S4AL - Semantic Segmentation with Active Semi-Supervised Learning

Training Env

We use PyTorch 1.8 with CUDA 11.1 for training all networks. These are the steps to set up the Anaconda environment:

conda create --name pyt18 anaconda python=3.7
conda activate pyt18
pip install torch==1.8.0+cu111 torchvision==0.9.0+cu111 torchaudio==0.8.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install opencv-python tensorboardX tabulate

Note: We observed on Windows, Pytorch throws some errors with respect to the dataloaders. As Ubuntu is our primary OS, a quick fix for the Windows error is to set num_workers to 0 in helpers\dataset_builder.py.

Datasets

We use CamVid and CityScapes in our paper. Please download CamVid from here and CityScapes from here. We preprocess CityScapes images to a resolution of 688 x 688, and keep original resolution for CamVid. The datasets folder is organized as:

• datasets
  • CamVid
    • annots
    • images
    • imagesets
  • CityScapes
    • annots
      • train
      • val
    • images
      • train
      • val
    • imagesets

Training Scripts

For training S4AL on CamVid:

python --dataset camvid --generations 2 --unsup_weight 1 --coldstart --region_size 30 --num_regions 4 --max_buffer_length 50 --save_dir camvid_s4al_run1

For training S4AL on CityScapes:

python --dataset cityscapes --generations 5 --unsup_weight 1 --coldstart --region_size 43 --num_regions 4 --max_buffer_length 500 --save_dir cityscapes_s4al_run1

References

List of open-source repositories used or referred for this code:

• ReCo
• DEAL
• ViewAL
• Nvidia SemSeg
• Pytorch Torchvision