deeplabv3

Why we included this project

This is a small, readable DeepLabV3 codebase in PyTorch, and it comes with a model already trained on Cityscapes. You can run it on street-scene images and get segmentation output immediately, no training required. The code is compact enough to read through in an afternoon, which helps if you want to see how the atrous-convolution backbone and decoder are put together. For autonomous driving or outdoor scene-parsing work, the training and evaluation pipeline on Cityscapes is easy to adapt to your own data. One caveat: it targets PyTorch 0.4, so expect some dependency updates before it runs on a modern environment.