
- Try with Conv2DTranspose