Gated Path Selection Network for Semantic Segmentation

Semantic segmentation is a challenging task that needs to handle large scale variations, deformations and different viewpoints. In this paper, we develop novel network named Gated Path Selection Network (GPSNet), which aims learn adaptive receptive fields. GPSNet, first design two-dimensional multi-scale - SuperNet, densely incorporates features from growing To dynamically select desirable semantic context, gate prediction module further introduced. contrast previous works focus on optimizing sample positions the regular grids, GPSNet can adaptively capture free form dense contexts. The derived fields are data-dependent, flexible model object geometric transformations. On two representative datasets, i.e., Cityscapes, ADE20K, show proposed approach consistently outperforms methods achieves competitive performance without bells whistles.