iS-MAP: Neural Implicit Mapping and Positioning for Structural Environments

This work presents iS-MAP, a neural implicit RGB-D SLAM approach based on multi-scale hybrid representation in structural environments. iS-MAP encodes the scene using an efficient hybrid feature representation, which combines a 3D hash grid and multi-scale 2D feature planes. This hybrid representation is then decoded into TSDF and RGB values, leading to robust reconstruction and multilevel detail understanding. Additionally, we introduce Manhattan matching loss and structural consistency loss to fully incorporate the prior constraints of structured planes and lines. Compared with only color and depth losses, our structured losses are capable of guiding network optimization at the semantic level, resulting in more reasonable scene regularization. Experimental results on synthetic and real-world scene datasets demonstrate that our approach performs either better or competitive to existing neural implicit RGB-D SLAM methods in mapping and tracking accuracy, and predicts the most plausible reconstruction results for the unobserved structural regions. The source code will be released soon.