AdaptSegNet

Why we included this project

Segmentation models trained on synthetic data often fall apart on real camera footage, and AdaptSegNet is a 2018 implementation that directly attacks that gap. It adapts a network's structured output space with an adversarial loss so predictions from a source domain like GTA5 line up with a target domain like Cityscapes, the standard benchmark for this problem. The PyTorch code is small enough to lift pieces from for your own experiments, and the authors' version ranked third in the VisDA Challenge when it came out. Just expect to update the dependencies, since this is research code tied to a CVPR 2018 paper rather than a maintained service.