GSCNN

Why we included this project

Researchers and engineers working on scene understanding will find a solid reference implementation of the ICCV 2019 Gated-SCNN paper, which splits shape information into its own stream so object boundaries stay sharp instead of getting smoothed over by the main semantic branch. The repo carries the full PyTorch training and evaluation pipeline with configs and pretrained Cityscapes weights, so it is practical to reproduce the authors' boundary-aware results without re-implementing the architecture yourself. It is a good baseline when urban or street-scene segmentation needs clean edges and fine structure, not just per-pixel accuracy. Because the method depends on a specific preprocessing arrangement and a dedicated shape stream, treat the code as a base for experimentation and study rather than a drop-in production service. Also note the CC BY-NC-SA non-commercial license before building anything on it.