Semantic Segmentation

Implemented the Unet model to do semantic segmentation

dataset: Cityscapes

Download packages

Download packaes listed in the offical website and unzip them to directory data.

    ├── data
    │   ├── gtCoarse
    │   │   ├── train
    │   │   ├── train_extra
    │   │   └── val
    │   ├── gtFine
    │   │   ├── test
    │   │   ├── train
    │   │   └── val
    │   └── leftImg8bit
    │       ├── test
    │       ├── train
    │       ├── train_extra
    │       └── val

Model utilization

Model loading

Type the following command to continue trainning

    python main.py --weights /save/checkpoint.pth.tar

Predicting

Type the following command to predict on the test set (it will automatically fetch the best model weights)

    python main.py --predict

Result

Since the network is relatively large, I have tried to simplify it a little bit so that we can train it faster. See the model details in model.py file. And we can find the results in log_epoch.csv. Here's the results (after 2 epoch).

    epoch, train loss, val loss, train acc, val acc, miou
    1, 0.68217, 0.72684, 1.23130, 1.23180, 0.22013

Reference

Thanks to github repository https://github.com/hoya012/semantic-segmentation-tutorial-pytorch, I used its interface and training tricks

Thanks to github repository https://github.com/aladdinpersson/Machine-Learning-Collection/tree/master/ML/Pytorch/image_segmentation/semantic_segmentation_unet, I have studied its network architecture

AnAppleCore / Semantic-Segmentation-Unet