Interactive Video Object Segmentation for Tracking and Mapping in Dynamic Environments*

Traditional SLAM algorithms based on static environment assumptions have made great progress in the past. However, when there are dynamic components in the scene, these static algorithms can not maintain their SoTA performance. Previous systems specially made for dynamic environments are hard to be implemented in real-world situations due to the constraint of object labels. In this paper, we propose to use iVOS( Interactive Video Object Segmentation) to memorize the dynamic objects and use a mask propagation network to propagate the dynamic masks by time. iVOS can recover masks of similar objects and does not attach any label to a specific object, which is suitable for problem with open-set identities. Compared with instance segmentation, iVOS can generate masks with more meticulous boundaries with better temporal continuity, which will leave less dynamic noise for tracking. The proposed algorithm has achieved impressive localization and mapping performance on the classic TUM RGB-D dataset and Bonn RGB-D dataset which is specifically made for dynamic scenes. Furthermore, we demonstrate the applicability by theoretical explanation.