End-to-End RGB-D SLAM With Multi-MLPs Dense Neural Implicit Representations
Mingrui Li, Jiaming He, Yangyang Wang, Hongyu Wang
Abstract
An accurate and generalizable dense 3D reconstruction system has attracted much attention. However, existing 3D dense reconstruction systems are constrained by pre-training, and there is a need for enhanced reconstruction of texture and shape details. We propose an end-to-end 3D reconstruction system which achieves fine scene reconstruction without prior information by utilizing a neural implicit encoding. Our proposed system successfully achieves the goal through improved multi-MLP decoders ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MLM</i> ) and an effective keyframe selection strategy. Experiments conducted on the commonly used Replica and TUM RGB-D datasets demonstrate that our approach can compete with widely adopted NeRF-based SLAM methods in terms of 3D reconstruction accuracy. Moreover, our approach shows a 40.8%(except Completion Ratio) improvement in accuracy compared to NICE-SLAM [14] and does not use prior information.