This is an all-in-one, pytorch-based framework for training and evaluation of state-of-the-art architectures on different open-source datasets and benchmarks.

• Cityscapes: 5000 images of urban scenes with high quality annotations (19 training classes)
• CamVid: The Cambridge-driving Labeled Video Database with complete metadata (32 classes)

• U-Net (2015)
• Fast-SCNN (2018)
• BiSeNetV1(ResNet18, 101) (2018)
• HRNetV2 (2019)
• SFNet(ResNet18, STDC1, STDC2) (2020)
• DDRNet-23, -23-slim, -39 (2021)
• RegSeg, -Large (2021)
• STDCNet1, 2 (2021)
• PIDNet-S, -M, -L (2022)
• PP-LiteSeg-B, -T (2022)

The code is developed under the following configurations:

• Hardware: >= 1 GPU for both training and inference
• Software: Windows, CUDA>=10.2, Python>=3.9, PyTorch>=1.8
• Dependencies: numpy, opencv, pandas, scipy, sklearn, tensorboard, tensorboardX, tqdm

• Download Cityscapes and CamVid and unzip them into data/cityscapes and data/CamVid respectively. Run prepare_data({dataset-dir}) from dataset/{dataset-name}.py to parse image files and save path lists.