IDD-Lite-Unet

Requirement

• Pytorch=0.4
• Visdom
• Opencv
• matplotlib
• pydensecrf

To directly create the conda virtual environemt please use the requirements.yml

Training procedure

1. Download the dataset in a folder "dataset" and arrange the data in the following structure: Combine the training and validation dataset. (annotations and the images are to be put in separate folders)

├── dataset
│   ├── images
│   │   ├── train
│   │   ├── val
│   │   ├── test
│   ├── annotation
│   │   ├── train
│   │   ├── val

2. Start the visdom server by the command

python -m visdom.server

2. Execute the command

bash run.sh

Weights

[https://drive.google.com/file/d/1yxCbStft75gTOriWQLGtqS_5l_OAvYnR/view?usp=sharing]

Testing procedure

1. For single scale testing:

python test.py

2. For multi-scale testing:

python multiscale_testing.py

3. For single scale testing with CRF post processing:

python test_with_postprocessing.py

Visualization

Left side : predicted classes : Right side : Ground Truth

Acknowledgement

One hundred layer Tiramisu [https://github.com/bfortuner/pytorch_tiramisu]