The needs for up-to-date maps are also growing fast. However, map production still heavily relies on manual annotation of geo-spatial images, i.e., aerial and satellite images, which is expensive and time consuming. This project aims to improve the mapping efficiency from geo-spatial images by over 10 times. The core technologies to be developed are semantic segmentation of geo-spatial images and shape modeling of segmentation results to generate object geometries. Accurate geometric boundaries of objects required by maps are difficult to extract using existing methods. We proposed a geometry-aware semantic segmentation method to generate accurate and clean boundaries from images, and a learning-based shape processing algorithm to convert segmentation results to geometries used by maps. The intern will learn state-of-the-art deep learning technologies and skills needed to solve real world problems. With the help of the intern, our research process will be accelerated and we can further increase the efficiency of map production.