Real-Time Dense Visual SLAM with Neural Factor Representation

• DOI: 10.3390/electronics13163332
• Publication Date: 2024-08-22T10:28:51Z
• Authors: Weifeng Wei; Jie Wang; Xiaolong Xie; Jie Liu; Pengxiang Su
• Keywords: Representation; Feature (linguistics)

Developing a high-quality, real-time, dense visual SLAM system poses significant challenge in the field of computer vision. NeRF introduces neural implicit representation, marking notable advancement research. However, existing methods suffer from long runtimes and face challenges when modeling complex structures scenes. In this paper, we propose method that enables high-quality real-time reconstruction even on desktop PC. Firstly, novel scene encoding geometry appearance information as combination basis coefficient factors. This representation allows for efficient memory usage accurate high-frequency detail regions. Secondly, introduce feature integration rendering to significantly improve speed while maintaining quality color rendering. Extensive experiments synthetic real-world datasets demonstrate our achieves an average improvement more than 60% Depth L1 ATE RMSE compared state-of-the-art running at 9.8 Hz PC with 3.20 GHz Intel Core i9-12900K CPU single NVIDIA RTX 3090 GPU. remarkable highlights crucial importance approach SLAM.