Geometry-Aware Network for Domain Adaptive Semantic Segmentation

Measuring and alleviating the discrepancies between synthetic (source) real scene (target) data is core issue for domain adaptive semantic segmentation. Though recent works have introduced depth information in source to reinforce geometric knowledge transfer, they cannot extract intrinsic 3D of objects, including positions shapes, merely based on 2D estimated depth. In this work, we propose a novel Geometry-Aware Network Domain Adaptation (GANDA), leveraging more compact point cloud representations shrink gaps. particular, first utilize auxiliary supervision from obtain prediction target accomplish structure-texture disentanglement. Beyond estimation, explicitly exploit topology clouds generated RGB-D images further coordinate-color disentanglement pseudo-labels refinement domain. Moreover, improve classifier domain, perform domain-invariant adaptation unify segmentation results two domains. Note that our GANDA plug-and-play any existing UDA framework. Qualitative quantitative demonstrate model outperforms state-of-the-arts GTA5->Cityscapes SYNTHIA->Cityscapes.