Optimized segmentation with image inpainting for semantic mapping in dynamic scenes

Moving objects will obscure static objects in a dynamic scene. When the existing semantic segmentation methods deal with these static objects, there are often missing or errors in segmentation results. To solve this problem, we propose a framework that combines image inpainting and semantic segmentation, termed SIS. Our framework adds an image inpainting network and an identical semantic segmentation network in series following an original semantic segmentation network, which can make full use of the two semantic segmentation results to obtain the optimized semantic segmentation results in this scene. Moreover, we combined our framework with Simultaneous Localization and Mapping (SLAM), and conducted experiments on the TUM RGB-D dataset. Experimental results show, the combined SLAM system can construct a semantic octree map with more complete and stable semantic information in dynamic scenes.