UPSNet

Why we included this project

UPSNet is a useful reference implementation if you work on scene understanding and want to see how panoptic segmentation was handled in the CVPR 2019 oral paper. Instead of running separate models for semantic and instance segmentation, it does both in a single end-to-end network, so you get per-pixel class labels and per-object masks from one pass. The repo ships pretrained models and reported results on COCO and Cityscapes, which makes it a convenient baseline when you are comparing newer panoptic methods or adapting the architecture to your own data. One thing to plan around: it targets PyTorch 0.4.1 and training expects 4 to 16 GPUs.