Abstract
In this paper, we introduce a self-supervised deep SLAM method that robustlyoperates in dynamic scenes while accurately identifying dynamic components. Ourmethod leverages a dual-flow representation for static flow and dynamic flow,facilitating effective scene decomposition in dynamic environments. We proposea dynamic update module based on this representation and develop a dense SLAMsystem that excels in dynamic scenarios. In addition, we design aself-supervised training scheme using DINO as a prior, enabling label-freetraining. Our method achieves superior accuracy compared to otherself-supervised methods. It also matches or even surpasses the performance ofexisting supervised methods in some cases. All code and data will be madepublicly available upon acceptance.